Loading deepchem/models/tensorgraph/chemnet_layers.py +35 −57 Original line number Diff line number Diff line Loading @@ -10,31 +10,9 @@ __license__ = "MIT" import numpy as np import tensorflow as tf from deepchem.models.tensorgraph.layers import KerasLayer from tensorflow.keras.layers import Conv2D, Concatenate, ReLU, Add, MaxPool2D class SmilesEmbedding(KerasLayer): def __init__(self, charset_size, embedding_dim, **kwargs): """ Parameters ---------- charset_size: int, Number of elements in the character to index dictionary embedding_dim: int, Dimensionality of the embedding """ super(SmilesEmbedding, self).__init__(**kwargs) self.charset_size = charset_size self.embedding_dim = embedding_dim def _build_layer(self): return tf.keras.layers.Embedding( input_dim=self.charset_size, output_dim=self.embedding_dim) class Stem(tf.keras.layers.Layer): def __init__(self, num_filters, **kwargs): Loading deepchem/models/tensorgraph/models/chemnet_models.py +55 −72 Original line number Diff line number Diff line """ Smiles2Vec model, described in https://arxiv.org/pdf/1712.02034.pdf ChemCeption model, described in https://arxiv.org/pdf/1706.06689.pdf Implementation of Smiles2Vec and ChemCeption models as part of the ChemNet transfer learning protocol. """ from __future__ import division Loading Loading @@ -38,16 +38,35 @@ RNN_DICT = {"GRU": GRU, "LSTM": LSTM} class Smiles2Vec(KerasModel): """ Smiles2Vec is a RNN basel model that ingests SMILES characters, get their embeddings. These embeddings are fed through a series of RNN (and optionally Conv) layers, with the representations from the final layer as input for classification or regression tasks. 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). """ def __init__(self, char_to_idx, n_tasks=10, batch_size=64, max_seq_len=270, embedding_dim=50, n_classes=2, Loading @@ -58,13 +77,7 @@ class Smiles2Vec(KerasModel): strides=1, rnn_sizes=[224, 384], rnn_types=["GRU", "GRU"], learning_rate=0.0001, model_dir=None, tensorboard=True, optimizer=None, mode="regression", tensorboard_log_frequency=5, restore_from=None, **kwargs): """ Parameters Loading @@ -85,20 +98,8 @@ class Smiles2Vec(KerasModel): Strides used in convolution rnn_sizes: list[int], default [224, 384] Number of hidden units in the RNN cells learning_rate: floatm default 0.0001, Default learning rate used for the model model_dir: str, default None Directory to save model to tensorboard: bool, default True Whether to use TensorBoard optimizer: tf.train.Optimizer, default None Optimizer used. mode: str, default regression Whether to use model for regression or classification tensorboard_log_frequency: int, default 5 How often to log to TensorBoard restore_from: str, default None Where to restore model from """ self.char_to_idx = char_to_idx Loading @@ -117,28 +118,17 @@ class Smiles2Vec(KerasModel): rnn_types), "Should have same number of hidden units as RNNs" self.n_tasks = n_tasks self.mode = mode self.batch_size = batch_size model, loss, output_types = self._build_graph() super(Smiles2Vec, self).__init__( model=model, loss=loss, output_types=output_types, model_dir=model_dir, tensorboard=tensorboard, batch_size=batch_size, optimizer=optimizer, tensorboard_log_frequency=tensorboard_log_frequency, learning_rate=learning_rate, **kwargs) model=model, loss=loss, output_types=output_types, **kwargs) def _build_graph(self): """Build the model.""" smiles_seqs = Input(dtype=tf.int32, shape=(self.max_seq_len,), name='Input') rnn_input = chemnet_layers.SmilesEmbedding( charset_size=len(self.char_to_idx), embedding_dim=self.embedding_dim, name='Embedding')(smiles_seqs) rnn_input = tf.keras.layers.Embedding( input_dim=len(self.char_to_idx), output_dim=self.embedding_dim)(smile_seqs) if self.use_conv: rnn_input = Conv1D( Loading Loading @@ -200,10 +190,30 @@ class Smiles2Vec(KerasModel): class ChemCeption(KerasModel): """ ChemCeption is a CNN based model composed of a series of Inception-ResNet and Reduction layers. The representations from the final layer serve as input for classification or regression tasks. The input for the ChemCeption model is an image representation of the molecule, described in SmilesToImage featurizer. Implements the ChemCeption model that leverages the representational capacties 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 as multiple resolutions, and use a ResNet style skip-connections, combining features from different resolutions. The Reduction layers reduce the spatial extent of the image by max-pooling and 2-strided convolutions. More details of 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. """ def __init__(self, Loading @@ -212,13 +222,7 @@ class ChemCeption(KerasModel): inception_blocks=DEFAULT_INCEPTION_BLOCKS, n_tasks=10, n_classes=2, batch_size=100, learning_rate=0.001, augment=False, tensorboard=True, tensorboard_log_frequency=5, model_dir=None, optimizer=None, mode="regression", **kwargs): """ Loading @@ -234,20 +238,8 @@ class ChemCeption(KerasModel): Number of classification or regression tasks n_classes: int, default 2 Number of classes (used only for classification) batch_size: int, default 100 Minibatch size used for model fitting learning_rate: float, default 0.0001 Learning rate used for training augment: bool, default False Whether to augment images tensorboard: bool, default True Whether to log to TensorBoard (does not work in eager mode.) tensorboard_log_frequency: int, default 5 Frequency to log to TensorBoard model_dir: str, default None, Directory to save the model in optimizer: tf.train.Optimizer, default None Optimizer used for training mode: str, default regression Whether the model is used for regression or classification """ Loading @@ -264,16 +256,7 @@ class ChemCeption(KerasModel): model, loss, output_types = self._build_graph() super(ChemCeption, self).__init__( model=model, loss=loss, output_types=output_types, model_dir=model_dir, tensorboard=tensorboard, batch_size=batch_size, optimizer=optimizer, tensorboard_log_frequency=tensorboard_log_frequency, learning_rate=learning_rate, **kwargs) model=model, loss=loss, output_types=output_types, **kwargs) def _build_graph(self): smile_images = Input(shape=self.input_shape) Loading Loading
deepchem/models/tensorgraph/chemnet_layers.py +35 −57 Original line number Diff line number Diff line Loading @@ -10,31 +10,9 @@ __license__ = "MIT" import numpy as np import tensorflow as tf from deepchem.models.tensorgraph.layers import KerasLayer from tensorflow.keras.layers import Conv2D, Concatenate, ReLU, Add, MaxPool2D class SmilesEmbedding(KerasLayer): def __init__(self, charset_size, embedding_dim, **kwargs): """ Parameters ---------- charset_size: int, Number of elements in the character to index dictionary embedding_dim: int, Dimensionality of the embedding """ super(SmilesEmbedding, self).__init__(**kwargs) self.charset_size = charset_size self.embedding_dim = embedding_dim def _build_layer(self): return tf.keras.layers.Embedding( input_dim=self.charset_size, output_dim=self.embedding_dim) class Stem(tf.keras.layers.Layer): def __init__(self, num_filters, **kwargs): Loading
deepchem/models/tensorgraph/models/chemnet_models.py +55 −72 Original line number Diff line number Diff line """ Smiles2Vec model, described in https://arxiv.org/pdf/1712.02034.pdf ChemCeption model, described in https://arxiv.org/pdf/1706.06689.pdf Implementation of Smiles2Vec and ChemCeption models as part of the ChemNet transfer learning protocol. """ from __future__ import division Loading Loading @@ -38,16 +38,35 @@ RNN_DICT = {"GRU": GRU, "LSTM": LSTM} class Smiles2Vec(KerasModel): """ Smiles2Vec is a RNN basel model that ingests SMILES characters, get their embeddings. These embeddings are fed through a series of RNN (and optionally Conv) layers, with the representations from the final layer as input for classification or regression tasks. 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). """ def __init__(self, char_to_idx, n_tasks=10, batch_size=64, max_seq_len=270, embedding_dim=50, n_classes=2, Loading @@ -58,13 +77,7 @@ class Smiles2Vec(KerasModel): strides=1, rnn_sizes=[224, 384], rnn_types=["GRU", "GRU"], learning_rate=0.0001, model_dir=None, tensorboard=True, optimizer=None, mode="regression", tensorboard_log_frequency=5, restore_from=None, **kwargs): """ Parameters Loading @@ -85,20 +98,8 @@ class Smiles2Vec(KerasModel): Strides used in convolution rnn_sizes: list[int], default [224, 384] Number of hidden units in the RNN cells learning_rate: floatm default 0.0001, Default learning rate used for the model model_dir: str, default None Directory to save model to tensorboard: bool, default True Whether to use TensorBoard optimizer: tf.train.Optimizer, default None Optimizer used. mode: str, default regression Whether to use model for regression or classification tensorboard_log_frequency: int, default 5 How often to log to TensorBoard restore_from: str, default None Where to restore model from """ self.char_to_idx = char_to_idx Loading @@ -117,28 +118,17 @@ class Smiles2Vec(KerasModel): rnn_types), "Should have same number of hidden units as RNNs" self.n_tasks = n_tasks self.mode = mode self.batch_size = batch_size model, loss, output_types = self._build_graph() super(Smiles2Vec, self).__init__( model=model, loss=loss, output_types=output_types, model_dir=model_dir, tensorboard=tensorboard, batch_size=batch_size, optimizer=optimizer, tensorboard_log_frequency=tensorboard_log_frequency, learning_rate=learning_rate, **kwargs) model=model, loss=loss, output_types=output_types, **kwargs) def _build_graph(self): """Build the model.""" smiles_seqs = Input(dtype=tf.int32, shape=(self.max_seq_len,), name='Input') rnn_input = chemnet_layers.SmilesEmbedding( charset_size=len(self.char_to_idx), embedding_dim=self.embedding_dim, name='Embedding')(smiles_seqs) rnn_input = tf.keras.layers.Embedding( input_dim=len(self.char_to_idx), output_dim=self.embedding_dim)(smile_seqs) if self.use_conv: rnn_input = Conv1D( Loading Loading @@ -200,10 +190,30 @@ class Smiles2Vec(KerasModel): class ChemCeption(KerasModel): """ ChemCeption is a CNN based model composed of a series of Inception-ResNet and Reduction layers. The representations from the final layer serve as input for classification or regression tasks. The input for the ChemCeption model is an image representation of the molecule, described in SmilesToImage featurizer. Implements the ChemCeption model that leverages the representational capacties 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 as multiple resolutions, and use a ResNet style skip-connections, combining features from different resolutions. The Reduction layers reduce the spatial extent of the image by max-pooling and 2-strided convolutions. More details of 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. """ def __init__(self, Loading @@ -212,13 +222,7 @@ class ChemCeption(KerasModel): inception_blocks=DEFAULT_INCEPTION_BLOCKS, n_tasks=10, n_classes=2, batch_size=100, learning_rate=0.001, augment=False, tensorboard=True, tensorboard_log_frequency=5, model_dir=None, optimizer=None, mode="regression", **kwargs): """ Loading @@ -234,20 +238,8 @@ class ChemCeption(KerasModel): Number of classification or regression tasks n_classes: int, default 2 Number of classes (used only for classification) batch_size: int, default 100 Minibatch size used for model fitting learning_rate: float, default 0.0001 Learning rate used for training augment: bool, default False Whether to augment images tensorboard: bool, default True Whether to log to TensorBoard (does not work in eager mode.) tensorboard_log_frequency: int, default 5 Frequency to log to TensorBoard model_dir: str, default None, Directory to save the model in optimizer: tf.train.Optimizer, default None Optimizer used for training mode: str, default regression Whether the model is used for regression or classification """ Loading @@ -264,16 +256,7 @@ class ChemCeption(KerasModel): model, loss, output_types = self._build_graph() super(ChemCeption, self).__init__( model=model, loss=loss, output_types=output_types, model_dir=model_dir, tensorboard=tensorboard, batch_size=batch_size, optimizer=optimizer, tensorboard_log_frequency=tensorboard_log_frequency, learning_rate=learning_rate, **kwargs) model=model, loss=loss, output_types=output_types, **kwargs) def _build_graph(self): smile_images = Input(shape=self.input_shape) Loading
