Model Classes

DeepChem maintains an extensive collection of models for scientific applications. DeepChem’s focus is on facilitating scientific applications, so we support a broad range of different machine learning frameworks (currently scikit-learn, xgboost, TensorFlow, and PyTorch) since different frameworks are more and less suited for different scientific applications.

Model Cheatsheet

If you’re just getting started with DeepChem, you’re probably interested in the basics. The place to get started is this “model cheatsheet” that lists various types of custom DeepChem models. Note that some wrappers like SklearnModel and GBDTModel which wrap external machine learning libraries are excluded, but this table is otherwise complete.

As a note about how to read this table, each row describes what’s needed to invoke a given model. Some models must be applied with given Transformer or Featurizer objects. Some models also have custom training methods. You can read off what’s needed to train the model from the table below.

Model

Type

Input Type

Transformations

Acceptable Featurizers

Fit Method

AtomicConvModel

Classifier/ Regressor

Tuple

ComplexNeighborListFragmentAtomicCoordinates

fit

ChemCeption

Classifier/ Regressor

Tensor of shape (N, M, c)

SmilesToImage

fit

CNN

Classifier/ Regressor

Tensor of shape (N, c) or (N, M, c) or (N, M, L, c)

fit

DTNNModel

Classifier/ Regressor

Matrix of shape (N, N)

CoulombMatrix

fit

DAGModel

Classifier/ Regressor

ConvMol

DAGTransformer

ConvMolFeaturizer

fit

GraphConvModel

Classifier/ Regressor

ConvMol

ConvMolFeaturizer

fit

MPNNModel

Classifier/ Regressor

WeaveMol

WeaveFeaturizer

fit

MultitaskClassifier

Classifier

Vector of shape (N,)

CircularFingerprint, RDKitDescriptors, CoulombMatrixEig, RdkitGridFeaturizer, BindingPocketFeaturizer, ElementPropertyFingerprint,

fit

MultitaskRegressor

Regressor

Vector of shape (N,)

CircularFingerprint, RDKitDescriptors, CoulombMatrixEig, RdkitGridFeaturizer, BindingPocketFeaturizer, ElementPropertyFingerprint,

fit

MultitaskFitTransformRegressor

Regressor

Vector of shape (N,)

Any

CircularFingerprint, RDKitDescriptors, CoulombMatrixEig, RdkitGridFeaturizer, BindingPocketFeaturizer, ElementPropertyFingerprint,

fit

MultitaskIRVClassifier

Classifier

Vector of shape (N,)

IRVTransformer

CircularFingerprint, RDKitDescriptors, CoulombMatrixEig, RdkitGridFeaturizer, BindingPocketFeaturizer, ElementPropertyFingerprint,

fit

ProgressiveMultitaskClassifier

Classifier

Vector of shape (N,)

CircularFingerprint, RDKitDescriptors, CoulombMatrixEig, RdkitGridFeaturizer, BindingPocketFeaturizer, ElementPropertyFingerprint,

fit

ProgressiveMultitaskRegressor

Regressor

Vector of shape (N,)

CircularFingerprint, RDKitDescriptors, CoulombMatrixEig, RdkitGridFeaturizer, BindingPocketFeaturizer, ElementPropertyFingerprint,

fit

RobustMultitaskClassifier

Classifier

Vector of shape (N,)

CircularFingerprint, RDKitDescriptors, CoulombMatrixEig, RdkitGridFeaturizer, BindingPocketFeaturizer, ElementPropertyFingerprint,

fit

RobustMultitaskRegressor

Regressor

Vector of shape (N,)

CircularFingerprint, RDKitDescriptors, CoulombMatrixEig, RdkitGridFeaturizer, BindingPocketFeaturizer, ElementPropertyFingerprint,

fit

ScScoreModel

Classifier

Vector of shape (N,)

CircularFingerprint, RDKitDescriptors, CoulombMatrixEig, RdkitGridFeaturizer, BindingPocketFeaturizer, ElementPropertyFingerprint,

fit

SeqToSeq

Sequence

Sequence

fit_sequences

Smiles2Vec

Classifier/ Regressor

Sequence

SmilesToSeq

fit

TextCNNModel

Classifier/ Regressor

String

fit

WGAN

Adversarial

Pair

fit_gan

CGCNNModel

Classifier/ Regressor

GraphData

CGCNNFeaturizer

fit

GATModel

Classifier/ Regressor

GraphData

MolGraphConvFeaturizer

fit

GCNModel

Classifier/ Regressor

GraphData

MolGraphConvFeaturizer

fit

AttentiveFPModel

Classifier/ Regressor

GraphData

MolGraphConvFeaturizer

fit

Model

class Model(model=None, model_dir: Optional[str] = None, **kwargs)[source]

Abstract base class for DeepChem models.

__init__(model=None, model_dir: Optional[str] = None, **kwargs) → None[source]

Abstract class for all models.

This is intended only for convenience of subclass implementations and should not be invoked directly.

Parameters
  • model (object) – Wrapper around ScikitLearn/Keras/Tensorflow model object.

  • model_dir (str, optional (default None)) – Path to directory where model will be stored. If not specified, model will be stored in a temporary directory.

fit_on_batch(X: Sequence, y: Sequence, w: Sequence)[source]

Perform a single step of training.

Parameters
  • X (np.ndarray) – the inputs for the batch

  • y (np.ndarray) – the labels for the batch

  • w (np.ndarray) – the weights for the batch

predict_on_batch(X: Sequence)[source]

Makes predictions on given batch of new data.

Parameters

X (np.ndarray) – Features

reload() → None[source]

Reload trained model from disk.

static get_model_filename(model_dir: str) → str[source]

Given model directory, obtain filename for the model itself.

static get_params_filename(model_dir: str) → str[source]

Given model directory, obtain filename for the model itself.

save() → None[source]

Dispatcher function for saving.

Each subclass is responsible for overriding this method.

fit(dataset: deepchem.data.datasets.Dataset)[source]

Fits a model on data in a Dataset object.

Parameters

dataset (Dataset) – the Dataset to train on

predict(dataset: deepchem.data.datasets.Dataset, transformers: List[transformers.Transformer] = []) → numpy.ndarray[source]

Uses self to make predictions on provided Dataset object.

Parameters
  • dataset (Dataset) – Dataset to make prediction on

  • transformers (List[Transformer]) – Transformers that the input data has been transformed by. The output is passed through these transformers to undo the transformations.

Returns

A numpy array of predictions the model produces.

Return type

np.ndarray

evaluate(dataset: deepchem.data.datasets.Dataset, metrics: List[deepchem.metrics.metric.Metric], transformers: List[transformers.Transformer] = [], per_task_metrics: bool = False, use_sample_weights: bool = False, n_classes: int = 2)[source]

Evaluates the performance of this model on specified dataset.

This function uses Evaluator under the hood to perform model evaluation. As a result, it inherits the same limitations of Evaluator. Namely, that only regression and classification models can be evaluated in this fashion. For generator models, you will need to overwrite this method to perform a custom evaluation.

Keyword arguments specified here will be passed to Evaluator.compute_model_performance.

Parameters
  • dataset (Dataset) – Dataset object.

  • metrics (Metric / List[Metric] / function) – The set of metrics provided. This class attempts to do some intelligent handling of input. If a single dc.metrics.Metric object is provided or a list is provided, it will evaluate self.model on these metrics. If a function is provided, it is assumed to be a metric function that this method will attempt to wrap in a dc.metrics.Metric object. A metric function must accept two arguments, y_true, y_pred both of which are np.ndarray objects and return a floating point score. The metric function may also accept a keyword argument sample_weight to account for per-sample weights.

  • transformers (List[Transformer]) – List of dc.trans.Transformer objects. These transformations must have been applied to dataset previously. The dataset will be untransformed for metric evaluation.

  • per_task_metrics (bool, optional (default False)) – If true, return computed metric for each task on multitask dataset.

  • use_sample_weights (bool, optional (default False)) – If set, use per-sample weights w.

  • n_classes (int, optional (default None)) – If specified, will use n_classes as the number of unique classes in self.dataset. Note that this argument will be ignored for regression metrics.

Returns

  • multitask_scores (dict) – Dictionary mapping names of metrics to metric scores.

  • all_task_scores (dict, optional) – If per_task_metrics == True is passed as a keyword argument, then returns a second dictionary of scores for each task separately.

get_task_type() → str[source]

Currently models can only be classifiers or regressors.

get_num_tasks() → int[source]

Get number of tasks.

Scikit-Learn Models

Scikit-learn’s models can be wrapped so that they can interact conveniently with DeepChem. Oftentimes scikit-learn models are more robust and easier to train and are a nice first model to train.

SklearnModel

class SklearnModel(model: sklearn.base.BaseEstimator, model_dir: Optional[str] = None, **kwargs)[source]

Wrapper class that wraps scikit-learn models as DeepChem models.

When you’re working with scikit-learn and DeepChem, at times it can be useful to wrap a scikit-learn model as a DeepChem model. The reason for this might be that you want to do an apples-to-apples comparison of a scikit-learn model to another DeepChem model, or perhaps you want to use the hyperparameter tuning capabilities in dc.hyper. The SklearnModel class provides a wrapper around scikit-learn models that allows scikit-learn models to be trained on Dataset objects and evaluated with the same metrics as other DeepChem models.

Notes

All SklearnModels perform learning solely in memory. This means that it may not be possible to train SklearnModel on large `Dataset`s.

__init__(model: sklearn.base.BaseEstimator, model_dir: Optional[str] = None, **kwargs)[source]
Parameters
  • model (BaseEstimator) – The model instance which inherits a scikit-learn BaseEstimator Class.

  • model_dir (str, optional (default None)) – If specified the model will be stored in this directory. Else, a temporary directory will be used.

  • model_instance (BaseEstimator (DEPRECATED)) – The model instance which inherits a scikit-learn BaseEstimator Class.

  • kwargs (dict) – kwargs[‘use_weights’] is a bool which determines if we pass weights into self.model.fit().

fit(dataset: deepchem.data.datasets.Dataset) → None[source]

Fits scikit-learn model to data.

Parameters

dataset (Dataset) – The Dataset to train this model on.

predict_on_batch(X: numpy.ndarray) → numpy.ndarray[source]

Makes predictions on batch of data.

Parameters

X (np.ndarray) – A numpy array of features.

Returns

The value is a return value of predict_proba or predict method of the scikit-learn model. If the scikit-learn model has both methods, the value is always a return value of predict_proba.

Return type

np.ndarray

predict(X: deepchem.data.datasets.Dataset, transformers: List[transformers.Transformer] = []) → numpy.ndarray[source]

Makes predictions on dataset.

Parameters
  • dataset (Dataset) – Dataset to make prediction on.

  • transformers (List[Transformer]) – Transformers that the input data has been transformed by. The output is passed through these transformers to undo the transformations.

save()[source]

Saves scikit-learn model to disk using joblib.

reload()[source]

Loads scikit-learn model from joblib file on disk.

Gradient Boosting Models

Gradient Boosting Models (LightGBM and XGBoost) can be wrapped so they can interact with DeepChem.

GBDTModel

class GBDTModel(model: sklearn.base.BaseEstimator, model_dir: Optional[str] = None, early_stopping_rounds: int = 50, eval_metric: Optional[Union[Callable, str]] = None, **kwargs)[source]

Wrapper class that wraps GBDT models as DeepChem models.

This class supports LightGBM/XGBoost models.

__init__(model: sklearn.base.BaseEstimator, model_dir: Optional[str] = None, early_stopping_rounds: int = 50, eval_metric: Optional[Union[Callable, str]] = None, **kwargs)[source]
Parameters
  • model (BaseEstimator) – The model instance of scikit-learn wrapper LightGBM/XGBoost models.

  • model_dir (str, optional (default None)) – Path to directory where model will be stored.

  • early_stopping_rounds (int, optional (default 50)) – Activates early stopping. Validation metric needs to improve at least once in every early_stopping_rounds round(s) to continue training.

  • eval_metric (Union[str, Callbale]) – If string, it should be a built-in evaluation metric to use. If callable, it should be a custom evaluation metric, see official note for more details.

fit(dataset: deepchem.data.datasets.Dataset)[source]

Fits GDBT model with all data.

First, this function splits all data into train and valid data (8:2), and finds the best n_estimators. And then, we retrain all data using best n_estimators * 1.25.

Parameters

dataset (Dataset) – The Dataset to train this model on.

fit_with_eval(train_dataset: deepchem.data.datasets.Dataset, valid_dataset: deepchem.data.datasets.Dataset)[source]

Fits GDBT model with valid data.

Parameters
  • train_dataset (Dataset) – The Dataset to train this model on.

  • valid_dataset (Dataset) – The Dataset to validate this model on.

Deep Learning Infrastructure

DeepChem maintains a lightweight layer of common deep learning model infrastructure that can be used for models built with different underlying frameworks. The losses and optimizers can be used for both TensorFlow and PyTorch models.

Losses

class Loss[source]

A loss function for use in training models.

class L1Loss[source]

The absolute difference between the true and predicted values.

class L2Loss[source]

The squared difference between the true and predicted values.

class HingeLoss[source]

The hinge loss function.

The ‘output’ argument should contain logits, and all elements of ‘labels’ should equal 0 or 1.

class BinaryCrossEntropy[source]

The cross entropy between pairs of probabilities.

The arguments should each have shape (batch_size) or (batch_size, tasks) and contain probabilities.

class CategoricalCrossEntropy[source]

The cross entropy between two probability distributions.

The arguments should each have shape (batch_size, classes) or (batch_size, tasks, classes), and represent a probability distribution over classes.

class SigmoidCrossEntropy[source]

The cross entropy between pairs of probabilities.

The arguments should each have shape (batch_size) or (batch_size, tasks). The labels should be probabilities, while the outputs should be logits that are converted to probabilities using a sigmoid function.

class SoftmaxCrossEntropy[source]

The cross entropy between two probability distributions.

The arguments should each have shape (batch_size, classes) or (batch_size, tasks, classes). The labels should be probabilities, while the outputs should be logits that are converted to probabilities using a softmax function.

class SparseSoftmaxCrossEntropy[source]

The cross entropy between two probability distributions.

The labels should have shape (batch_size) or (batch_size, tasks), and be integer class labels. The outputs have shape (batch_size, classes) or (batch_size, tasks, classes) and be logits that are converted to probabilities using a softmax function.

class VAE_ELBO[source]

The Variational AutoEncoder loss, KL Divergence Regularize + marginal log-likelihood.

This losses based on _[1]. ELBO(Evidence lower bound) lexically replaced Variational lower bound. BCE means marginal log-likelihood, and KLD means KL divergence with normal distribution. Added hyper parameter ‘kl_scale’ for KLD.

The logvar and mu should have shape (batch_size, hidden_space). The x and reconstruction_x should have (batch_size, attribute). The kl_scale should be float.

Examples

Examples for calculating loss using constant tensor.

batch_size = 2, hidden_space = 2, num of original attribute = 3 >>> import numpy as np >>> import torch >>> import tensorflow as tf >>> logvar = np.array([[1.0,1.3],[0.6,1.2]]) >>> mu = np.array([[0.2,0.7],[1.2,0.4]]) >>> x = np.array([[0.9,0.4,0.8],[0.3,0,1]]) >>> reconstruction_x = np.array([[0.8,0.3,0.7],[0.2,0,0.9]])

Case tensorflow >>> VAE_ELBO()._compute_tf_loss(tf.constant(logvar), tf.constant(mu), tf.constant(x), tf.constant(reconstruction_x)) <tf.Tensor: shape=(2,), dtype=float64, numpy=array([0.70165154, 0.76238271])>

Case pytorch >>> (VAE_ELBO()._create_pytorch_loss())(torch.tensor(logvar), torch.tensor(mu), torch.tensor(x), torch.tensor(reconstruction_x)) tensor([0.7017, 0.7624], dtype=torch.float64)

References

1

Kingma, Diederik P., and Max Welling. “Auto-encoding variational bayes.” arXiv preprint arXiv:1312.6114 (2013).

class VAE_KLDivergence[source]

The KL_divergence between hidden distribution and normal distribution.

This loss represents KL divergence losses between normal distribution(using parameter of distribution) based on _[1].

The logvar should have shape (batch_size, hidden_space) and each term represents standard deviation of hidden distribution. The mean shuold have (batch_size, hidden_space) and each term represents mean of hidden distribtuon.

Examples

Examples for calculating loss using constant tensor.

batch_size = 2, hidden_space = 2, >>> import numpy as np >>> import torch >>> import tensorflow as tf >>> logvar = np.array([[1.0,1.3],[0.6,1.2]]) >>> mu = np.array([[0.2,0.7],[1.2,0.4]])

Case tensorflow >>> VAE_KLDivergence()._compute_tf_loss(tf.constant(logvar), tf.constant(mu)) <tf.Tensor: shape=(2,), dtype=float64, numpy=array([0.17381787, 0.51425203])>

Case pytorch >>> (VAE_KLDivergence()._create_pytorch_loss())(torch.tensor(logvar), torch.tensor(mu)) tensor([0.1738, 0.5143], dtype=torch.float64)

References

1

Kingma, Diederik P., and Max Welling. “Auto-encoding variational bayes.” arXiv preprint arXiv:1312.6114 (2013).

class ShannonEntropy[source]

The ShannonEntropy of discrete-distribution.

This loss represents shannon entropy based on _[1].

The inputs should have shape (batch size, num of variable) and represents probabilites distribution.

Examples

Examples for calculating loss using constant tensor.

batch_size = 2, num_of variable = variable, >>> import numpy as np >>> import torch >>> import tensorflow as tf >>> inputs = np.array([[0.7,0.3],[0.9,0.1]])

Case tensorflow >>> ShannonEntropy()._compute_tf_loss(tf.constant(inputs)) <tf.Tensor: shape=(2,), dtype=float64, numpy=array([0.30543215, 0.16254149])>

Case pytorch >>> (ShannonEntropy()._create_pytorch_loss())(torch.tensor(inputs)) tensor([0.3054, 0.1625], dtype=torch.float64)

References

1

Chen, Ricky Xiaofeng. “A Brief Introduction to Shannon’s Information Theory.” arXiv preprint arXiv:1612.09316 (2016).

Optimizers

class Optimizer(learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule])[source]

An algorithm for optimizing a model.

This is an abstract class. Subclasses represent specific optimization algorithms.

__init__(learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule])[source]

This constructor should only be called by subclasses.

Parameters

learning_rate (float or LearningRateSchedule) – the learning rate to use for optimization

class LearningRateSchedule[source]

A schedule for changing the learning rate over the course of optimization.

This is an abstract class. Subclasses represent specific schedules.

class AdaGrad(learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule] = 0.001, initial_accumulator_value: float = 0.1, epsilon: float = 1e-07)[source]

The AdaGrad optimization algorithm.

Adagrad is an optimizer with parameter-specific learning rates, which are adapted relative to how frequently a parameter gets updated during training. The more updates a parameter receives, the smaller the updates. See [1]_ for a full reference for the algorithm.

References

1

Duchi, John, Elad Hazan, and Yoram Singer. “Adaptive subgradient methods for online learning and stochastic optimization.” Journal of machine learning research 12.7 (2011).

__init__(learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule] = 0.001, initial_accumulator_value: float = 0.1, epsilon: float = 1e-07)[source]

Construct an AdaGrad optimizer. :param learning_rate: the learning rate to use for optimization :type learning_rate: float or LearningRateSchedule :param initial_accumulator_value: a parameter of the AdaGrad algorithm :type initial_accumulator_value: float :param epsilon: a parameter of the AdaGrad algorithm :type epsilon: float

class Adam(learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule] = 0.001, beta1: float = 0.9, beta2: float = 0.999, epsilon: float = 1e-08)[source]

The Adam optimization algorithm.

__init__(learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule] = 0.001, beta1: float = 0.9, beta2: float = 0.999, epsilon: float = 1e-08)[source]

Construct an Adam optimizer.

Parameters
  • learning_rate (float or LearningRateSchedule) – the learning rate to use for optimization

  • beta1 (float) – a parameter of the Adam algorithm

  • beta2 (float) – a parameter of the Adam algorithm

  • epsilon (float) – a parameter of the Adam algorithm

class RMSProp(learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule] = 0.001, momentum: float = 0.0, decay: float = 0.9, epsilon: float = 1e-10)[source]

RMSProp Optimization algorithm.

__init__(learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule] = 0.001, momentum: float = 0.0, decay: float = 0.9, epsilon: float = 1e-10)[source]

Construct an RMSProp Optimizer.

Parameters
  • learning_rate (float or LearningRateSchedule) – the learning_rate used for optimization

  • momentum (float, default 0.0) – a parameter of the RMSProp algorithm

  • decay (float, default 0.9) – a parameter of the RMSProp algorithm

  • epsilon (float, default 1e-10) – a parameter of the RMSProp algorithm

class GradientDescent(learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule] = 0.001)[source]

The gradient descent optimization algorithm.

__init__(learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule] = 0.001)[source]

Construct a gradient descent optimizer.

Parameters

learning_rate (float or LearningRateSchedule) – the learning rate to use for optimization

class ExponentialDecay(initial_rate: float, decay_rate: float, decay_steps: int, staircase: bool = True)[source]

A learning rate that decreases exponentially with the number of training steps.

__init__(initial_rate: float, decay_rate: float, decay_steps: int, staircase: bool = True)[source]

Create an exponentially decaying learning rate.

The learning rate starts as initial_rate. Every decay_steps training steps, it is multiplied by decay_rate.

Parameters
  • initial_rate (float) – the initial learning rate

  • decay_rate (float) – the base of the exponential

  • decay_steps (int) – the number of training steps over which the rate decreases by decay_rate

  • staircase (bool) – if True, the learning rate decreases by discrete jumps every decay_steps. if False, the learning rate decreases smoothly every step

class PolynomialDecay(initial_rate: float, final_rate: float, decay_steps: int, power: float = 1.0)[source]

A learning rate that decreases from an initial value to a final value over a fixed number of training steps.

__init__(initial_rate: float, final_rate: float, decay_steps: int, power: float = 1.0)[source]

Create a smoothly decaying learning rate.

The learning rate starts as initial_rate. It smoothly decreases to final_rate over decay_steps training steps. It decays as a function of (1-step/decay_steps)**power. Once the final rate is reached, it remains there for the rest of optimization.

Parameters
  • initial_rate (float) – the initial learning rate

  • final_rate (float) – the final learning rate

  • decay_steps (int) – the number of training steps over which the rate decreases from initial_rate to final_rate

  • power (float) – the exponent controlling the shape of the decay

class LinearCosineDecay(initial_rate: float, decay_steps: int, alpha: float = 0.0, beta: float = 0.001, num_periods: float = 0.5)[source]

Applies linear cosine decay to the learning rate

__init__(initial_rate: float, decay_steps: int, alpha: float = 0.0, beta: float = 0.001, num_periods: float = 0.5)[source]
Parameters
  • learning_rate (float) –

  • learning rate (initial) –

  • decay_steps (int) –

  • of steps to decay over (number) –

  • num_periods (number of periods in the cosine part of the decay) –

Keras Models

DeepChem extensively uses Keras to build deep learning models.

KerasModel

Training loss and validation metrics can be automatically logged to Weights & Biases with the following commands:

# Install wandb in shell
pip install wandb

# Login in shell (required only once)
wandb login

# Start a W&B run in your script (refer to docs for optional parameters)
wandb.init(project="my project")

# Set `wandb` arg when creating `KerasModel`
model = KerasModel(…, wandb=True)
class KerasModel(model: tensorflow.python.keras.engine.training.Model, loss: Union[deepchem.models.losses.Loss, Callable[[List, List, List], Any]], output_types: Optional[List[str]] = None, batch_size: int = 100, model_dir: Optional[str] = None, learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule] = 0.001, optimizer: Optional[deepchem.models.optimizers.Optimizer] = None, tensorboard: bool = False, wandb: bool = False, log_frequency: int = 100, **kwargs)[source]

This is a DeepChem model implemented by a Keras model.

This class provides several advantages over using the Keras model’s fitting and prediction methods directly.

  1. It provides better integration with the rest of DeepChem, such as direct support for Datasets and Transformers.

  2. It defines the loss in a more flexible way. In particular, Keras does not support multidimensional weight matrices, which makes it impossible to implement most multitask models with Keras.

  3. It provides various additional features not found in the Keras Model class, such as uncertainty prediction and saliency mapping.

Here is a simple example of code that uses KerasModel to train a Keras model on a DeepChem dataset.

>> keras_model = tf.keras.Sequential([ >> tf.keras.layers.Dense(1000, activation=’tanh’), >> tf.keras.layers.Dense(1) >> ]) >> model = KerasModel(keras_model, loss=dc.models.losses.L2Loss()) >> model.fit(dataset)

The loss function for a model can be defined in two different ways. For models that have only a single output and use a standard loss function, you can simply provide a dc.models.losses.Loss object. This defines the loss for each sample or sample/task pair. The result is automatically multiplied by the weights and averaged over the batch. Any additional losses computed by model layers, such as weight decay penalties, are also added.

For more complicated cases, you can instead provide a function that directly computes the total loss. It must be of the form f(outputs, labels, weights), taking the list of outputs from the model, the expected values, and any weight matrices. It should return a scalar equal to the value of the loss function for the batch. No additional processing is done to the result; it is up to you to do any weighting, averaging, adding of penalty terms, etc.

You can optionally provide an output_types argument, which describes how to interpret the model’s outputs. This should be a list of strings, one for each output. You can use an arbitrary output_type for a output, but some output_types are special and will undergo extra processing:

  • ‘prediction’: This is a normal output, and will be returned by predict(). If output types are not specified, all outputs are assumed to be of this type.

  • ‘loss’: This output will be used in place of the normal outputs for computing the loss function. For example, models that output probability distributions usually do it by computing unbounded numbers (the logits), then passing them through a softmax function to turn them into probabilities. When computing the cross entropy, it is more numerically stable to use the logits directly rather than the probabilities. You can do this by having the model produce both probabilities and logits as outputs, then specifying output_types=[‘prediction’, ‘loss’]. When predict() is called, only the first output (the probabilities) will be returned. But during training, it is the second output (the logits) that will be passed to the loss function.

  • ‘variance’: This output is used for estimating the uncertainty in another output. To create a model that can estimate uncertainty, there must be the same number of ‘prediction’ and ‘variance’ outputs. Each variance output must have the same shape as the corresponding prediction output, and each element is an estimate of the variance in the corresponding prediction. Also be aware that if a model supports uncertainty, it MUST use dropout on every layer, and dropout most be enabled during uncertainty prediction. Otherwise, the uncertainties it computes will be inaccurate.

  • other: Arbitrary output_types can be used to extract outputs produced by the model, but will have no additional processing performed.

__init__(model: tensorflow.python.keras.engine.training.Model, loss: Union[deepchem.models.losses.Loss, Callable[[List, List, List], Any]], output_types: Optional[List[str]] = None, batch_size: int = 100, model_dir: Optional[str] = None, learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule] = 0.001, optimizer: Optional[deepchem.models.optimizers.Optimizer] = None, tensorboard: bool = False, wandb: bool = False, log_frequency: int = 100, **kwargs) → None[source]

Create a new KerasModel.

Parameters
  • model (tf.keras.Model) – the Keras model implementing the calculation

  • loss (dc.models.losses.Loss or function) – a Loss or function defining how to compute the training loss for each batch, as described above

  • output_types (list of strings) – the type of each output from the model, as described above

  • batch_size (int) – default batch size for training and evaluating

  • model_dir (str) – the directory on disk where the model will be stored. If this is None, a temporary directory is created.

  • learning_rate (float or LearningRateSchedule) – the learning rate to use for fitting. If optimizer is specified, this is ignored.

  • optimizer (Optimizer) – the optimizer to use for fitting. If this is specified, learning_rate is ignored.

  • tensorboard (bool) – whether to log progress to TensorBoard during training

  • wandb (bool) – whether to log progress to Weights & Biases during training

  • log_frequency (int) – The frequency at which to log data. Data is logged using logging by default. If tensorboard is set, data is also logged to TensorBoard. If wandb is set, data is also logged to Weights & Biases. Logging happens at global steps. Roughly, a global step corresponds to one batch of training. If you’d like a printout every 10 batch steps, you’d set log_frequency=10 for example.

fit(dataset: deepchem.data.datasets.Dataset, nb_epoch: int = 10, max_checkpoints_to_keep: int = 5, checkpoint_interval: int = 1000, deterministic: bool = False, restore: bool = False, variables: Optional[List[tensorflow.python.ops.variables.Variable]] = None, loss: Optional[Callable[[List, List, List], Any]] = None, callbacks: Union[Callable, List[Callable]] = [], all_losses: Optional[List[float]] = None) → float[source]

Train this model on a dataset.

Parameters
  • dataset (Dataset) – the Dataset to train on

  • nb_epoch (int) – the number of epochs to train for

  • max_checkpoints_to_keep (int) – the maximum number of checkpoints to keep. Older checkpoints are discarded.

  • checkpoint_interval (int) – the frequency at which to write checkpoints, measured in training steps. Set this to 0 to disable automatic checkpointing.

  • deterministic (bool) – if True, the samples are processed in order. If False, a different random order is used for each epoch.

  • restore (bool) – if True, restore the model from the most recent checkpoint and continue training from there. If False, retrain the model from scratch.

  • variables (list of tf.Variable) – the variables to train. If None (the default), all trainable variables in the model are used.

  • loss (function) – a function of the form f(outputs, labels, weights) that computes the loss for each batch. If None (the default), the model’s standard loss function is used.

  • callbacks (function or list of functions) – one or more functions of the form f(model, step) that will be invoked after every step. This can be used to perform validation, logging, etc.

  • all_losses (Optional[List[float]], optional (default None)) – If specified, all logged losses are appended into this list. Note that you can call fit() repeatedly with the same list and losses will continue to be appended.

Returns

Return type

The average loss over the most recent checkpoint interval

fit_generator(generator: Iterable[Tuple[Any, Any, Any]], max_checkpoints_to_keep: int = 5, checkpoint_interval: int = 1000, restore: bool = False, variables: Optional[List[tensorflow.python.ops.variables.Variable]] = None, loss: Optional[Callable[[List, List, List], Any]] = None, callbacks: Union[Callable, List[Callable]] = [], all_losses: Optional[List[float]] = None) → float[source]

Train this model on data from a generator.

Parameters
  • generator (generator) – this should generate batches, each represented as a tuple of the form (inputs, labels, weights).

  • max_checkpoints_to_keep (int) – the maximum number of checkpoints to keep. Older checkpoints are discarded.

  • checkpoint_interval (int) – the frequency at which to write checkpoints, measured in training steps. Set this to 0 to disable automatic checkpointing.

  • restore (bool) – if True, restore the model from the most recent checkpoint and continue training from there. If False, retrain the model from scratch.

  • variables (list of tf.Variable) – the variables to train. If None (the default), all trainable variables in the model are used.

  • loss (function) – a function of the form f(outputs, labels, weights) that computes the loss for each batch. If None (the default), the model’s standard loss function is used.

  • callbacks (function or list of functions) – one or more functions of the form f(model, step) that will be invoked after every step. This can be used to perform validation, logging, etc.

  • all_losses (Optional[List[float]], optional (default None)) – If specified, all logged losses are appended into this list. Note that you can call fit() repeatedly with the same list and losses will continue to be appended.

Returns

Return type

The average loss over the most recent checkpoint interval

fit_on_batch(X: Sequence, y: Sequence, w: Sequence, variables: Optional[List[tensorflow.python.ops.variables.Variable]] = None, loss: Optional[Callable[[List, List, List], Any]] = None, callbacks: Union[Callable, List[Callable]] = [], checkpoint: bool = True, max_checkpoints_to_keep: int = 5) → float[source]

Perform a single step of training.

Parameters
  • X (ndarray) – the inputs for the batch

  • y (ndarray) – the labels for the batch

  • w (ndarray) – the weights for the batch

  • variables (list of tf.Variable) – the variables to train. If None (the default), all trainable variables in the model are used.

  • loss (function) – a function of the form f(outputs, labels, weights) that computes the loss for each batch. If None (the default), the model’s standard loss function is used.

  • callbacks (function or list of functions) – one or more functions of the form f(model, step) that will be invoked after every step. This can be used to perform validation, logging, etc.

  • checkpoint (bool) – if true, save a checkpoint after performing the training step

  • max_checkpoints_to_keep (int) – the maximum number of checkpoints to keep. Older checkpoints are discarded.

Returns

Return type

the loss on the batch

predict_on_generator(generator: Iterable[Tuple[Any, Any, Any]], transformers: List[transformers.Transformer] = [], outputs: Optional[Union[tensorflow.python.framework.ops.Tensor, Sequence[tensorflow.python.framework.ops.Tensor]]] = None, output_types: Optional[Union[str, Sequence[str]]] = None) → Union[numpy.ndarray, Sequence[numpy.ndarray]][source]
Parameters
  • generator (generator) – this should generate batches, each represented as a tuple of the form (inputs, labels, weights).

  • transformers (list of dc.trans.Transformers) – Transformers that the input data has been transformed by. The output is passed through these transformers to undo the transformations.

  • outputs (Tensor or list of Tensors) – The outputs to return. If this is None, the model’s standard prediction outputs will be returned. Alternatively one or more Tensors within the model may be specified, in which case the output of those Tensors will be returned. If outputs is specified, output_types must be None.

  • output_types (String or list of Strings) – If specified, all outputs of this type will be retrieved from the model. If output_types is specified, outputs must be None.

  • Returns – a NumPy array of the model produces a single output, or a list of arrays if it produces multiple outputs

predict_on_batch(X: Sequence, transformers: List[transformers.Transformer] = [], outputs: Optional[Union[tensorflow.python.framework.ops.Tensor, Sequence[tensorflow.python.framework.ops.Tensor]]] = None) → Union[numpy.ndarray, Sequence[numpy.ndarray]][source]

Generates predictions for input samples, processing samples in a batch.

Parameters
  • X (ndarray) – the input data, as a Numpy array.

  • transformers (list of dc.trans.Transformers) – Transformers that the input data has been transformed by. The output is passed through these transformers to undo the transformations.

  • outputs (Tensor or list of Tensors) – The outputs to return. If this is None, the model’s standard prediction outputs will be returned. Alternatively one or more Tensors within the model may be specified, in which case the output of those Tensors will be returned.

Returns

  • a NumPy array of the model produces a single output, or a list of arrays

  • if it produces multiple outputs

predict_uncertainty_on_batch(X: Sequence, masks: int = 50) → Union[Tuple[numpy.ndarray, numpy.ndarray], Sequence[Tuple[numpy.ndarray, numpy.ndarray]]][source]

Predict the model’s outputs, along with the uncertainty in each one.

The uncertainty is computed as described in https://arxiv.org/abs/1703.04977. It involves repeating the prediction many times with different dropout masks. The prediction is computed as the average over all the predictions. The uncertainty includes both the variation among the predicted values (epistemic uncertainty) and the model’s own estimates for how well it fits the data (aleatoric uncertainty). Not all models support uncertainty prediction.

Parameters
  • X (ndarray) – the input data, as a Numpy array.

  • masks (int) – the number of dropout masks to average over

Returns

  • for each output, a tuple (y_pred, y_std) where y_pred is the predicted

  • value of the output, and each element of y_std estimates the standard

  • deviation of the corresponding element of y_pred

predict(dataset: deepchem.data.datasets.Dataset, transformers: List[transformers.Transformer] = [], outputs: Optional[Union[tensorflow.python.framework.ops.Tensor, Sequence[tensorflow.python.framework.ops.Tensor]]] = None, output_types: Optional[List[str]] = None) → Union[numpy.ndarray, Sequence[numpy.ndarray]][source]

Uses self to make predictions on provided Dataset object.

Parameters
  • dataset (dc.data.Dataset) – Dataset to make prediction on

  • transformers (list of dc.trans.Transformers) – Transformers that the input data has been transformed by. The output is passed through these transformers to undo the transformations.

  • outputs (Tensor or list of Tensors) – The outputs to return. If this is None, the model’s standard prediction outputs will be returned. Alternatively one or more Tensors within the model may be specified, in which case the output of those Tensors will be returned.

  • output_types (String or list of Strings) – If specified, all outputs of this type will be retrieved from the model. If output_types is specified, outputs must be None.

Returns

  • a NumPy array of the model produces a single output, or a list of arrays

  • if it produces multiple outputs

predict_embedding(dataset: deepchem.data.datasets.Dataset) → Union[numpy.ndarray, Sequence[numpy.ndarray]][source]

Predicts embeddings created by underlying model if any exist. An embedding must be specified to have output_type of ‘embedding’ in the model definition.

Parameters

dataset (dc.data.Dataset) – Dataset to make prediction on

Returns

  • a NumPy array of the embeddings model produces, or a list

  • of arrays if it produces multiple embeddings

predict_uncertainty(dataset: deepchem.data.datasets.Dataset, masks: int = 50) → Union[Tuple[numpy.ndarray, numpy.ndarray], Sequence[Tuple[numpy.ndarray, numpy.ndarray]]][source]

Predict the model’s outputs, along with the uncertainty in each one.

The uncertainty is computed as described in https://arxiv.org/abs/1703.04977. It involves repeating the prediction many times with different dropout masks. The prediction is computed as the average over all the predictions. The uncertainty includes both the variation among the predicted values (epistemic uncertainty) and the model’s own estimates for how well it fits the data (aleatoric uncertainty). Not all models support uncertainty prediction.

Parameters
  • dataset (dc.data.Dataset) – Dataset to make prediction on

  • masks (int) – the number of dropout masks to average over

Returns

  • for each output, a tuple (y_pred, y_std) where y_pred is the predicted

  • value of the output, and each element of y_std estimates the standard

  • deviation of the corresponding element of y_pred

evaluate_generator(generator: Iterable[Tuple[Any, Any, Any]], metrics: List[deepchem.metrics.metric.Metric], transformers: List[transformers.Transformer] = [], per_task_metrics: bool = False)[source]

Evaluate the performance of this model on the data produced by a generator.

Parameters
  • generator (generator) – this should generate batches, each represented as a tuple of the form (inputs, labels, weights).

  • metric (list of deepchem.metrics.Metric) – Evaluation metric

  • transformers (list of dc.trans.Transformers) – Transformers that the input data has been transformed by. The output is passed through these transformers to undo the transformations.

  • per_task_metrics (bool) – If True, return per-task scores.

Returns

Maps tasks to scores under metric.

Return type

dict

compute_saliency(X: numpy.ndarray) → Union[numpy.ndarray, Sequence[numpy.ndarray]][source]

Compute the saliency map for an input sample.

This computes the Jacobian matrix with the derivative of each output element with respect to each input element. More precisely,

  • If this model has a single output, it returns a matrix of shape (output_shape, input_shape) with the derivatives.

  • If this model has multiple outputs, it returns a list of matrices, one for each output.

This method cannot be used on models that take multiple inputs.

Parameters

X (ndarray) – the input data for a single sample

Returns

Return type

the Jacobian matrix, or a list of matrices

default_generator(dataset: deepchem.data.datasets.Dataset, epochs: int = 1, mode: str = 'fit', deterministic: bool = True, pad_batches: bool = True) → Iterable[Tuple[List, List, List]][source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

save_checkpoint(max_checkpoints_to_keep: int = 5, model_dir: Optional[str] = None) → None[source]

Save a checkpoint to disk.

Usually you do not need to call this method, since fit() saves checkpoints automatically. If you have disabled automatic checkpointing during fitting, this can be called to manually write checkpoints.

Parameters
  • max_checkpoints_to_keep (int) – the maximum number of checkpoints to keep. Older checkpoints are discarded.

  • model_dir (str, default None) – Model directory to save checkpoint to. If None, revert to self.model_dir

get_checkpoints(model_dir: Optional[str] = None)[source]

Get a list of all available checkpoint files.

Parameters

model_dir (str, default None) – Directory to get list of checkpoints from. Reverts to self.model_dir if None

restore(checkpoint: Optional[str] = None, model_dir: Optional[str] = None) → None[source]

Reload the values of all variables from a checkpoint file.

Parameters
  • checkpoint (str) – the path to the checkpoint file to load. If this is None, the most recent checkpoint will be chosen automatically. Call get_checkpoints() to get a list of all available checkpoints.

  • model_dir (str, default None) – Directory to restore checkpoint from. If None, use self.model_dir.

get_global_step() → int[source]

Get the number of steps of fitting that have been performed.

load_from_pretrained(source_model: deepchem.models.keras_model.KerasModel, assignment_map: Optional[Dict[Any, Any]] = None, value_map: Optional[Dict[Any, Any]] = None, checkpoint: Optional[str] = None, model_dir: Optional[str] = None, include_top: bool = True, inputs: Optional[Sequence[Any]] = None, **kwargs) → None[source]

Copies variable values from a pretrained model. source_model can either be a pretrained model or a model with the same architecture. value_map is a variable-value dictionary. If no value_map is provided, the variable values are restored to the source_model from a checkpoint and a default value_map is created. assignment_map is a dictionary mapping variables from the source_model to the current model. If no assignment_map is provided, one is made from scratch and assumes the model is composed of several different layers, with the final one being a dense layer. include_top is used to control whether or not the final dense layer is used. The default assignment map is useful in cases where the type of task is different (classification vs regression) and/or number of tasks in the setting.

Parameters
  • source_model (dc.KerasModel, required) – source_model can either be the pretrained model or a dc.KerasModel with the same architecture as the pretrained model. It is used to restore from a checkpoint, if value_map is None and to create a default assignment map if assignment_map is None

  • assignment_map (Dict, default None) – Dictionary mapping the source_model variables and current model variables

  • value_map (Dict, default None) – Dictionary containing source_model trainable variables mapped to numpy arrays. If value_map is None, the values are restored and a default variable map is created using the restored values

  • checkpoint (str, default None) – the path to the checkpoint file to load. If this is None, the most recent checkpoint will be chosen automatically. Call get_checkpoints() to get a list of all available checkpoints

  • model_dir (str, default None) – Restore model from custom model directory if needed

  • include_top (bool, default True) – if True, copies the weights and bias associated with the final dense layer. Used only when assignment map is None

  • inputs (List, input tensors for model) – if not None, then the weights are built for both the source and self. This option is useful only for models that are built by subclassing tf.keras.Model, and not using the functional API by tf.keras

MultitaskRegressor

class MultitaskRegressor(n_tasks: int, n_features: int, layer_sizes: Sequence[int] = [1000], weight_init_stddevs: Union[float, Sequence[float]] = 0.02, bias_init_consts: Union[float, Sequence[float]] = 1.0, weight_decay_penalty: float = 0.0, weight_decay_penalty_type: str = 'l2', dropouts: Union[float, Sequence[float]] = 0.5, activation_fns: Union[Callable, str, Sequence[Union[Callable, str]]] = <function relu>, uncertainty: bool = False, residual: bool = False, **kwargs)[source]

A fully connected network for multitask regression.

This class provides lots of options for customizing aspects of the model: the number and widths of layers, the activation functions, regularization methods, etc.

It optionally can compose the model from pre-activation residual blocks, as described in https://arxiv.org/abs/1603.05027, rather than a simple stack of dense layers. This often leads to easier training, especially when using a large number of layers. Note that residual blocks can only be used when successive layers have the same width. Wherever the layer width changes, a simple dense layer will be used even if residual=True.

__init__(n_tasks: int, n_features: int, layer_sizes: Sequence[int] = [1000], weight_init_stddevs: Union[float, Sequence[float]] = 0.02, bias_init_consts: Union[float, Sequence[float]] = 1.0, weight_decay_penalty: float = 0.0, weight_decay_penalty_type: str = 'l2', dropouts: Union[float, Sequence[float]] = 0.5, activation_fns: Union[Callable, str, Sequence[Union[Callable, str]]] = <function relu>, uncertainty: bool = False, residual: bool = False, **kwargs) → None[source]

Create a MultitaskRegressor.

In addition to the following arguments, this class also accepts all the keywork arguments from TensorGraph.

Parameters
  • n_tasks (int) – number of tasks

  • n_features (int) – number of features

  • layer_sizes (list) – the size of each dense layer in the network. The length of this list determines the number of layers.

  • weight_init_stddevs (list or float) – the standard deviation of the distribution to use for weight initialization of each layer. The length of this list should equal len(layer_sizes)+1. The final element corresponds to the output layer. Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • bias_init_consts (list or float) – the value to initialize the biases in each layer to. The length of this list should equal len(layer_sizes)+1. The final element corresponds to the output layer. Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • weight_decay_penalty (float) – the magnitude of the weight decay penalty to use

  • weight_decay_penalty_type (str) – the type of penalty to use for weight decay, either ‘l1’ or ‘l2’

  • dropouts (list or float) – the dropout probablity to use for each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • activation_fns (list or object) – the Tensorflow activation function to apply to each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • uncertainty (bool) – if True, include extra outputs and loss terms to enable the uncertainty in outputs to be predicted

  • residual (bool) – if True, the model will be composed of pre-activation residual blocks instead of a simple stack of dense layers.

default_generator(dataset: deepchem.data.datasets.Dataset, epochs: int = 1, mode: str = 'fit', deterministic: bool = True, pad_batches: bool = True) → Iterable[Tuple[List, List, List]][source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

MultitaskFitTransformRegressor

class MultitaskFitTransformRegressor(n_tasks: int, n_features: int, fit_transformers: Sequence[transformers.Transformer] = [], batch_size: int = 50, **kwargs)[source]

Implements a MultitaskRegressor that performs on-the-fly transformation during fit/predict.

Examples

>>> n_samples = 10
>>> n_features = 3
>>> n_tasks = 1
>>> ids = np.arange(n_samples)
>>> X = np.random.rand(n_samples, n_features, n_features)
>>> y = np.zeros((n_samples, n_tasks))
>>> w = np.ones((n_samples, n_tasks))
>>> dataset = dc.data.NumpyDataset(X, y, w, ids)
>>> fit_transformers = [dc.trans.CoulombFitTransformer(dataset)]
>>> model = dc.models.MultitaskFitTransformRegressor(n_tasks, [n_features, n_features],
...     dropouts=[0.], learning_rate=0.003, weight_init_stddevs=[np.sqrt(6)/np.sqrt(1000)],
...     batch_size=n_samples, fit_transformers=fit_transformers)
>>> model.n_features
12
__init__(n_tasks: int, n_features: int, fit_transformers: Sequence[transformers.Transformer] = [], batch_size: int = 50, **kwargs)[source]

Create a MultitaskFitTransformRegressor.

In addition to the following arguments, this class also accepts all the keywork arguments from MultitaskRegressor.

Parameters
  • n_tasks (int) – number of tasks

  • n_features (list or int) – number of features

  • fit_transformers (list) – List of dc.trans.FitTransformer objects

default_generator(dataset: deepchem.data.datasets.Dataset, epochs: int = 1, mode: str = 'fit', deterministic: bool = True, pad_batches: bool = True) → Iterable[Tuple[List, List, List]][source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

predict_on_generator(generator: Iterable[Tuple[Any, Any, Any]], transformers: List[transformers.Transformer] = [], outputs: Optional[Union[tensorflow.python.framework.ops.Tensor, Sequence[tensorflow.python.framework.ops.Tensor]]] = None, output_types: Optional[Union[str, Sequence[str]]] = None) → Union[numpy.ndarray, Sequence[numpy.ndarray]][source]
Parameters
  • generator (generator) – this should generate batches, each represented as a tuple of the form (inputs, labels, weights).

  • transformers (list of dc.trans.Transformers) – Transformers that the input data has been transformed by. The output is passed through these transformers to undo the transformations.

  • outputs (Tensor or list of Tensors) – The outputs to return. If this is None, the model’s standard prediction outputs will be returned. Alternatively one or more Tensors within the model may be specified, in which case the output of those Tensors will be returned. If outputs is specified, output_types must be None.

  • output_types (String or list of Strings) – If specified, all outputs of this type will be retrieved from the model. If output_types is specified, outputs must be None.

  • Returns – a NumPy array of the model produces a single output, or a list of arrays if it produces multiple outputs

MultitaskClassifier

class MultitaskClassifier(n_tasks: int, n_features: int, layer_sizes: Sequence[int] = [1000], weight_init_stddevs: Union[float, Sequence[float]] = 0.02, bias_init_consts: Union[float, Sequence[float]] = 1.0, weight_decay_penalty: float = 0.0, weight_decay_penalty_type: str = 'l2', dropouts: Union[float, Sequence[float]] = 0.5, activation_fns: Union[Callable, str, Sequence[Union[Callable, str]]] = <function relu>, n_classes: int = 2, residual: bool = False, **kwargs)[source]

A fully connected network for multitask classification.

This class provides lots of options for customizing aspects of the model: the number and widths of layers, the activation functions, regularization methods, etc.

It optionally can compose the model from pre-activation residual blocks, as described in https://arxiv.org/abs/1603.05027, rather than a simple stack of dense layers. This often leads to easier training, especially when using a large number of layers. Note that residual blocks can only be used when successive layers have the same width. Wherever the layer width changes, a simple dense layer will be used even if residual=True.

__init__(n_tasks: int, n_features: int, layer_sizes: Sequence[int] = [1000], weight_init_stddevs: Union[float, Sequence[float]] = 0.02, bias_init_consts: Union[float, Sequence[float]] = 1.0, weight_decay_penalty: float = 0.0, weight_decay_penalty_type: str = 'l2', dropouts: Union[float, Sequence[float]] = 0.5, activation_fns: Union[Callable, str, Sequence[Union[Callable, str]]] = <function relu>, n_classes: int = 2, residual: bool = False, **kwargs) → None[source]

Create a MultitaskClassifier.

In addition to the following arguments, this class also accepts all the keyword arguments from TensorGraph.

Parameters
  • n_tasks (int) – number of tasks

  • n_features (int) – number of features

  • layer_sizes (list) – the size of each dense layer in the network. The length of this list determines the number of layers.

  • weight_init_stddevs (list or float) – the standard deviation of the distribution to use for weight initialization of each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • bias_init_consts (list or float) – the value to initialize the biases in each layer to. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • weight_decay_penalty (float) – the magnitude of the weight decay penalty to use

  • weight_decay_penalty_type (str) – the type of penalty to use for weight decay, either ‘l1’ or ‘l2’

  • dropouts (list or float) – the dropout probablity to use for each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • activation_fns (list or object) – the Tensorflow activation function to apply to each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • n_classes (int) – the number of classes

  • residual (bool) – if True, the model will be composed of pre-activation residual blocks instead of a simple stack of dense layers.

default_generator(dataset: deepchem.data.datasets.Dataset, epochs: int = 1, mode: str = 'fit', deterministic: bool = True, pad_batches: bool = True) → Iterable[Tuple[List, List, List]][source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

TensorflowMultitaskIRVClassifier

class TensorflowMultitaskIRVClassifier(*args, **kwargs)[source]
__init__(*args, **kwargs)[source]

Initialize MultitaskIRVClassifier

Parameters
  • n_tasks (int) – Number of tasks

  • K (int) – Number of nearest neighbours used in classification

  • penalty (float) – Amount of penalty (l2 or l1 applied)

RobustMultitaskClassifier

class RobustMultitaskClassifier(n_tasks, n_features, layer_sizes=[1000], weight_init_stddevs=0.02, bias_init_consts=1.0, weight_decay_penalty=0.0, weight_decay_penalty_type='l2', dropouts=0.5, activation_fns=<function relu>, n_classes=2, bypass_layer_sizes=[100], bypass_weight_init_stddevs=[0.02], bypass_bias_init_consts=[1.0], bypass_dropouts=[0.5], **kwargs)[source]

Implements a neural network for robust multitasking.

The key idea of this model is to have bypass layers that feed directly from features to task output. This might provide some flexibility toroute around challenges in multitasking with destructive interference.

References

This technique was introduced in [1]_

1

Ramsundar, Bharath, et al. “Is multitask deep learning practical for pharma?.” Journal of chemical information and modeling 57.8 (2017): 2068-2076.

__init__(n_tasks, n_features, layer_sizes=[1000], weight_init_stddevs=0.02, bias_init_consts=1.0, weight_decay_penalty=0.0, weight_decay_penalty_type='l2', dropouts=0.5, activation_fns=<function relu>, n_classes=2, bypass_layer_sizes=[100], bypass_weight_init_stddevs=[0.02], bypass_bias_init_consts=[1.0], bypass_dropouts=[0.5], **kwargs)[source]

Create a RobustMultitaskClassifier.

Parameters
  • n_tasks (int) – number of tasks

  • n_features (int) – number of features

  • layer_sizes (list) – the size of each dense layer in the network. The length of this list determines the number of layers.

  • weight_init_stddevs (list or float) – the standard deviation of the distribution to use for weight initialization of each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • bias_init_consts (list or loat) – the value to initialize the biases in each layer to. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • weight_decay_penalty (float) – the magnitude of the weight decay penalty to use

  • weight_decay_penalty_type (str) – the type of penalty to use for weight decay, either ‘l1’ or ‘l2’

  • dropouts (list or float) – the dropout probablity to use for each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • activation_fns (list or object) – the Tensorflow activation function to apply to each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • n_classes (int) – the number of classes

  • bypass_layer_sizes (list) – the size of each dense layer in the bypass network. The length of this list determines the number of bypass layers.

  • bypass_weight_init_stddevs (list or float) – the standard deviation of the distribution to use for weight initialization of bypass layers. same requirements as weight_init_stddevs

  • bypass_bias_init_consts (list or float) – the value to initialize the biases in bypass layers same requirements as bias_init_consts

  • bypass_dropouts (list or float) – the dropout probablity to use for bypass layers. same requirements as dropouts

default_generator(dataset, epochs=1, mode='fit', deterministic=True, pad_batches=True)[source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

RobustMultitaskRegressor

class RobustMultitaskRegressor(n_tasks, n_features, layer_sizes=[1000], weight_init_stddevs=0.02, bias_init_consts=1.0, weight_decay_penalty=0.0, weight_decay_penalty_type='l2', dropouts=0.5, activation_fns=<function relu>, bypass_layer_sizes=[100], bypass_weight_init_stddevs=[0.02], bypass_bias_init_consts=[1.0], bypass_dropouts=[0.5], **kwargs)[source]

Implements a neural network for robust multitasking.

The key idea of this model is to have bypass layers that feed directly from features to task output. This might provide some flexibility toroute around challenges in multitasking with destructive interference.

References

1

Ramsundar, Bharath, et al. “Is multitask deep learning practical for pharma?.” Journal of chemical information and modeling 57.8 (2017): 2068-2076.

__init__(n_tasks, n_features, layer_sizes=[1000], weight_init_stddevs=0.02, bias_init_consts=1.0, weight_decay_penalty=0.0, weight_decay_penalty_type='l2', dropouts=0.5, activation_fns=<function relu>, bypass_layer_sizes=[100], bypass_weight_init_stddevs=[0.02], bypass_bias_init_consts=[1.0], bypass_dropouts=[0.5], **kwargs)[source]

Create a RobustMultitaskRegressor.

Parameters
  • n_tasks (int) – number of tasks

  • n_features (int) – number of features

  • layer_sizes (list) – the size of each dense layer in the network. The length of this list determines the number of layers.

  • weight_init_stddevs (list or float) – the standard deviation of the distribution to use for weight initialization of each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • bias_init_consts (list or loat) – the value to initialize the biases in each layer to. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • weight_decay_penalty (float) – the magnitude of the weight decay penalty to use

  • weight_decay_penalty_type (str) – the type of penalty to use for weight decay, either ‘l1’ or ‘l2’

  • dropouts (list or float) – the dropout probablity to use for each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • activation_fns (list or object) – the Tensorflow activation function to apply to each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • bypass_layer_sizes (list) – the size of each dense layer in the bypass network. The length of this list determines the number of bypass layers.

  • bypass_weight_init_stddevs (list or float) – the standard deviation of the distribution to use for weight initialization of bypass layers. same requirements as weight_init_stddevs

  • bypass_bias_init_consts (list or float) – the value to initialize the biases in bypass layers same requirements as bias_init_consts

  • bypass_dropouts (list or float) – the dropout probablity to use for bypass layers. same requirements as dropouts

default_generator(dataset: deepchem.data.datasets.Dataset, epochs: int = 1, mode: str = 'fit', deterministic: bool = True, pad_batches: bool = True) → Iterable[Tuple[List, List, List]][source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

ProgressiveMultitaskClassifier

class ProgressiveMultitaskClassifier(n_tasks, n_features, alpha_init_stddevs=0.02, layer_sizes=[1000], weight_init_stddevs=0.02, bias_init_consts=1.0, weight_decay_penalty=0.0, weight_decay_penalty_type='l2', dropouts=0.5, activation_fns=<function relu>, **kwargs)[source]

Implements a progressive multitask neural network for classification.

Progressive Networks: https://arxiv.org/pdf/1606.04671v3.pdf

Progressive networks allow for multitask learning where each task gets a new column of weights. As a result, there is no exponential forgetting where previous tasks are ignored.

__init__(n_tasks, n_features, alpha_init_stddevs=0.02, layer_sizes=[1000], weight_init_stddevs=0.02, bias_init_consts=1.0, weight_decay_penalty=0.0, weight_decay_penalty_type='l2', dropouts=0.5, activation_fns=<function relu>, **kwargs)[source]

Creates a progressive network.

Only listing parameters specific to progressive networks here.

Parameters
  • n_tasks (int) – Number of tasks

  • n_features (int) – Number of input features

  • alpha_init_stddevs (list) – List of standard-deviations for alpha in adapter layers.

  • layer_sizes (list) – the size of each dense layer in the network. The length of this list determines the number of layers.

  • weight_init_stddevs (list or float) – the standard deviation of the distribution to use for weight initialization of each layer. The length of this list should equal len(layer_sizes)+1. The final element corresponds to the output layer. Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • bias_init_consts (list or float) – the value to initialize the biases in each layer to. The length of this list should equal len(layer_sizes)+1. The final element corresponds to the output layer. Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • weight_decay_penalty (float) – the magnitude of the weight decay penalty to use

  • weight_decay_penalty_type (str) – the type of penalty to use for weight decay, either ‘l1’ or ‘l2’

  • dropouts (list or float) – the dropout probablity to use for each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • activation_fns (list or object) – the Tensorflow activation function to apply to each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

ProgressiveMultitaskRegressor

class ProgressiveMultitaskRegressor(n_tasks, n_features, alpha_init_stddevs=0.02, layer_sizes=[1000], weight_init_stddevs=0.02, bias_init_consts=1.0, weight_decay_penalty=0.0, weight_decay_penalty_type='l2', dropouts=0.5, activation_fns=<function relu>, n_outputs=1, **kwargs)[source]

Implements a progressive multitask neural network for regression.

Progressive networks allow for multitask learning where each task gets a new column of weights. As a result, there is no exponential forgetting where previous tasks are ignored.

References

See [1]_ for a full description of the progressive architecture

1

Rusu, Andrei A., et al. “Progressive neural networks.” arXiv preprint arXiv:1606.04671 (2016).

__init__(n_tasks, n_features, alpha_init_stddevs=0.02, layer_sizes=[1000], weight_init_stddevs=0.02, bias_init_consts=1.0, weight_decay_penalty=0.0, weight_decay_penalty_type='l2', dropouts=0.5, activation_fns=<function relu>, n_outputs=1, **kwargs)[source]

Creates a progressive network.

Only listing parameters specific to progressive networks here.

Parameters
  • n_tasks (int) – Number of tasks

  • n_features (int) – Number of input features

  • alpha_init_stddevs (list) – List of standard-deviations for alpha in adapter layers.

  • layer_sizes (list) – the size of each dense layer in the network. The length of this list determines the number of layers.

  • weight_init_stddevs (list or float) – the standard deviation of the distribution to use for weight initialization of each layer. The length of this list should equal len(layer_sizes)+1. The final element corresponds to the output layer. Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • bias_init_consts (list or float) – the value to initialize the biases in each layer to. The length of this list should equal len(layer_sizes)+1. The final element corresponds to the output layer. Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • weight_decay_penalty (float) – the magnitude of the weight decay penalty to use

  • weight_decay_penalty_type (str) – the type of penalty to use for weight decay, either ‘l1’ or ‘l2’

  • dropouts (list or float) – the dropout probablity to use for each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • activation_fns (list or object) – the Tensorflow activation function to apply to each layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

add_adapter(all_layers, task, layer_num)[source]

Add an adapter connection for given task/layer combo

fit(dataset, nb_epoch=10, max_checkpoints_to_keep=5, checkpoint_interval=1000, deterministic=False, restore=False, **kwargs)[source]

Train this model on a dataset.

Parameters
  • dataset (Dataset) – the Dataset to train on

  • nb_epoch (int) – the number of epochs to train for

  • max_checkpoints_to_keep (int) – the maximum number of checkpoints to keep. Older checkpoints are discarded.

  • checkpoint_interval (int) – the frequency at which to write checkpoints, measured in training steps. Set this to 0 to disable automatic checkpointing.

  • deterministic (bool) – if True, the samples are processed in order. If False, a different random order is used for each epoch.

  • restore (bool) – if True, restore the model from the most recent checkpoint and continue training from there. If False, retrain the model from scratch.

  • variables (list of tf.Variable) – the variables to train. If None (the default), all trainable variables in the model are used.

  • loss (function) – a function of the form f(outputs, labels, weights) that computes the loss for each batch. If None (the default), the model’s standard loss function is used.

  • callbacks (function or list of functions) – one or more functions of the form f(model, step) that will be invoked after every step. This can be used to perform validation, logging, etc.

  • all_losses (Optional[List[float]], optional (default None)) – If specified, all logged losses are appended into this list. Note that you can call fit() repeatedly with the same list and losses will continue to be appended.

Returns

Return type

The average loss over the most recent checkpoint interval

fit_task(dataset, task, nb_epoch=10, max_checkpoints_to_keep=5, checkpoint_interval=1000, deterministic=False, restore=False, **kwargs)[source]

Fit one task.

WeaveModel

class WeaveModel(n_tasks: int, n_atom_feat: Union[int, Sequence[int]] = 75, n_pair_feat: Union[int, Sequence[int]] = 14, n_hidden: int = 50, n_graph_feat: int = 128, n_weave: int = 2, fully_connected_layer_sizes: List[int] = [2000, 100], conv_weight_init_stddevs: Union[float, Sequence[float]] = 0.03, weight_init_stddevs: Union[float, Sequence[float]] = 0.01, bias_init_consts: Union[float, Sequence[float]] = 0.0, weight_decay_penalty: float = 0.0, weight_decay_penalty_type: str = 'l2', dropouts: Union[float, Sequence[float]] = 0.25, final_conv_activation_fn: Optional[Union[Callable, str]] = <function tanh>, activation_fns: Union[Callable, str, Sequence[Union[Callable, str]]] = <function relu>, batch_normalize: bool = True, batch_normalize_kwargs: Dict = {'fused': False, 'renorm': True}, gaussian_expand: bool = True, compress_post_gaussian_expansion: bool = False, mode: str = 'classification', n_classes: int = 2, batch_size: int = 100, **kwargs)[source]

Implements Google-style Weave Graph Convolutions

This model implements the Weave style graph convolutions from [1]_.

The biggest difference between WeaveModel style convolutions and GraphConvModel style convolutions is that Weave convolutions model bond features explicitly. This has the side effect that it needs to construct a NxN matrix explicitly to model bond interactions. This may cause scaling issues, but may possibly allow for better modeling of subtle bond effects.

Note that [1]_ introduces a whole variety of different architectures for Weave models. The default settings in this class correspond to the W2N2 variant from [1]_ which is the most commonly used variant..

Examples

Here’s an example of how to fit a WeaveModel on a tiny sample dataset.

>>> import numpy as np
>>> import deepchem as dc
>>> featurizer = dc.feat.WeaveFeaturizer()
>>> X = featurizer(["C", "CC"])
>>> y = np.array([1, 0])
>>> dataset = dc.data.NumpyDataset(X, y)
>>> model = dc.models.WeaveModel(n_tasks=1, n_weave=2, fully_connected_layer_sizes=[2000, 1000], mode="classification")
>>> loss = model.fit(dataset)

Note

In general, the use of batch normalization can cause issues with NaNs. If you’re having trouble with NaNs while using this model, consider setting batch_normalize_kwargs={“trainable”: False} or turning off batch normalization entirely with batch_normalize=False.

References

1

Kearnes, Steven, et al. “Molecular graph convolutions: moving beyond

fingerprints.” Journal of computer-aided molecular design 30.8 (2016): 595-608.

__init__(n_tasks: int, n_atom_feat: Union[int, Sequence[int]] = 75, n_pair_feat: Union[int, Sequence[int]] = 14, n_hidden: int = 50, n_graph_feat: int = 128, n_weave: int = 2, fully_connected_layer_sizes: List[int] = [2000, 100], conv_weight_init_stddevs: Union[float, Sequence[float]] = 0.03, weight_init_stddevs: Union[float, Sequence[float]] = 0.01, bias_init_consts: Union[float, Sequence[float]] = 0.0, weight_decay_penalty: float = 0.0, weight_decay_penalty_type: str = 'l2', dropouts: Union[float, Sequence[float]] = 0.25, final_conv_activation_fn: Optional[Union[Callable, str]] = <function tanh>, activation_fns: Union[Callable, str, Sequence[Union[Callable, str]]] = <function relu>, batch_normalize: bool = True, batch_normalize_kwargs: Dict = {'fused': False, 'renorm': True}, gaussian_expand: bool = True, compress_post_gaussian_expansion: bool = False, mode: str = 'classification', n_classes: int = 2, batch_size: int = 100, **kwargs)[source]
Parameters
  • n_tasks (int) – Number of tasks

  • n_atom_feat (int, optional (default 75)) – Number of features per atom. Note this is 75 by default and should be 78 if chirality is used by WeaveFeaturizer.

  • n_pair_feat (int, optional (default 14)) – Number of features per pair of atoms.

  • n_hidden (int, optional (default 50)) – Number of units(convolution depths) in corresponding hidden layer

  • n_graph_feat (int, optional (default 128)) – Number of output features for each molecule(graph)

  • n_weave (int, optional (default 2)) – The number of weave layers in this model.

  • fully_connected_layer_sizes (list (default [2000, 100])) – The size of each dense layer in the network. The length of this list determines the number of layers.

  • conv_weight_init_stddevs (list or float (default 0.03)) – The standard deviation of the distribution to use for weight initialization of each convolutional layer. The length of this lisst should equal n_weave. Alternatively, this may be a single value instead of a list, in which case the same value is used for each layer.

  • weight_init_stddevs (list or float (default 0.01)) – The standard deviation of the distribution to use for weight initialization of each fully connected layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • bias_init_consts (list or float (default 0.0)) – The value to initialize the biases in each fully connected layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • weight_decay_penalty (float (default 0.0)) – The magnitude of the weight decay penalty to use

  • weight_decay_penalty_type (str (default "l2")) – The type of penalty to use for weight decay, either ‘l1’ or ‘l2’

  • dropouts (list or float (default 0.25)) – The dropout probablity to use for each fully connected layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • final_conv_activation_fn (Optional[KerasActivationFn] (default tf.nn.tanh)) – The Tensorflow activation funcntion to apply to the final convolution at the end of the weave convolutions. If None, then no activate is applied (hence linear).

  • activation_fns (list or object (default tf.nn.relu)) – The Tensorflow activation function to apply to each fully connected layer. The length of this list should equal len(layer_sizes). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • batch_normalize (bool, optional (default True)) – If this is turned on, apply batch normalization before applying activation functions on convolutional and fully connected layers.

  • batch_normalize_kwargs (Dict, optional (default {“renorm”=True, “fused”: False})) – Batch normalization is a complex layer which has many potential argumentswhich change behavior. This layer accepts user-defined parameters which are passed to all BatchNormalization layers in WeaveModel, WeaveLayer, and WeaveGather.

  • gaussian_expand (boolean, optional (default True)) – Whether to expand each dimension of atomic features by gaussian histogram

  • compress_post_gaussian_expansion (bool, optional (default False)) – If True, compress the results of the Gaussian expansion back to the original dimensions of the input.

  • mode (str (default "classification")) – Either “classification” or “regression” for type of model.

  • n_classes (int (default 2)) – Number of classes to predict (only used in classification mode)

  • batch_size (int (default 100)) – Batch size used by this model for training.

compute_features_on_batch(X_b)[source]

Compute tensors that will be input into the model from featurized representation.

The featurized input to WeaveModel is instances of WeaveMol created by WeaveFeaturizer. This method converts input WeaveMol objects into tensors used by the Keras implementation to compute WeaveModel outputs.

Parameters

X_b (np.ndarray) – A numpy array with dtype=object where elements are WeaveMol objects.

Returns

  • atom_feat (np.ndarray) – Of shape (N_atoms, N_atom_feat).

  • pair_feat (np.ndarray) – Of shape (N_pairs, N_pair_feat). Note that N_pairs will depend on the number of pairs being considered. If max_pair_distance is None, then this will be N_atoms**2. Else it will be the number of pairs within the specifed graph distance.

  • pair_split (np.ndarray) – Of shape (N_pairs,). The i-th entry in this array will tell you the originating atom for this pair (the “source”). Note that pairs are symmetric so for a pair (a, b), both a and b will separately be sources at different points in this array.

  • atom_split (np.ndarray) – Of shape (N_atoms,). The i-th entry in this array will be the molecule with the i-th atom belongs to.

  • atom_to_pair (np.ndarray) – Of shape (N_pairs, 2). The i-th row in this array will be the array [a, b] if (a, b) is a pair to be considered. (Note by symmetry, this implies some other row will contain [b, a].

default_generator(dataset: deepchem.data.datasets.Dataset, epochs: int = 1, mode: str = 'fit', deterministic: bool = True, pad_batches: bool = True) → Iterable[Tuple[List, List, List]][source]

Convert a dataset into the tensors needed for learning.

Parameters
  • dataset (dc.data.Dataset) – Dataset to convert

  • epochs (int, optional (Default 1)) – Number of times to walk over dataset

  • mode (str, optional (Default 'fit')) – Ignored in this implementation.

  • deterministic (bool, optional (Default True)) – Whether the dataset should be walked in a deterministic fashion

  • pad_batches (bool, optional (Default True)) – If true, each returned batch will have size self.batch_size.

Returns

Return type

Iterator which walks over the batches

DTNNModel

class DTNNModel(n_tasks, n_embedding=30, n_hidden=100, n_distance=100, distance_min=- 1, distance_max=18, output_activation=True, mode='regression', dropout=0.0, **kwargs)[source]

Deep Tensor Neural Networks

This class implements deep tensor neural networks as first defined in [1]_

References

1

Schütt, Kristof T., et al. “Quantum-chemical insights from deep

tensor neural networks.” Nature communications 8.1 (2017): 1-8.

__init__(n_tasks, n_embedding=30, n_hidden=100, n_distance=100, distance_min=- 1, distance_max=18, output_activation=True, mode='regression', dropout=0.0, **kwargs)[source]
Parameters
  • n_tasks (int) – Number of tasks

  • n_embedding (int, optional) – Number of features per atom.

  • n_hidden (int, optional) – Number of features for each molecule after DTNNStep

  • n_distance (int, optional) – granularity of distance matrix step size will be (distance_max-distance_min)/n_distance

  • distance_min (float, optional) – minimum distance of atom pairs, default = -1 Angstorm

  • distance_max (float, optional) – maximum distance of atom pairs, default = 18 Angstorm

  • mode (str) – Only “regression” is currently supported.

  • dropout (float) – the dropout probablity to use.

compute_features_on_batch(X_b)[source]

Computes the values for different Feature Layers on given batch

A tf.py_func wrapper is written around this when creating the input_fn for tf.Estimator

default_generator(dataset, epochs=1, mode='fit', deterministic=True, pad_batches=True)[source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

DAGModel

class DAGModel(n_tasks, max_atoms=50, n_atom_feat=75, n_graph_feat=30, n_outputs=30, layer_sizes=[100], layer_sizes_gather=[100], dropout=None, mode='classification', n_classes=2, uncertainty=False, batch_size=100, **kwargs)[source]

Directed Acyclic Graph models for molecular property prediction.

This model is based on the following paper:

Lusci, Alessandro, Gianluca Pollastri, and Pierre Baldi. “Deep architectures and deep learning in chemoinformatics: the prediction of aqueous solubility for drug-like molecules.” Journal of chemical information and modeling 53.7 (2013): 1563-1575.

The basic idea for this paper is that a molecule is usually viewed as an undirected graph. However, you can convert it to a series of directed graphs. The idea is that for each atom, you make a DAG using that atom as the vertex of the DAG and edges pointing “inwards” to it. This transformation is implemented in dc.trans.transformers.DAGTransformer.UG_to_DAG.

This model accepts ConvMols as input, just as GraphConvModel does, but these ConvMol objects must be transformed by dc.trans.DAGTransformer.

As a note, performance of this model can be a little sensitive to initialization. It might be worth training a few different instantiations to get a stable set of parameters.

__init__(n_tasks, max_atoms=50, n_atom_feat=75, n_graph_feat=30, n_outputs=30, layer_sizes=[100], layer_sizes_gather=[100], dropout=None, mode='classification', n_classes=2, uncertainty=False, batch_size=100, **kwargs)[source]
Parameters
  • n_tasks (int) – Number of tasks.

  • max_atoms (int, optional) – Maximum number of atoms in a molecule, should be defined based on dataset.

  • n_atom_feat (int, optional) – Number of features per atom.

  • n_graph_feat (int, optional) – Number of features for atom in the graph.

  • n_outputs (int, optional) – Number of features for each molecule.

  • layer_sizes (list of int, optional) – List of hidden layer size(s) in the propagation step: length of this list represents the number of hidden layers, and each element is the width of corresponding hidden layer.

  • layer_sizes_gather (list of int, optional) – List of hidden layer size(s) in the gather step.

  • dropout (None or float, optional) – Dropout probability, applied after each propagation step and gather step.

  • mode (str, optional) – Either “classification” or “regression” for type of model.

  • n_classes (int) – the number of classes to predict (only used in classification mode)

  • uncertainty (bool) – if True, include extra outputs and loss terms to enable the uncertainty in outputs to be predicted

default_generator(dataset, epochs=1, mode='fit', deterministic=True, pad_batches=True)[source]

Convert a dataset into the tensors needed for learning

GraphConvModel

class GraphConvModel(n_tasks: int, graph_conv_layers: List[int] = [64, 64], dense_layer_size: int = 128, dropout: float = 0.0, mode: str = 'classification', number_atom_features: int = 75, n_classes: int = 2, batch_size: int = 100, batch_normalize: bool = True, uncertainty: bool = False, **kwargs)[source]

Graph Convolutional Models.

This class implements the graph convolutional model from the following paper [1]_. These graph convolutions start with a per-atom set of descriptors for each atom in a molecule, then combine and recombine these descriptors over convolutional layers. following [1]_.

References

1

Duvenaud, David K., et al. “Convolutional networks on graphs for

learning molecular fingerprints.” Advances in neural information processing systems. 2015.

__init__(n_tasks: int, graph_conv_layers: List[int] = [64, 64], dense_layer_size: int = 128, dropout: float = 0.0, mode: str = 'classification', number_atom_features: int = 75, n_classes: int = 2, batch_size: int = 100, batch_normalize: bool = True, uncertainty: bool = False, **kwargs)[source]

The wrapper class for graph convolutions.

Note that since the underlying _GraphConvKerasModel class is specified using imperative subclassing style, this model cannout make predictions for arbitrary outputs.

Parameters
  • n_tasks (int) – Number of tasks

  • graph_conv_layers (list of int) – Width of channels for the Graph Convolution Layers

  • dense_layer_size (int) – Width of channels for Atom Level Dense Layer before GraphPool

  • dropout (list or float) – the dropout probablity to use for each layer. The length of this list should equal len(graph_conv_layers)+1 (one value for each convolution layer, and one for the dense layer). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • mode (str) – Either “classification” or “regression”

  • number_atom_features (int) – 75 is the default number of atom features created, but this can vary if various options are passed to the function atom_features in graph_features

  • n_classes (int) – the number of classes to predict (only used in classification mode)

  • batch_normalize (True) – if True, apply batch normalization to model

  • uncertainty (bool) – if True, include extra outputs and loss terms to enable the uncertainty in outputs to be predicted

default_generator(dataset, epochs=1, mode='fit', deterministic=True, pad_batches=True)[source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

MPNNModel

class MPNNModel(n_tasks, n_atom_feat=70, n_pair_feat=8, n_hidden=100, T=5, M=10, mode='regression', dropout=0.0, n_classes=2, uncertainty=False, batch_size=100, **kwargs)[source]

Message Passing Neural Network,

Message Passing Neural Networks treat graph convolutional operations as an instantiation of a more general message passing schem. Recall that message passing in a graph is when nodes in a graph send each other “messages” and update their internal state as a consequence of these messages.

Ordering structures in this model are built according to [1]_

References

1

Vinyals, Oriol, Samy Bengio, and Manjunath Kudlur. “Order matters:

Sequence to sequence for sets.” arXiv preprint arXiv:1511.06391 (2015).

__init__(n_tasks, n_atom_feat=70, n_pair_feat=8, n_hidden=100, T=5, M=10, mode='regression', dropout=0.0, n_classes=2, uncertainty=False, batch_size=100, **kwargs)[source]
Parameters
  • n_tasks (int) – Number of tasks

  • n_atom_feat (int, optional) – Number of features per atom.

  • n_pair_feat (int, optional) – Number of features per pair of atoms.

  • n_hidden (int, optional) – Number of units(convolution depths) in corresponding hidden layer

  • n_graph_feat (int, optional) – Number of output features for each molecule(graph)

  • dropout (float) – the dropout probablity to use.

  • n_classes (int) – the number of classes to predict (only used in classification mode)

  • uncertainty (bool) – if True, include extra outputs and loss terms to enable the uncertainty in outputs to be predicted

default_generator(dataset, epochs=1, mode='fit', deterministic=True, pad_batches=True)[source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

ScScoreModel

class ScScoreModel(n_features, layer_sizes=[300, 300, 300], dropouts=0.0, **kwargs)[source]

https://pubs.acs.org/doi/abs/10.1021/acs.jcim.7b00622 Several definitions of molecular complexity exist to facilitate prioritization of lead compounds, to identify diversity-inducing and complexifying reactions, and to guide retrosynthetic searches. In this work, we focus on synthetic complexity and reformalize its definition to correlate with the expected number of reaction steps required to produce a target molecule, with implicit knowledge about what compounds are reasonable starting materials. We train a neural network model on 12 million reactions from the Reaxys database to impose a pairwise inequality constraint enforcing the premise of this definition: that on average, the products of published chemical reactions should be more synthetically complex than their corresponding reactants. The learned metric (SCScore) exhibits highly desirable nonlinear behavior, particularly in recognizing increases in synthetic complexity throughout a number of linear synthetic routes.

Our model here actually uses hingeloss instead of the shifted relu loss in https://github.com/connorcoley/scscore.

This could cause issues differentiation issues with compounds that are “close” to each other in “complexity”

__init__(n_features, layer_sizes=[300, 300, 300], dropouts=0.0, **kwargs)[source]
Parameters
  • n_features (int) – number of features per molecule

  • layer_sizes (list of int) – size of each hidden layer

  • dropouts (int) – droupout to apply to each hidden layer

  • kwargs – This takes all kwards as TensorGraph

default_generator(dataset, epochs=1, mode='fit', deterministic=True, pad_batches=True)[source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

SeqToSeq

class SeqToSeq(input_tokens, output_tokens, max_output_length, encoder_layers=4, decoder_layers=4, embedding_dimension=512, dropout=0.0, reverse_input=True, variational=False, annealing_start_step=5000, annealing_final_step=10000, **kwargs)[source]

Implements sequence to sequence translation models.

The model is based on the description in Sutskever et al., “Sequence to Sequence Learning with Neural Networks” (https://arxiv.org/abs/1409.3215), although this implementation uses GRUs instead of LSTMs. The goal is to take sequences of tokens as input, and translate each one into a different output sequence. The input and output sequences can both be of variable length, and an output sequence need not have the same length as the input sequence it was generated from. For example, these models were originally developed for use in natural language processing. In that context, the input might be a sequence of English words, and the output might be a sequence of French words. The goal would be to train the model to translate sentences from English to French.

The model consists of two parts called the “encoder” and “decoder”. Each one consists of a stack of recurrent layers. The job of the encoder is to transform the input sequence into a single, fixed length vector called the “embedding”. That vector contains all relevant information from the input sequence. The decoder then transforms the embedding vector into the output sequence.

These models can be used for various purposes. First and most obviously, they can be used for sequence to sequence translation. In any case where you have sequences of tokens, and you want to translate each one into a different sequence, a SeqToSeq model can be trained to perform the translation.

Another possible use case is transforming variable length sequences into fixed length vectors. Many types of models require their inputs to have a fixed shape, which makes it difficult to use them with variable sized inputs (for example, when the input is a molecule, and different molecules have different numbers of atoms). In that case, you can train a SeqToSeq model as an autoencoder, so that it tries to make the output sequence identical to the input one. That forces the embedding vector to contain all information from the original sequence. You can then use the encoder for transforming sequences into fixed length embedding vectors, suitable to use as inputs to other types of models.

Another use case is to train the decoder for use as a generative model. Here again you begin by training the SeqToSeq model as an autoencoder. Once training is complete, you can supply arbitrary embedding vectors, and transform each one into an output sequence. When used in this way, you typically train it as a variational autoencoder. This adds random noise to the encoder, and also adds a constraint term to the loss that forces the embedding vector to have a unit Gaussian distribution. You can then pick random vectors from a Gaussian distribution, and the output sequences should follow the same distribution as the training data.

When training as a variational autoencoder, it is best to use KL cost annealing, as described in https://arxiv.org/abs/1511.06349. The constraint term in the loss is initially set to 0, so the optimizer just tries to minimize the reconstruction loss. Once it has made reasonable progress toward that, the constraint term can be gradually turned back on. The range of steps over which this happens is configurable.

__init__(input_tokens, output_tokens, max_output_length, encoder_layers=4, decoder_layers=4, embedding_dimension=512, dropout=0.0, reverse_input=True, variational=False, annealing_start_step=5000, annealing_final_step=10000, **kwargs)[source]

Construct a SeqToSeq model.

In addition to the following arguments, this class also accepts all the keyword arguments from TensorGraph.

Parameters
  • input_tokens (list) – a list of all tokens that may appear in input sequences

  • output_tokens (list) – a list of all tokens that may appear in output sequences

  • max_output_length (int) – the maximum length of output sequence that may be generated

  • encoder_layers (int) – the number of recurrent layers in the encoder

  • decoder_layers (int) – the number of recurrent layers in the decoder

  • embedding_dimension (int) – the width of the embedding vector. This also is the width of all recurrent layers.

  • dropout (float) – the dropout probability to use during training

  • reverse_input (bool) – if True, reverse the order of input sequences before sending them into the encoder. This can improve performance when working with long sequences.

  • variational (bool) – if True, train the model as a variational autoencoder. This adds random noise to the encoder, and also constrains the embedding to follow a unit Gaussian distribution.

  • annealing_start_step (int) – the step (that is, batch) at which to begin turning on the constraint term for KL cost annealing

  • annealing_final_step (int) – the step (that is, batch) at which to finish turning on the constraint term for KL cost annealing

fit_sequences(sequences, max_checkpoints_to_keep=5, checkpoint_interval=1000, restore=False)[source]

Train this model on a set of sequences

Parameters
  • sequences (iterable) – the training samples to fit to. Each sample should be represented as a tuple of the form (input_sequence, output_sequence).

  • max_checkpoints_to_keep (int) – the maximum number of checkpoints to keep. Older checkpoints are discarded.

  • checkpoint_interval (int) – the frequency at which to write checkpoints, measured in training steps.

  • restore (bool) – if True, restore the model from the most recent checkpoint and continue training from there. If False, retrain the model from scratch.

predict_from_sequences(sequences, beam_width=5)[source]

Given a set of input sequences, predict the output sequences.

The prediction is done using a beam search with length normalization.

Parameters
  • sequences (iterable) – the input sequences to generate a prediction for

  • beam_width (int) – the beam width to use for searching. Set to 1 to use a simple greedy search.

predict_from_embeddings(embeddings, beam_width=5)[source]

Given a set of embedding vectors, predict the output sequences.

The prediction is done using a beam search with length normalization.

Parameters
  • embeddings (iterable) – the embedding vectors to generate predictions for

  • beam_width (int) – the beam width to use for searching. Set to 1 to use a simple greedy search.

predict_embeddings(sequences)[source]

Given a set of input sequences, compute the embedding vectors.

Parameters

sequences (iterable) – the input sequences to generate an embedding vector for

GAN

class GAN(n_generators=1, n_discriminators=1, **kwargs)[source]

Implements Generative Adversarial Networks.

A Generative Adversarial Network (GAN) is a type of generative model. It consists of two parts called the “generator” and the “discriminator”. The generator takes random noise as input and transforms it into an output that (hopefully) resembles the training data. The discriminator takes a set of samples as input and tries to distinguish the real training samples from the ones created by the generator. Both of them are trained together. The discriminator tries to get better and better at telling real from false data, while the generator tries to get better and better at fooling the discriminator.

In many cases there also are additional inputs to the generator and discriminator. In that case it is known as a Conditional GAN (CGAN), since it learns a distribution that is conditional on the values of those inputs. They are referred to as “conditional inputs”.

Many variations on this idea have been proposed, and new varieties of GANs are constantly being proposed. This class tries to make it very easy to implement straightforward GANs of the most conventional types. At the same time, it tries to be flexible enough that it can be used to implement many (but certainly not all) variations on the concept.

To define a GAN, you must create a subclass that provides implementations of the following methods:

get_noise_input_shape() get_data_input_shapes() create_generator() create_discriminator()

If you want your GAN to have any conditional inputs you must also implement:

get_conditional_input_shapes()

The following methods have default implementations that are suitable for most conventional GANs. You can override them if you want to customize their behavior:

create_generator_loss() create_discriminator_loss() get_noise_batch()

This class allows a GAN to have multiple generators and discriminators, a model known as MIX+GAN. It is described in Arora et al., “Generalization and Equilibrium in Generative Adversarial Nets (GANs)” (https://arxiv.org/abs/1703.00573). This can lead to better models, and is especially useful for reducing mode collapse, since different generators can learn different parts of the distribution. To use this technique, simply specify the number of generators and discriminators when calling the constructor. You can then tell predict_gan_generator() which generator to use for predicting samples.

__init__(n_generators=1, n_discriminators=1, **kwargs)[source]

Construct a GAN.

In addition to the parameters listed below, this class accepts all the keyword arguments from KerasModel.

Parameters
  • n_generators (int) – the number of generators to include

  • n_discriminators (int) – the number of discriminators to include

get_noise_input_shape()[source]

Get the shape of the generator’s noise input layer.

Subclasses must override this to return a tuple giving the shape of the noise input. The actual Input layer will be created automatically. The dimension corresponding to the batch size should be omitted.

get_data_input_shapes()[source]

Get the shapes of the inputs for training data.

Subclasses must override this to return a list of tuples, each giving the shape of one of the inputs. The actual Input layers will be created automatically. This list of shapes must also match the shapes of the generator’s outputs. The dimension corresponding to the batch size should be omitted.

get_conditional_input_shapes()[source]

Get the shapes of any conditional inputs.

Subclasses may override this to return a list of tuples, each giving the shape of one of the conditional inputs. The actual Input layers will be created automatically. The dimension corresponding to the batch size should be omitted.

The default implementation returns an empty list, meaning there are no conditional inputs.

get_noise_batch(batch_size)[source]

Get a batch of random noise to pass to the generator.

This should return a NumPy array whose shape matches the one returned by get_noise_input_shape(). The default implementation returns normally distributed values. Subclasses can override this to implement a different distribution.

create_generator()[source]

Create and return a generator.

Subclasses must override this to construct the generator. The returned value should be a tf.keras.Model whose inputs are a batch of noise, followed by any conditional inputs. The number and shapes of its outputs must match the return value from get_data_input_shapes(), since generated data must have the same form as training data.

create_discriminator()[source]

Create and return a discriminator.

Subclasses must override this to construct the discriminator. The returned value should be a tf.keras.Model whose inputs are all data inputs, followed by any conditional inputs. Its output should be a one dimensional tensor containing the probability of each sample being a training sample.

create_generator_loss(discrim_output)[source]

Create the loss function for the generator.

The default implementation is appropriate for most cases. Subclasses can override this if the need to customize it.

Parameters

discrim_output (Tensor) – the output from the discriminator on a batch of generated data. This is its estimate of the probability that each sample is training data.

Returns

Return type

A Tensor equal to the loss function to use for optimizing the generator.

create_discriminator_loss(discrim_output_train, discrim_output_gen)[source]

Create the loss function for the discriminator.

The default implementation is appropriate for most cases. Subclasses can override this if the need to customize it.

Parameters
  • discrim_output_train (Tensor) – the output from the discriminator on a batch of training data. This is its estimate of the probability that each sample is training data.

  • discrim_output_gen (Tensor) – the output from the discriminator on a batch of generated data. This is its estimate of the probability that each sample is training data.

Returns

Return type

A Tensor equal to the loss function to use for optimizing the discriminator.

fit_gan(batches, generator_steps=1.0, max_checkpoints_to_keep=5, checkpoint_interval=1000, restore=False)[source]

Train this model on data.

Parameters
  • batches (iterable) – batches of data to train the discriminator on, each represented as a dict that maps Inputs to values. It should specify values for all members of data_inputs and conditional_inputs.

  • generator_steps (float) – the number of training steps to perform for the generator for each batch. This can be used to adjust the ratio of training steps for the generator and discriminator. For example, 2.0 will perform two training steps for every batch, while 0.5 will only perform one training step for every two batches.

  • max_checkpoints_to_keep (int) – the maximum number of checkpoints to keep. Older checkpoints are discarded.

  • checkpoint_interval (int) – the frequency at which to write checkpoints, measured in batches. Set this to 0 to disable automatic checkpointing.

  • restore (bool) – if True, restore the model from the most recent checkpoint before training it.

predict_gan_generator(batch_size=1, noise_input=None, conditional_inputs=[], generator_index=0)[source]

Use the GAN to generate a batch of samples.

Parameters
  • batch_size (int) – the number of samples to generate. If either noise_input or conditional_inputs is specified, this argument is ignored since the batch size is then determined by the size of that argument.

  • noise_input (array) – the value to use for the generator’s noise input. If None (the default), get_noise_batch() is called to generate a random input, so each call will produce a new set of samples.

  • conditional_inputs (list of arrays) – the values to use for all conditional inputs. This must be specified if the GAN has any conditional inputs.

  • generator_index (int) – the index of the generator (between 0 and n_generators-1) to use for generating the samples.

Returns

  • An array (if the generator has only one output) or list of arrays (if it has

  • multiple outputs) containing the generated samples.

WGAN

class WGAN(gradient_penalty=10.0, **kwargs)[source]

Implements Wasserstein Generative Adversarial Networks.

This class implements Wasserstein Generative Adversarial Networks (WGANs) as described in Arjovsky et al., “Wasserstein GAN” (https://arxiv.org/abs/1701.07875). A WGAN is conceptually rather different from a conventional GAN, but in practical terms very similar. It reinterprets the discriminator (often called the “critic” in this context) as learning an approximation to the Earth Mover distance between the training and generated distributions. The generator is then trained to minimize that distance. In practice, this just means using slightly different loss functions for training the generator and discriminator.

WGANs have theoretical advantages over conventional GANs, and they often work better in practice. In addition, the discriminator’s loss function can be directly interpreted as a measure of the quality of the model. That is an advantage over conventional GANs, where the loss does not directly convey information about the quality of the model.

The theory WGANs are based on requires the discriminator’s gradient to be bounded. The original paper achieved this by clipping its weights. This class instead does it by adding a penalty term to the discriminator’s loss, as described in https://arxiv.org/abs/1704.00028. This is sometimes found to produce better results.

There are a few other practical differences between GANs and WGANs. In a conventional GAN, the discriminator’s output must be between 0 and 1 so it can be interpreted as a probability. In a WGAN, it should produce an unbounded output that can be interpreted as a distance.

When training a WGAN, you also should usually use a smaller value for generator_steps. Conventional GANs rely on keeping the generator and discriminator “in balance” with each other. If the discriminator ever gets too good, it becomes impossible for the generator to fool it and training stalls. WGANs do not have this problem, and in fact the better the discriminator is, the easier it is for the generator to improve. It therefore usually works best to perform several training steps on the discriminator for each training step on the generator.

__init__(gradient_penalty=10.0, **kwargs)[source]

Construct a WGAN.

In addition to the following, this class accepts all the keyword arguments from GAN and KerasModel.

Parameters

gradient_penalty (float) – the magnitude of the gradient penalty loss

create_generator_loss(discrim_output)[source]

Create the loss function for the generator.

The default implementation is appropriate for most cases. Subclasses can override this if the need to customize it.

Parameters

discrim_output (Tensor) – the output from the discriminator on a batch of generated data. This is its estimate of the probability that each sample is training data.

Returns

Return type

A Tensor equal to the loss function to use for optimizing the generator.

create_discriminator_loss(discrim_output_train, discrim_output_gen)[source]

Create the loss function for the discriminator.

The default implementation is appropriate for most cases. Subclasses can override this if the need to customize it.

Parameters
  • discrim_output_train (Tensor) – the output from the discriminator on a batch of training data. This is its estimate of the probability that each sample is training data.

  • discrim_output_gen (Tensor) – the output from the discriminator on a batch of generated data. This is its estimate of the probability that each sample is training data.

Returns

Return type

A Tensor equal to the loss function to use for optimizing the discriminator.

CNN

class CNN(n_tasks, n_features, dims, layer_filters=[100], kernel_size=5, strides=1, weight_init_stddevs=0.02, bias_init_consts=1.0, weight_decay_penalty=0.0, weight_decay_penalty_type='l2', dropouts=0.5, activation_fns=<function relu>, dense_layer_size=1000, pool_type='max', mode='classification', n_classes=2, uncertainty=False, residual=False, padding='valid', **kwargs)[source]

A 1, 2, or 3 dimensional convolutional network for either regression or classification.

The network consists of the following sequence of layers:

  • A configurable number of convolutional layers

  • A global pooling layer (either max pool or average pool)

  • A final dense layer to compute the output

It optionally can compose the model from pre-activation residual blocks, as described in https://arxiv.org/abs/1603.05027, rather than a simple stack of convolution layers. This often leads to easier training, especially when using a large number of layers. Note that residual blocks can only be used when successive layers have the same output shape. Wherever the output shape changes, a simple convolution layer will be used even if residual=True.

__init__(n_tasks, n_features, dims, layer_filters=[100], kernel_size=5, strides=1, weight_init_stddevs=0.02, bias_init_consts=1.0, weight_decay_penalty=0.0, weight_decay_penalty_type='l2', dropouts=0.5, activation_fns=<function relu>, dense_layer_size=1000, pool_type='max', mode='classification', n_classes=2, uncertainty=False, residual=False, padding='valid', **kwargs)[source]

Create a CNN.

In addition to the following arguments, this class also accepts all the keyword arguments from TensorGraph.

Parameters
  • n_tasks (int) – number of tasks

  • n_features (int) – number of features

  • dims (int) – the number of dimensions to apply convolutions over (1, 2, or 3)

  • layer_filters (list) – the number of output filters for each convolutional layer in the network. The length of this list determines the number of layers.

  • kernel_size (int, tuple, or list) – a list giving the shape of the convolutional kernel for each layer. Each element may be either an int (use the same kernel width for every dimension) or a tuple (the kernel width along each dimension). Alternatively this may be a single int or tuple instead of a list, in which case the same kernel shape is used for every layer.

  • strides (int, tuple, or list) – a list giving the stride between applications of the kernel for each layer. Each element may be either an int (use the same stride for every dimension) or a tuple (the stride along each dimension). Alternatively this may be a single int or tuple instead of a list, in which case the same stride is used for every layer.

  • weight_init_stddevs (list or float) – the standard deviation of the distribution to use for weight initialization of each layer. The length of this list should equal len(layer_filters)+1, where the final element corresponds to the dense layer. Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • bias_init_consts (list or loat) – the value to initialize the biases in each layer to. The length of this list should equal len(layer_filters)+1, where the final element corresponds to the dense layer. Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • weight_decay_penalty (float) – the magnitude of the weight decay penalty to use

  • weight_decay_penalty_type (str) – the type of penalty to use for weight decay, either ‘l1’ or ‘l2’

  • dropouts (list or float) – the dropout probablity to use for each layer. The length of this list should equal len(layer_filters). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • activation_fns (list or object) – the Tensorflow activation function to apply to each layer. The length of this list should equal len(layer_filters). Alternatively this may be a single value instead of a list, in which case the same value is used for every layer.

  • pool_type (str) – the type of pooling layer to use, either ‘max’ or ‘average’

  • mode (str) – Either ‘classification’ or ‘regression’

  • n_classes (int) – the number of classes to predict (only used in classification mode)

  • uncertainty (bool) – if True, include extra outputs and loss terms to enable the uncertainty in outputs to be predicted

  • residual (bool) – if True, the model will be composed of pre-activation residual blocks instead of a simple stack of convolutional layers.

  • padding (str) – the type of padding to use for convolutional layers, either ‘valid’ or ‘same’

default_generator(dataset, epochs=1, mode='fit', deterministic=True, pad_batches=True)[source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

TextCNNModel

class TextCNNModel(n_tasks, char_dict, seq_length, n_embedding=75, kernel_sizes=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20], num_filters=[100, 200, 200, 200, 200, 100, 100, 100, 100, 100, 160, 160], dropout=0.25, mode='classification', **kwargs)[source]

A Convolutional neural network on smiles strings

Reimplementation of the discriminator module in ORGAN [1]_ . Originated from 2.

This model applies multiple 1D convolutional filters to the padded strings, then max-over-time pooling is applied on all filters, extracting one feature per filter. All features are concatenated and transformed through several hidden layers to form predictions.

This model is initially developed for sentence-level classification tasks, with words represented as vectors. In this implementation, SMILES strings are dissected into characters and transformed to one-hot vectors in a similar way. The model can be used for general molecular-level classification or regression tasks. It is also used in the ORGAN model as discriminator.

Training of the model only requires SMILES strings input, all featurized datasets that include SMILES in the ids attribute are accepted. PDBbind, QM7 and QM7b are not supported. To use the model, build_char_dict should be called first before defining the model to build character dict of input dataset, example can be found in examples/delaney/delaney_textcnn.py

References

1

Guimaraes, Gabriel Lima, et al. “Objective-reinforced generative adversarial networks (ORGAN) for sequence generation models.” arXiv preprint arXiv:1705.10843 (2017).

2

Kim, Yoon. “Convolutional neural networks for sentence classification.” arXiv preprint arXiv:1408.5882 (2014).

__init__(n_tasks, char_dict, seq_length, n_embedding=75, kernel_sizes=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20], num_filters=[100, 200, 200, 200, 200, 100, 100, 100, 100, 100, 160, 160], dropout=0.25, mode='classification', **kwargs)[source]
Parameters
  • n_tasks (int) – Number of tasks

  • char_dict (dict) – Mapping from characters in smiles to integers

  • seq_length (int) – Length of sequences(after padding)

  • n_embedding (int, optional) – Length of embedding vector

  • filter_sizes (list of int, optional) – Properties of filters used in the conv net

  • num_filters (list of int, optional) – Properties of filters used in the conv net

  • dropout (float, optional) – Dropout rate

  • mode (str) – Either “classification” or “regression” for type of model.

static build_char_dict(dataset, default_dict={'#': 1, '(': 2, ')': 3, '+': 4, '-': 5, '/': 6, '1': 7, '2': 8, '3': 9, '4': 10, '5': 11, '6': 12, '7': 13, '8': 14, '=': 15, 'Br': 30, 'C': 16, 'Cl': 29, 'F': 17, 'H': 18, 'I': 19, 'N': 20, 'O': 21, 'P': 22, 'S': 23, '[': 24, '\\': 25, ']': 26, '_': 27, 'c': 28, 'n': 31, 'o': 32, 's': 33})[source]

Collect all unique characters(in smiles) from the dataset. This method should be called before defining the model to build appropriate char_dict

smiles_to_seq_batch(ids_b)[source]

Converts SMILES strings to np.array sequence.

A tf.py_func wrapper is written around this when creating the input_fn for make_estimator

default_generator(dataset, epochs=1, mode='fit', deterministic=True, pad_batches=True)[source]

Transfer smiles strings to fixed length integer vectors

smiles_to_seq(smiles)[source]

Tokenize characters in smiles to integers

AtomicConvModel

class AtomicConvModel(frag1_num_atoms=70, frag2_num_atoms=634, complex_num_atoms=701, max_num_neighbors=12, batch_size=24, atom_types=[6, 7.0, 8.0, 9.0, 11.0, 12.0, 15.0, 16.0, 17.0, 20.0, 25.0, 30.0, 35.0, 53.0, - 1.0], radial=[[1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5, 5.0, 5.5, 6.0, 6.5, 7.0, 7.5, 8.0, 8.5, 9.0, 9.5, 10.0, 10.5, 11.0, 11.5, 12.0], [0.0, 4.0, 8.0], [0.4]], layer_sizes=[32, 32, 16], learning_rate=0.001, **kwargs)[source]

Implements an Atomic Convolution Model.

Implements the atomic convolutional networks as introduced in

Gomes, Joseph, et al. “Atomic convolutional networks for predicting protein-ligand binding affinity.” arXiv preprint arXiv:1703.10603 (2017).

The atomic convolutional networks function as a variant of graph convolutions. The difference is that the “graph” here is the nearest neighbors graph in 3D space. The AtomicConvModel leverages these connections in 3D space to train models that learn to predict energetic state starting from the spatial geometry of the model.

__init__(frag1_num_atoms=70, frag2_num_atoms=634, complex_num_atoms=701, max_num_neighbors=12, batch_size=24, atom_types=[6, 7.0, 8.0, 9.0, 11.0, 12.0, 15.0, 16.0, 17.0, 20.0, 25.0, 30.0, 35.0, 53.0, - 1.0], radial=[[1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5, 5.0, 5.5, 6.0, 6.5, 7.0, 7.5, 8.0, 8.5, 9.0, 9.5, 10.0, 10.5, 11.0, 11.5, 12.0], [0.0, 4.0, 8.0], [0.4]], layer_sizes=[32, 32, 16], learning_rate=0.001, **kwargs)[source]
Parameters
  • frag1_num_atoms (int) – Number of atoms in first fragment

  • frag2_num_atoms (int) – Number of atoms in sec

  • max_num_neighbors (int) – Maximum number of neighbors possible for an atom. Recall neighbors are spatial neighbors.

  • atom_types (list) – List of atoms recognized by model. Atoms are indicated by their nuclear numbers.

  • radial (list) – TODO: add description

  • layer_sizes (list) – TODO: add description

  • learning_rate (float) – Learning rate for the model.

default_generator(dataset, epochs=1, mode='fit', deterministic=True, pad_batches=True)[source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

Smiles2Vec

class Smiles2Vec(char_to_idx, n_tasks=10, max_seq_len=270, embedding_dim=50, n_classes=2, use_bidir=True, use_conv=True, filters=192, kernel_size=3, strides=1, rnn_sizes=[224, 384], rnn_types=['GRU', 'GRU'], mode='regression', **kwargs)[source]

Implements the Smiles2Vec model, that learns neural representations of SMILES strings which can be used for downstream tasks.

The model is based on the description in Goh et al., “SMILES2vec: An Interpretable General-Purpose Deep Neural Network for Predicting Chemical Properties” (https://arxiv.org/pdf/1712.02034.pdf). The goal here is to take SMILES strings as inputs, turn them into vector representations which can then be used in predicting molecular properties.

The model consists of an Embedding layer that retrieves embeddings for each character in the SMILES string. These embeddings are learnt jointly with the rest of the model. The output from the embedding layer is a tensor of shape (batch_size, seq_len, embedding_dim). This tensor can optionally be fed through a 1D convolutional layer, before being passed to a series of RNN cells (optionally bidirectional). The final output from the RNN cells aims to have learnt the temporal dependencies in the SMILES string, and in turn information about the structure of the molecule, which is then used for molecular property prediction.

In the paper, the authors also train an explanation mask to endow the model with interpretability and gain insights into its decision making. This segment is currently not a part of this implementation as this was developed for the purpose of investigating a transfer learning protocol, ChemNet (which can be found at https://arxiv.org/abs/1712.02734).

__init__(char_to_idx, n_tasks=10, max_seq_len=270, embedding_dim=50, n_classes=2, use_bidir=True, use_conv=True, filters=192, kernel_size=3, strides=1, rnn_sizes=[224, 384], rnn_types=['GRU', 'GRU'], mode='regression', **kwargs)[source]
Parameters
  • char_to_idx (dict,) – char_to_idx contains character to index mapping for SMILES characters

  • embedding_dim (int, default 50) – Size of character embeddings used.

  • use_bidir (bool, default True) – Whether to use BiDirectional RNN Cells

  • use_conv (bool, default True) – Whether to use a conv-layer

  • kernel_size (int, default 3) – Kernel size for convolutions

  • filters (int, default 192) – Number of filters

  • strides (int, default 1) – Strides used in convolution

  • rnn_sizes (list[int], default [224, 384]) – Number of hidden units in the RNN cells

  • mode (str, default regression) – Whether to use model for regression or classification

default_generator(dataset, epochs=1, mode='fit', deterministic=True, pad_batches=True)[source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

ChemCeption

class ChemCeption(img_spec='std', base_filters=16, inception_blocks={'A': 3, 'B': 3, 'C': 3}, n_tasks=10, n_classes=2, augment=False, mode='regression', **kwargs)[source]

Implements the ChemCeption model that leverages the representational capacities of convolutional neural networks (CNNs) to predict molecular properties.

The model is based on the description in Goh et al., “Chemception: A Deep Neural Network with Minimal Chemistry Knowledge Matches the Performance of Expert-developed QSAR/QSPR Models” (https://arxiv.org/pdf/1706.06689.pdf). The authors use an image based representation of the molecule, where pixels encode different atomic and bond properties. More details on the image repres- entations can be found at https://arxiv.org/abs/1710.02238

The model consists of a Stem Layer that reduces the image resolution for the layers to follow. The output of the Stem Layer is followed by a series of Inception-Resnet blocks & a Reduction layer. Layers in the Inception-Resnet blocks process image tensors at multiple resolutions and use a ResNet style skip-connection, combining features from different resolutions. The Reduction layers reduce the spatial extent of the image by max-pooling and 2-strided convolutions. More details on these layers can be found in the ChemCeption paper referenced above. The output of the final Reduction layer is subject to a Global Average Pooling, and a fully-connected layer maps the features to downstream outputs.

In the ChemCeption paper, the authors perform real-time image augmentation by rotating images between 0 to 180 degrees. This can be done during model training by setting the augment argument to True.

__init__(img_spec='std', base_filters=16, inception_blocks={'A': 3, 'B': 3, 'C': 3}, n_tasks=10, n_classes=2, augment=False, mode='regression', **kwargs)[source]
Parameters
  • img_spec (str, default std) – Image specification used

  • base_filters (int, default 16) – Base filters used for the different inception and reduction layers

  • inception_blocks (dict,) – Dictionary containing number of blocks for every inception layer

  • n_tasks (int, default 10) – Number of classification or regression tasks

  • n_classes (int, default 2) – Number of classes (used only for classification)

  • augment (bool, default False) – Whether to augment images

  • mode (str, default regression) – Whether the model is used for regression or classification

build_inception_module(inputs, type='A')[source]

Inception module is a series of inception layers of similar type. This function builds that.

default_generator(dataset, epochs=1, mode='fit', deterministic=True, pad_batches=True)[source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

NormalizingFlowModel

The purpose of a normalizing flow is to map a simple distribution (that is easy to sample from and evaluate probability densities for) to a more complex distribution that is learned from data. Normalizing flows combine the advantages of autoregressive models (which provide likelihood estimation but do not learn features) and variational autoencoders (which learn feature representations but do not provide marginal likelihoods). They are effective for any application requiring a probabilistic model with these capabilities, e.g. generative modeling, unsupervised learning, or probabilistic inference.

class NormalizingFlowModel(model: deepchem.models.normalizing_flows.NormalizingFlow, **kwargs)[source]

A base distribution and normalizing flow for applying transformations.

Normalizing flows are effective for any application requiring a probabilistic model that can both sample from a distribution and compute marginal likelihoods, e.g. generative modeling, unsupervised learning, or probabilistic inference. For a thorough review of normalizing flows, see [1]_.

A distribution implements two main operations:
  1. Sampling from the transformed distribution

  2. Calculating log probabilities

A normalizing flow implements three main operations:
  1. Forward transformation

  2. Inverse transformation

  3. Calculating the Jacobian

Deep Normalizing Flow models require normalizing flow layers where input and output dimensions are the same, the transformation is invertible, and the determinant of the Jacobian is efficient to compute and differentiable. The determinant of the Jacobian of the transformation gives a factor that preserves the probability volume to 1 when transforming between probability densities of different random variables.

References

1

Papamakarios, George et al. “Normalizing Flows for Probabilistic Modeling and Inference.” (2019). https://arxiv.org/abs/1912.02762.

__init__(model: deepchem.models.normalizing_flows.NormalizingFlow, **kwargs) → None[source]

Creates a new NormalizingFlowModel.

In addition to the following arguments, this class also accepts all the keyword arguments from KerasModel.

Parameters

model (NormalizingFlow) – An instance of NormalizingFlow.

Examples

>> import tensorflow_probability as tfp >> tfd = tfp.distributions >> tfb = tfp.bijectors >> flow_layers = [ .. tfb.RealNVP( .. num_masked=2, .. shift_and_log_scale_fn=tfb.real_nvp_default_template( .. hidden_layers=[8, 8])) ..] >> base_distribution = tfd.MultivariateNormalDiag(loc=[0., 0., 0.]) >> nf = NormalizingFlow(base_distribution, flow_layers) >> nfm = NormalizingFlowModel(nf) >> dataset = NumpyDataset( .. X=np.random.rand(5, 3).astype(np.float32), .. y=np.random.rand(5,), .. ids=np.arange(5)) >> nfm.fit(dataset)

create_nll(input: Union[tensorflow.python.framework.ops.Tensor, Sequence[tensorflow.python.framework.ops.Tensor]]) → tensorflow.python.framework.ops.Tensor[source]

Create the negative log likelihood loss function.

The default implementation is appropriate for most cases. Subclasses can override this if there is a need to customize it.

Parameters

input (OneOrMany[tf.Tensor]) – A batch of data.

Returns

Return type

A Tensor equal to the loss function to use for optimization.

save()[source]

Saves model to disk using joblib.

reload()[source]

Loads model from joblib file on disk.

PyTorch Models

DeepChem supports the use of PyTorch to build deep learning models.

TorchModel

You can wrap an arbitrary torch.nn.Module in a TorchModel object.

class TorchModel(model: torch.nn.modules.module.Module, loss: Union[deepchem.models.losses.Loss, Callable[[List, List, List], Any]], output_types: Optional[List[str]] = None, batch_size: int = 100, model_dir: Optional[str] = None, learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule] = 0.001, optimizer: Optional[deepchem.models.optimizers.Optimizer] = None, tensorboard: bool = False, wandb: bool = False, log_frequency: int = 100, device: Optional[torch.device] = None, **kwargs)[source]

This is a DeepChem model implemented by a PyTorch model.

Here is a simple example of code that uses TorchModel to train a PyTorch model on a DeepChem dataset.

>> pytorch_model = torch.nn.Sequential( >> torch.nn.Linear(100, 1000), >> torch.nn.Tanh(), >> torch.nn.Linear(1000, 1)) >> model = TorchModel(pytorch_model, loss=dc.models.losses.L2Loss()) >> model.fit(dataset)

The loss function for a model can be defined in two different ways. For models that have only a single output and use a standard loss function, you can simply provide a dc.models.losses.Loss object. This defines the loss for each sample or sample/task pair. The result is automatically multiplied by the weights and averaged over the batch.

For more complicated cases, you can instead provide a function that directly computes the total loss. It must be of the form f(outputs, labels, weights), taking the list of outputs from the model, the expected values, and any weight matrices. It should return a scalar equal to the value of the loss function for the batch. No additional processing is done to the result; it is up to you to do any weighting, averaging, adding of penalty terms, etc.

You can optionally provide an output_types argument, which describes how to interpret the model’s outputs. This should be a list of strings, one for each output. You can use an arbitrary output_type for a output, but some output_types are special and will undergo extra processing:

  • ‘prediction’: This is a normal output, and will be returned by predict(). If output types are not specified, all outputs are assumed to be of this type.

  • ‘loss’: This output will be used in place of the normal outputs for computing the loss function. For example, models that output probability distributions usually do it by computing unbounded numbers (the logits), then passing them through a softmax function to turn them into probabilities. When computing the cross entropy, it is more numerically stable to use the logits directly rather than the probabilities. You can do this by having the model produce both probabilities and logits as outputs, then specifying output_types=[‘prediction’, ‘loss’]. When predict() is called, only the first output (the probabilities) will be returned. But during training, it is the second output (the logits) that will be passed to the loss function.

  • ‘variance’: This output is used for estimating the uncertainty in another output. To create a model that can estimate uncertainty, there must be the same number of ‘prediction’ and ‘variance’ outputs. Each variance output must have the same shape as the corresponding prediction output, and each element is an estimate of the variance in the corresponding prediction. Also be aware that if a model supports uncertainty, it MUST use dropout on every layer, and dropout most be enabled during uncertainty prediction. Otherwise, the uncertainties it computes will be inaccurate.

  • other: Arbitrary output_types can be used to extract outputs produced by the model, but will have no additional processing performed.

__init__(model: torch.nn.modules.module.Module, loss: Union[deepchem.models.losses.Loss, Callable[[List, List, List], Any]], output_types: Optional[List[str]] = None, batch_size: int = 100, model_dir: Optional[str] = None, learning_rate: Union[float, deepchem.models.optimizers.LearningRateSchedule] = 0.001, optimizer: Optional[deepchem.models.optimizers.Optimizer] = None, tensorboard: bool = False, wandb: bool = False, log_frequency: int = 100, device: Optional[torch.device] = None, **kwargs) → None[source]

Create a new TorchModel.

Parameters
  • model (torch.nn.Module) – the PyTorch model implementing the calculation

  • loss (dc.models.losses.Loss or function) – a Loss or function defining how to compute the training loss for each batch, as described above

  • output_types (list of strings, optional (default None)) – the type of each output from the model, as described above

  • batch_size (int, optional (default 100)) – default batch size for training and evaluating

  • model_dir (str, optional (default None)) – the directory on disk where the model will be stored. If this is None, a temporary directory is created.

  • learning_rate (float or LearningRateSchedule, optional (default 0.001)) – the learning rate to use for fitting. If optimizer is specified, this is ignored.

  • optimizer (Optimizer, optional (default None)) – the optimizer to use for fitting. If this is specified, learning_rate is ignored.

  • tensorboard (bool, optional (default False)) – whether to log progress to TensorBoard during training

  • wandb (bool, optional (default False)) – whether to log progress to Weights & Biases during training

  • log_frequency (int, optional (default 100)) – The frequency at which to log data. Data is logged using logging by default. If tensorboard is set, data is also logged to TensorBoard. If wandb is set, data is also logged to Weights & Biases. Logging happens at global steps. Roughly, a global step corresponds to one batch of training. If you’d like a printout every 10 batch steps, you’d set log_frequency=10 for example.

  • device (torch.device, optional (default None)) – the device on which to run computations. If None, a device is chosen automatically.

fit(dataset: deepchem.data.datasets.Dataset, nb_epoch: int = 10, max_checkpoints_to_keep: int = 5, checkpoint_interval: int = 1000, deterministic: bool = False, restore: bool = False, variables: Optional[List[torch.nn.parameter.Parameter]] = None, loss: Optional[Callable[[List, List, List], Any]] = None, callbacks: Union[Callable, List[Callable]] = [], all_losses: Optional[List[float]] = None) → float[source]

Train this model on a dataset.

Parameters
  • dataset (Dataset) – the Dataset to train on

  • nb_epoch (int) – the number of epochs to train for

  • max_checkpoints_to_keep (int) – the maximum number of checkpoints to keep. Older checkpoints are discarded.

  • checkpoint_interval (int) – the frequency at which to write checkpoints, measured in training steps. Set this to 0 to disable automatic checkpointing.

  • deterministic (bool) – if True, the samples are processed in order. If False, a different random order is used for each epoch.

  • restore (bool) – if True, restore the model from the most recent checkpoint and continue training from there. If False, retrain the model from scratch.

  • variables (list of torch.nn.Parameter) – the variables to train. If None (the default), all trainable variables in the model are used.

  • loss (function) – a function of the form f(outputs, labels, weights) that computes the loss for each batch. If None (the default), the model’s standard loss function is used.

  • callbacks (function or list of functions) – one or more functions of the form f(model, step) that will be invoked after every step. This can be used to perform validation, logging, etc.

  • all_losses (Optional[List[float]], optional (default None)) – If specified, all logged losses are appended into this list. Note that you can call fit() repeatedly with the same list and losses will continue to be appended.

Returns

Return type

The average loss over the most recent checkpoint interval

fit_generator(generator: Iterable[Tuple[Any, Any, Any]], max_checkpoints_to_keep: int = 5, checkpoint_interval: int = 1000, restore: bool = False, variables: Optional[List[torch.nn.parameter.Parameter]] = None, loss: Optional[Callable[[List, List, List], Any]] = None, callbacks: Union[Callable, List[Callable]] = [], all_losses: Optional[List[float]] = None) → float[source]

Train this model on data from a generator.

Parameters
  • generator (generator) – this should generate batches, each represented as a tuple of the form (inputs, labels, weights).

  • max_checkpoints_to_keep (int) – the maximum number of checkpoints to keep. Older checkpoints are discarded.

  • checkpoint_interval (int) – the frequency at which to write checkpoints, measured in training steps. Set this to 0 to disable automatic checkpointing.

  • restore (bool) – if True, restore the model from the most recent checkpoint and continue training from there. If False, retrain the model from scratch.

  • variables (list of torch.nn.Parameter) – the variables to train. If None (the default), all trainable variables in the model are used.

  • loss (function) – a function of the form f(outputs, labels, weights) that computes the loss for each batch. If None (the default), the model’s standard loss function is used.

  • callbacks (function or list of functions) – one or more functions of the form f(model, step) that will be invoked after every step. This can be used to perform validation, logging, etc.

  • all_losses (Optional[List[float]], optional (default None)) – If specified, all logged losses are appended into this list. Note that you can call fit() repeatedly with the same list and losses will continue to be appended.

Returns

Return type

The average loss over the most recent checkpoint interval

fit_on_batch(X: Sequence, y: Sequence, w: Sequence, variables: Optional[List[torch.nn.parameter.Parameter]] = None, loss: Optional[Callable[[List, List, List], Any]] = None, callbacks: Union[Callable, List[Callable]] = [], checkpoint: bool = True, max_checkpoints_to_keep: int = 5) → float[source]

Perform a single step of training.

Parameters
  • X (ndarray) – the inputs for the batch

  • y (ndarray) – the labels for the batch

  • w (ndarray) – the weights for the batch

  • variables (list of torch.nn.Parameter) – the variables to train. If None (the default), all trainable variables in the model are used.

  • loss (function) – a function of the form f(outputs, labels, weights) that computes the loss for each batch. If None (the default), the model’s standard loss function is used.

  • callbacks (function or list of functions) – one or more functions of the form f(model, step) that will be invoked after every step. This can be used to perform validation, logging, etc.

  • checkpoint (bool) – if true, save a checkpoint after performing the training step

  • max_checkpoints_to_keep (int) – the maximum number of checkpoints to keep. Older checkpoints are discarded.

Returns

Return type

the loss on the batch

predict_on_generator(generator: Iterable[Tuple[Any, Any, Any]], transformers: List[transformers.Transformer] = [], output_types: Optional[Union[str, Sequence[str]]] = None) → Union[numpy.ndarray, Sequence[numpy.ndarray]][source]
Parameters
  • generator (generator) – this should generate batches, each represented as a tuple of the form (inputs, labels, weights).

  • transformers (list of dc.trans.Transformers) – Transformers that the input data has been transformed by. The output is passed through these transformers to undo the transformations.

  • output_types (String or list of Strings) – If specified, all outputs of this type will be retrieved from the model. If output_types is specified, outputs must be None.

  • Returns – a NumPy array of the model produces a single output, or a list of arrays if it produces multiple outputs

predict_on_batch(X: Sequence, transformers: List[transformers.Transformer] = []) → Union[numpy.ndarray, Sequence[numpy.ndarray]][source]

Generates predictions for input samples, processing samples in a batch.

Parameters
  • X (ndarray) – the input data, as a Numpy array.

  • transformers (list of dc.trans.Transformers) – Transformers that the input data has been transformed by. The output is passed through these transformers to undo the transformations.

Returns

  • a NumPy array of the model produces a single output, or a list of arrays

  • if it produces multiple outputs

predict_uncertainty_on_batch(X: Sequence, masks: int = 50) → Union[Tuple[numpy.ndarray, numpy.ndarray], Sequence[Tuple[numpy.ndarray, numpy.ndarray]]][source]

Predict the model’s outputs, along with the uncertainty in each one.

The uncertainty is computed as described in https://arxiv.org/abs/1703.04977. It involves repeating the prediction many times with different dropout masks. The prediction is computed as the average over all the predictions. The uncertainty includes both the variation among the predicted values (epistemic uncertainty) and the model’s own estimates for how well it fits the data (aleatoric uncertainty). Not all models support uncertainty prediction.

Parameters
  • X (ndarray) – the input data, as a Numpy array.

  • masks (int) – the number of dropout masks to average over

Returns

  • for each output, a tuple (y_pred, y_std) where y_pred is the predicted

  • value of the output, and each element of y_std estimates the standard

  • deviation of the corresponding element of y_pred

predict(dataset: deepchem.data.datasets.Dataset, transformers: List[transformers.Transformer] = [], output_types: Optional[List[str]] = None) → Union[numpy.ndarray, Sequence[numpy.ndarray]][source]

Uses self to make predictions on provided Dataset object.

Parameters
  • dataset (dc.data.Dataset) – Dataset to make prediction on

  • transformers (list of dc.trans.Transformers) – Transformers that the input data has been transformed by. The output is passed through these transformers to undo the transformations.

  • output_types (String or list of Strings) – If specified, all outputs of this type will be retrieved from the model. If output_types is specified, outputs must be None.

Returns

  • a NumPy array of the model produces a single output, or a list of arrays

  • if it produces multiple outputs

predict_embedding(dataset: deepchem.data.datasets.Dataset) → Union[numpy.ndarray, Sequence[numpy.ndarray]][source]

Predicts embeddings created by underlying model if any exist. An embedding must be specified to have output_type of ‘embedding’ in the model definition.

Parameters

dataset (dc.data.Dataset) – Dataset to make prediction on

Returns

  • a NumPy array of the embeddings model produces, or a list

  • of arrays if it produces multiple embeddings

predict_uncertainty(dataset: deepchem.data.datasets.Dataset, masks: int = 50) → Union[Tuple[numpy.ndarray, numpy.ndarray], Sequence[Tuple[numpy.ndarray, numpy.ndarray]]][source]

Predict the model’s outputs, along with the uncertainty in each one.

The uncertainty is computed as described in https://arxiv.org/abs/1703.04977. It involves repeating the prediction many times with different dropout masks. The prediction is computed as the average over all the predictions. The uncertainty includes both the variation among the predicted values (epistemic uncertainty) and the model’s own estimates for how well it fits the data (aleatoric uncertainty). Not all models support uncertainty prediction.

Parameters
  • dataset (dc.data.Dataset) – Dataset to make prediction on

  • masks (int) – the number of dropout masks to average over

Returns

  • for each output, a tuple (y_pred, y_std) where y_pred is the predicted

  • value of the output, and each element of y_std estimates the standard

  • deviation of the corresponding element of y_pred

evaluate_generator(generator: Iterable[Tuple[Any, Any, Any]], metrics: List[deepchem.metrics.metric.Metric], transformers: List[transformers.Transformer] = [], per_task_metrics: bool = False)[source]

Evaluate the performance of this model on the data produced by a generator.

Parameters
  • generator (generator) – this should generate batches, each represented as a tuple of the form (inputs, labels, weights).

  • metric (list of deepchem.metrics.Metric) – Evaluation metric

  • transformers (list of dc.trans.Transformers) – Transformers that the input data has been transformed by. The output is passed through these transformers to undo the transformations.

  • per_task_metrics (bool) – If True, return per-task scores.

Returns

Maps tasks to scores under metric.

Return type

dict

compute_saliency(X: numpy.ndarray) → Union[numpy.ndarray, Sequence[numpy.ndarray]][source]

Compute the saliency map for an input sample.

This computes the Jacobian matrix with the derivative of each output element with respect to each input element. More precisely,

  • If this model has a single output, it returns a matrix of shape (output_shape, input_shape) with the derivatives.

  • If this model has multiple outputs, it returns a list of matrices, one for each output.

This method cannot be used on models that take multiple inputs.

Parameters

X (ndarray) – the input data for a single sample

Returns

Return type

the Jacobian matrix, or a list of matrices

default_generator(dataset: deepchem.data.datasets.Dataset, epochs: int = 1, mode: str = 'fit', deterministic: bool = True, pad_batches: bool = True) → Iterable[Tuple[List, List, List]][source]

Create a generator that iterates batches for a dataset.

Subclasses may override this method to customize how model inputs are generated from the data.

Parameters
  • dataset (Dataset) – the data to iterate

  • epochs (int) – the number of times to iterate over the full dataset

  • mode (str) – allowed values are ‘fit’ (called during training), ‘predict’ (called during prediction), and ‘uncertainty’ (called during uncertainty prediction)

  • deterministic (bool) – whether to iterate over the dataset in order, or randomly shuffle the data for each epoch

  • pad_batches (bool) – whether to pad each batch up to this model’s preferred batch size

Returns

  • a generator that iterates batches, each represented as a tuple of lists

  • ([inputs], [outputs], [weights])

save_checkpoint(max_checkpoints_to_keep: int = 5, model_dir: Optional[str] = None) → None[source]

Save a checkpoint to disk.

Usually you do not need to call this method, since fit() saves checkpoints automatically. If you have disabled automatic checkpointing during fitting, this can be called to manually write checkpoints.

Parameters
  • max_checkpoints_to_keep (int) – the maximum number of checkpoints to keep. Older checkpoints are discarded.

  • model_dir (str, default None) – Model directory to save checkpoint to. If None, revert to self.model_dir

get_checkpoints(model_dir: Optional[str] = None)[source]

Get a list of all available checkpoint files.

Parameters

model_dir (str, default None) – Directory to get list of checkpoints from. Reverts to self.model_dir if None

restore(checkpoint: Optional[str] = None, model_dir: Optional[str] = None) → None[source]

Reload the values of all variables from a checkpoint file.

Parameters
  • checkpoint (str) – the path to the checkpoint file to load. If this is None, the most recent checkpoint will be chosen automatically. Call get_checkpoints() to get a list of all available checkpoints.

  • model_dir (str, default None) – Directory to restore checkpoint from. If None, use self.model_dir. If checkpoint is not None, this is ignored.

get_global_step() → int[source]

Get the number of steps of fitting that have been performed.

load_from_pretrained(source_model: deepchem.models.torch_models.torch_model.TorchModel, assignment_map: Optional[Dict[Any, Any]] = None, value_map: Optional[Dict[Any, Any]] = None, checkpoint: Optional[str] = None, model_dir: Optional[str] = None, include_top: bool = True, inputs: Optional[Sequence[Any]] = None, **kwargs) → None[source]

Copies parameter values from a pretrained model. source_model can either be a pretrained model or a model with the same architecture. value_map is a parameter-value dictionary. If no value_map is provided, the parameter values are restored to the source_model from a checkpoint and a default value_map is created. assignment_map is a dictionary mapping parameters from the source_model to the current model. If no assignment_map is provided, one is made from scratch and assumes the model is composed of several different layers, with the final one being a dense layer. include_top is used to control whether or not the final dense layer is used. The default assignment map is useful in cases where the type of task is different (classification vs regression) and/or number of tasks in the setting.

Parameters
  • source_model (dc.TorchModel, required) – source_model can either be the pretrained model or a dc.TorchModel with the same architecture as the pretrained model. It is used to restore from a checkpoint, if value_map is None and to create a default assignment map if assignment_map is None

  • assignment_map (Dict, default None) – Dictionary mapping the source_model parameters and current model parameters

  • value_map (Dict, default None) – Dictionary containing source_model trainable parameters mapped to numpy arrays. If value_map is None, the values are restored and a default parameter map is created using the restored values

  • checkpoint (str, default None) – the path to the checkpoint file to load. If this is None, the most recent checkpoint will be chosen automatically. Call get_checkpoints() to get a list of all available checkpoints

  • model_dir (str, default None) – Restore model from custom model directory if needed

  • include_top (bool, default True) – if True, copies the weights and bias associated with the final dense layer. Used only when assignment map is None

  • inputs (List, input tensors for model) – if not None, then the weights are built for both the source and self.

CGCNNModel

class CGCNNModel(in_node_dim: int = 92, hidden_node_dim: int = 64, in_edge_dim: int = 41, num_conv: int = 3, predictor_hidden_feats: int = 128, n_tasks: int = 1, mode: str = 'regression', n_classes: int = 2, **kwargs)[source]

Crystal Graph Convolutional Neural Network (CGCNN).

Here is a simple example of code that uses the CGCNNModel with materials dataset.

>> import deepchem as dc >> dataset_config = {“reload”: False, “featurizer”: dc.feat.CGCNNFeaturizer, “transformers”: []} >> tasks, datasets, transformers = dc.molnet.load_perovskite(**dataset_config) >> train, valid, test = datasets >> model = dc.models.CGCNNModel(mode=’regression’, batch_size=32, learning_rate=0.001) >> model.fit(train, nb_epoch=50)

This model takes arbitary crystal structures as an input, and predict material properties using the element information and connection of atoms in the crystal. If you want to get some material properties which has a high computational cost like band gap in the case of DFT, this model may be useful. This model is one of variants of Graph Convolutional Networks. The main differences between other GCN models are how to construct graphs and how to update node representations. This model defines the crystal graph from structures using distances between atoms. The crystal graph is an undirected multigraph which is defined by nodes representing atom properties and edges representing connections between atoms in a crystal. And, this model updates the node representations using both neighbor node and edge representations. Please confirm the detail algorithms from [1]_.

References

1

Xie, Tian, and Jeffrey C. Grossman. “Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties.” Physical review letters 120.14 (2018): 145301.

Notes

This class requires DGL and PyTorch to be installed.

__init__(in_node_dim: int = 92, hidden_node_dim: int = 64, in_edge_dim: int = 41, num_conv: int = 3, predictor_hidden_feats: int = 128, n_tasks: int = 1, mode: str = 'regression', n_classes: int = 2, **kwargs)[source]

This class accepts all the keyword arguments from TorchModel.

Parameters
  • in_node_dim (int, default 92) – The length of the initial node feature vectors. The 92 is based on length of vectors in the atom_init.json.

  • hidden_node_dim (int, default 64) – The length of the hidden node feature vectors.

  • in_edge_dim (int, default 41) – The length of the initial edge feature vectors. The 41 is based on default setting of CGCNNFeaturizer.

  • num_conv (int, default 3) – The number of convolutional layers.

  • predictor_hidden_feats (int, default 128) – The size for hidden representations in the output MLP predictor.

  • n_tasks (int, default 1) – The number of the output size.

  • mode (str, default 'regression') – The model type, ‘classification’ or ‘regression’.

  • n_classes (int, default 2) – The number of classes to predict (only used in classification mode).

  • kwargs (Dict) – This class accepts all the keyword arguments from TorchModel.

GATModel

class GATModel(n_tasks: int, graph_attention_layers: Optional[list] = None, n_attention_heads: int = 8, agg_modes: Optional[list] = None, activation=<function elu>, residual: bool = True, dropout: float = 0.0, alpha: float = 0.2, predictor_hidden_feats: int = 128, predictor_dropout: float = 0.0, mode: str = 'regression', number_atom_features: int = 30, n_classes: int = 2, self_loop: bool = True, **kwargs)[source]

Model for Graph Property Prediction Based on Graph Attention Networks (GAT).

This model proceeds as follows:

  • Update node representations in graphs with a variant of GAT

  • For each graph, compute its representation by 1) a weighted sum of the node representations in the graph, where the weights are computed by applying a gating function to the node representations 2) a max pooling of the node representations 3) concatenating the output of 1) and 2)

  • Perform the final prediction using an MLP

Examples

>>>
>> import deepchem as dc
>> from deepchem.models import GATModel
>> featurizer = dc.feat.MolGraphConvFeaturizer()
>> tasks, datasets, transformers = dc.molnet.load_tox21(
..     reload=False, featurizer=featurizer, transformers=[])
>> train, valid, test = datasets
>> model = GATModel(mode='classification', n_tasks=len(tasks),
..                  batch_size=32, learning_rate=0.001)
>> model.fit(train, nb_epoch=50)

References

1

Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. “Graph Attention Networks.” ICLR 2018.

Notes

This class requires DGL (https://github.com/dmlc/dgl) and DGL-LifeSci (https://github.com/awslabs/dgl-lifesci) to be installed.

__init__(n_tasks: int, graph_attention_layers: Optional[list] = None, n_attention_heads: int = 8, agg_modes: Optional[list] = None, activation=<function elu>, residual: bool = True, dropout: float = 0.0, alpha: float = 0.2, predictor_hidden_feats: int = 128, predictor_dropout: float = 0.0, mode: str = 'regression', number_atom_features: int = 30, n_classes: int = 2, self_loop: bool = True, **kwargs)[source]
Parameters
  • n_tasks (int) – Number of tasks.

  • graph_attention_layers (list of int) – Width of channels per attention head for GAT layers. graph_attention_layers[i] gives the width of channel for each attention head for the i-th GAT layer. If both graph_attention_layers and agg_modes are specified, they should have equal length. If not specified, the default value will be [8, 8].

  • n_attention_heads (int) – Number of attention heads in each GAT layer.

  • agg_modes (list of str) – The way to aggregate multi-head attention results for each GAT layer, which can be either ‘flatten’ for concatenating all-head results or ‘mean’ for averaging all-head results. agg_modes[i] gives the way to aggregate multi-head attention results for the i-th GAT layer. If both graph_attention_layers and agg_modes are specified, they should have equal length. If not specified, the model will flatten multi-head results for intermediate GAT layers and compute mean of multi-head results for the last GAT layer.

  • activation (activation function or None) – The activation function to apply to the aggregated multi-head results for each GAT layer. If not specified, the default value will be ELU.

  • residual (bool) – Whether to add a residual connection within each GAT layer. Default to True.

  • dropout (float) – The dropout probability within each GAT layer. Default to 0.

  • alpha (float) – A hyperparameter in LeakyReLU, which is the slope for negative values. Default to 0.2.

  • predictor_hidden_feats (int) – The size for hidden representations in the output MLP predictor. Default to 128.

  • predictor_dropout (float) – The dropout probability in the output MLP predictor. Default to 0.

  • mode (str) – The model type, ‘classification’ or ‘regression’. Default to ‘regression’.

  • number_atom_features (int) – The length of the initial atom feature vectors. Default to 30.

  • n_classes (int) – The number of classes to predict per task (only used when mode is ‘classification’). Default to 2.

  • self_loop (bool) – Whether to add self loops for the nodes, i.e. edges from nodes to themselves. When input graphs have isolated nodes, self loops allow preserving the original feature of them in message passing. Default to True.

  • kwargs – This can include any keyword argument of TorchModel.

GCNModel

class GCNModel(n_tasks: int, graph_conv_layers: Optional[list] = None, activation=None, residual: bool = True, batchnorm: bool = False, dropout: float = 0.0, predictor_hidden_feats: int = 128, predictor_dropout: float = 0.0, mode: str = 'regression', number_atom_features=30, n_classes: int = 2, self_loop: bool = True, **kwargs)[source]

Model for Graph Property Prediction Based on Graph Convolution Networks (GCN).

This model proceeds as follows:

  • Update node representations in graphs with a variant of GCN

  • For each graph, compute its representation by 1) a weighted sum of the node representations in the graph, where the weights are computed by applying a gating function to the node representations 2) a max pooling of the node representations 3) concatenating the output of 1) and 2)

  • Perform the final prediction using an MLP

Examples

>>>
>> import deepchem as dc
>> from deepchem.models import GCNModel
>> featurizer = dc.feat.MolGraphConvFeaturizer()
>> tasks, datasets, transformers = dc.molnet.load_tox21(
..     reload=False, featurizer=featurizer, transformers=[])
>> train, valid, test = datasets
>> model = GCNModel(mode='classification', n_tasks=len(tasks),
..                  batch_size=32, learning_rate=0.001)
>> model.fit(train, nb_epoch=50)

References

1

Thomas N. Kipf and Max Welling. “Semi-Supervised Classification with Graph Convolutional Networks.” ICLR 2017.

Notes

This class requires DGL (https://github.com/dmlc/dgl) and DGL-LifeSci (https://github.com/awslabs/dgl-lifesci) to be installed.

This model is different from deepchem.models.GraphConvModel as follows:

  • For each graph convolution, the learnable weight in this model is shared across all nodes. GraphConvModel employs separate learnable weights for nodes of different degrees. A learnable weight is shared across all nodes of a particular degree.

  • For GraphConvModel, there is an additional GraphPool operation after each graph convolution. The operation updates the representation of a node by applying an element-wise maximum over the representations of its neighbors and itself.

  • For computing graph-level representations, this model computes a weighted sum and an element-wise maximum of the representations of all nodes in a graph and concatenates them. The node weights are obtained by using a linear/dense layer followd by a sigmoid function. For GraphConvModel, the sum over node representations is unweighted.

  • There are various minor differences in using dropout, skip connection and batch normalization.

__init__(n_tasks: int, graph_conv_layers: Optional[list] = None, activation=None, residual: bool = True, batchnorm: bool = False, dropout: float = 0.0, predictor_hidden_feats: int = 128, predictor_dropout: float = 0.0, mode: str = 'regression', number_atom_features=30, n_classes: int = 2, self_loop: bool = True, **kwargs)[source]
Parameters
  • n_tasks (int) – Number of tasks.

  • graph_conv_layers (list of int) – Width of channels for GCN layers. graph_conv_layers[i] gives the width of channel for the i-th GCN layer. If not specified, the default value will be [64, 64].

  • activation (callable) – The activation function to apply to the output of each GCN layer. By default, no activation function will be applied.

  • residual (bool) – Whether to add a residual connection within each GCN layer. Default to True.

  • batchnorm (bool) – Whether to apply batch normalization to the output of each GCN layer. Default to False.

  • dropout (float) – The dropout probability for the output of each GCN layer. Default to 0.

  • predictor_hidden_feats (int) – The size for hidden representations in the output MLP predictor. Default to 128.

  • predictor_dropout (float) – The dropout probability in the output MLP predictor. Default to 0.

  • mode (str) – The model type, ‘classification’ or ‘regression’. Default to ‘regression’.

  • number_atom_features (int) – The length of the initial atom feature vectors. Default to 30.

  • n_classes (int) – The number of classes to predict per task (only used when mode is ‘classification’). Default to 2.

  • self_loop (bool) – Whether to add self loops for the nodes, i.e. edges from nodes to themselves. When input graphs have isolated nodes, self loops allow preserving the original feature of them in message passing. Default to True.

  • kwargs – This can include any keyword argument of TorchModel.

AttentiveFPModel

class AttentiveFPModel(n_tasks: int, num_layers: int = 2, num_timesteps: int = 2, graph_feat_size: int = 200, dropout: float = 0.0, mode: str = 'regression', number_atom_features: int = 30, number_bond_features: int = 11, n_classes: int = 2, self_loop: bool = True, **kwargs)[source]

Model for Graph Property Prediction.

This model proceeds as follows:

  • Combine node features and edge features for initializing node representations, which involves a round of message passing

  • Update node representations with multiple rounds of message passing

  • For each graph, compute its representation by combining the representations of all nodes in it, which involves a gated recurrent unit (GRU).

  • Perform the final prediction using a linear layer

Examples

>>>
>> import deepchem as dc
>> from deepchem.models import AttentiveFPModel
>> featurizer = dc.feat.MolGraphConvFeaturizer(use_edges=True)
>> tasks, datasets, transformers = dc.molnet.load_tox21(
..     reload=False, featurizer=featurizer, transformers=[])
>> train, valid, test = datasets
>> model = AttentiveFPModel(mode='classification', n_tasks=len(tasks),
..                          batch_size=32, learning_rate=0.001)
>> model.fit(train, nb_epoch=50)

References

1

Zhaoping Xiong, Dingyan Wang, Xiaohong Liu, Feisheng Zhong, Xiaozhe Wan, Xutong Li, Zhaojun Li, Xiaomin Luo, Kaixian Chen, Hualiang Jiang, and Mingyue Zheng. “Pushing the Boundaries of Molecular Representation for Drug Discovery with the Graph Attention Mechanism.” Journal of Medicinal Chemistry. 2020, 63, 16, 8749–8760.

Notes

This class requires DGL (https://github.com/dmlc/dgl) and DGL-LifeSci (https://github.com/awslabs/dgl-lifesci) to be installed.

__init__(n_tasks: int, num_layers: int = 2, num_timesteps: int = 2, graph_feat_size: int = 200, dropout: float = 0.0, mode: str = 'regression', number_atom_features: int = 30, number_bond_features: int = 11, n_classes: int = 2, self_loop: bool = True, **kwargs)[source]
Parameters
  • n_tasks (int) – Number of tasks.

  • num_layers (int) – Number of graph neural network layers, i.e. number of rounds of message passing. Default to 2.

  • num_timesteps (int) – Number of time steps for updating graph representations with a GRU. Default to 2.

  • graph_feat_size (int) – Size for graph representations. Default to 200.

  • dropout (float) – Dropout probability. Default to 0.

  • mode (str) – The model type, ‘classification’ or ‘regression’. Default to ‘regression’.

  • number_atom_features (int) – The length of the initial atom feature vectors. Default to 30.

  • number_bond_features (int) – The length of the initial bond feature vectors. Default to 11.

  • n_classes (int) – The number of classes to predict per task (only used when mode is ‘classification’). Default to 2.

  • self_loop (bool) – Whether to add self loops for the nodes, i.e. edges from nodes to themselves. When input graphs have isolated nodes, self loops allow preserving the original feature of them in message passing. Default to True.

  • kwargs – This can include any keyword argument of TorchModel.

MPNNModel

Note that this is an alternative implementation for MPNN and currently you can only import it from deepchem.models.torch_models.

class MPNNModel(n_tasks: int, node_out_feats: int = 64, edge_hidden_feats: int = 128, num_step_message_passing: int = 3, num_step_set2set: int = 6, num_layer_set2set: int = 3, mode: str = 'regression', number_atom_features: int = 30, number_bond_features: int = 11, n_classes: int = 2, self_loop: bool = False, **kwargs)[source]

Model for graph property prediction

This model proceeds as follows:

  • Combine latest node representations and edge features in updating node representations, which involves multiple rounds of message passing

  • For each graph, compute its representation by combining the representations of all nodes in it, which involves a Set2Set layer.

  • Perform the final prediction using an MLP

Examples

>>>
>> import deepchem as dc
>> from deepchem.models.torch_models import MPNNModel
>> featurizer = dc.feat.MolGraphConvFeaturizer(use_edges=True)
>> tasks, datasets, transformers = dc.molnet.load_tox21(
..     reload=False, featurizer=featurizer, transformers=[])
>> train, valid, test = datasets
>> model = MPNNModel(mode='classification', n_tasks=len(tasks),
..                   batch_size=32, learning_rate=0.001)
>> model.fit(train, nb_epoch=50)

References

1

Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, George E. Dahl. “Neural Message Passing for Quantum Chemistry.” ICML 2017.

Notes

This class requires DGL (https://github.com/dmlc/dgl) and DGL-LifeSci (https://github.com/awslabs/dgl-lifesci) to be installed.

__init__(n_tasks: int, node_out_feats: int = 64, edge_hidden_feats: int = 128, num_step_message_passing: int = 3, num_step_set2set: int = 6, num_layer_set2set: int = 3, mode: str = 'regression', number_atom_features: int = 30, number_bond_features: int = 11, n_classes: int = 2, self_loop: bool = False, **kwargs)[source]
Parameters
  • n_tasks (int) – Number of tasks.

  • node_out_feats (int) – The length of the final node representation vectors. Default to 64.

  • edge_hidden_feats (int) – The length of the hidden edge representation vectors. Default to 128.

  • num_step_message_passing (int) – The number of rounds of message passing. Default to 3.

  • num_step_set2set (int) – The number of set2set steps. Default to 6.

  • num_layer_set2set (int) – The number of set2set layers. Default to 3.

  • mode (str) – The model type, ‘classification’ or ‘regression’. Default to ‘regression’.

  • number_atom_features (int) – The length of the initial atom feature vectors. Default to 30.

  • number_bond_features (int) – The length of the initial bond feature vectors. Default to 11.

  • n_classes (int) – The number of classes to predict per task (only used when mode is ‘classification’). Default to 2.

  • self_loop (bool) – Whether to add self loops for the nodes, i.e. edges from nodes to themselves. Generally, an MPNNModel does not require self loops. Default to False.

  • kwargs – This can include any keyword argument of TorchModel.