Loading deepchem/models/__init__.py +1 −0 Original line number Diff line number Diff line Loading @@ -32,4 +32,5 @@ from deepchem.models.tensorgraph.models.graph_models import WeaveTensorGraph, DT from deepchem.models.tensorgraph.models.symmetry_function_regression import BPSymmetryFunctionRegression, ANIRegression from deepchem.models.tensorgraph.models.seqtoseq import SeqToSeq from deepchem.models.tensorgraph.models.gan import GAN, WGAN from deepchem.models.tensorgraph.models.text_cnn import TextCNNTensorGraph deepchem/models/tensorgraph/layers.py +288 −61 Original line number Diff line number Diff line Loading @@ -27,6 +27,7 @@ class Layer(object): self.op_type = "gpu" self.variable_scope = '' self.variable_values = None self.out_tensor = None self.rnn_initial_states = [] self.rnn_final_states = [] self.rnn_zero_states = [] Loading @@ -51,9 +52,24 @@ class Layer(object): def create_tensor(self, in_layers=None, set_tensors=True, **kwargs): raise NotImplementedError("Subclasses must implement for themselves") def clone(self, in_layers): """Create a copy of this layer with different inputs.""" saved_inputs = self.in_layers self.in_layers = [] saved_tensors = self.none_tensors() copy = deepcopy(self) self.in_layers = saved_inputs self.set_tensors(saved_tensors) copy.in_layers = in_layers return copy def shared(self, in_layers): """ Share weights with different in tensors and a new out tensor Create a copy of this layer that shares variables with it. This is similar to clone(), but where clone() creates two independent layers, this causes the layers to share variables with each other. Parameters ---------- in_layers: list tensor Loading @@ -63,7 +79,9 @@ class Layer(object): ------- Layer """ raise NotImplementedError("Each Layer must implement shared for itself") if self.variable_scope == '': return self.clone(in_layers) raise ValueError('%s does not implement shared()' % self.__class__.__name__) def __call__(self, *in_layers): return self.create_tensor(in_layers=in_layers, set_tensors=False) Loading Loading @@ -173,14 +191,14 @@ class Layer(object): elif self.summary_op == 'histogram': tf.summary.histogram(self.name, self.tb_input, self.collections) def copy(self, replacements={}, variables_graph=None): def copy(self, replacements={}, variables_graph=None, shared=False): """Duplicate this Layer and all its inputs. This creates and returns a clone of this layer. It also recursively calls copy() on all of this layer's inputs to clone the entire hierarchy of layers. In the process, you can optionally tell it to replace particular layers with specific existing ones. For example, you can clone a stack of layers, while connecting the topmost ones to different inputs. This is similar to clone(), but instead of only cloning one layer, it also recursively calls copy() on all of this layer's inputs to clone the entire hierarchy of layers. In the process, you can optionally tell it to replace particular layers with specific existing ones. For example, you can clone a stack of layers, while connecting the topmost ones to different inputs. For example, consider a stack of dense layers that depend on an input: Loading Loading @@ -210,20 +228,24 @@ class Layer(object): the current value of each variable in each layer is recorded, and the copy has that value specified as its initial value. This allows a piece of a pre-trained model to be copied to another model. shared: bool if True, create new layers by calling shared() on the input layers. This means the newly created layers will share variables with the original ones. """ if self in replacements: return replacements[self] copied_inputs = [ layer.copy(replacements, variables_graph) for layer in self.in_layers layer.copy(replacements, variables_graph, shared) for layer in self.in_layers ] saved_inputs = self.in_layers self.in_layers = [] saved_tensors = self.none_tensors() copy = deepcopy(self) self.in_layers = saved_inputs self.set_tensors(saved_tensors) copy.in_layers = copied_inputs if shared: copy = self.shared(copied_inputs) else: copy = self.clone(copied_inputs) if variables_graph is not None: if shared: raise ValueError('Cannot specify variables_graph when shared==True') variables = variables_graph.get_layer_variables(self) if len(variables) > 0: with variables_graph._get_tf("Graph").as_default(): Loading Loading @@ -292,9 +314,9 @@ class TensorWrapper(Layer): """Used to wrap a tensorflow tensor.""" def __init__(self, out_tensor, **kwargs): super(TensorWrapper, self).__init__(**kwargs) self.out_tensor = out_tensor self._shape = out_tensor.get_shape().as_list() super(TensorWrapper, self).__init__(**kwargs) def create_tensor(self, in_layers=None, **kwargs): """Take no actions.""" Loading @@ -314,6 +336,33 @@ def convert_to_layers(in_layers): return layers class SharedVariableScope(Layer): """A Layer that can share variables with another layer via name scope. This abstract class can be used as a parent for any layer that implements shared() by means of the variable name scope. It exists to avoid duplicated code. """ def __init__(self, **kwargs): super(SharedVariableScope, self).__init__(**kwargs) self._reuse = False self._shared_with = None def shared(self, in_layers): copy = self.clone(in_layers) self._reuse = True copy._reuse = True copy._shared_with = self return copy def _get_scope_name(self): if self._shared_with is None: return self.name else: return self._shared_with._get_scope_name() class Conv1D(Layer): """A 1D convolution on the input. Loading Loading @@ -379,7 +428,7 @@ class Conv1D(Layer): return out_tensor class Dense(Layer): class Dense(SharedVariableScope): def __init__( self, Loading Loading @@ -422,8 +471,6 @@ class Dense(Layer): self._shape = tuple(parent_shape[:-1]) + (out_channels,) except: pass self._reuse = False self._shared_with = None def create_tensor(self, in_layers=None, set_tensors=True, **kwargs): inputs = self._get_input_tensors(in_layers) Loading Loading @@ -460,25 +507,6 @@ class Dense(Layer): self.out_tensor = out_tensor return out_tensor def shared(self, in_layers): copy = Dense( self.out_channels, self.activation_fn, self.biases_initializer, self.weights_initializer, time_series=self.time_series, in_layers=in_layers) self._reuse = True copy._reuse = True copy._shared_with = self return copy def _get_scope_name(self): if self._shared_with is None: return self.name else: return self._shared_with._get_scope_name() class Highway(Layer): """ Create a highway layer. y = H(x) * T(x) + x * (1 - T(x)) Loading Loading @@ -972,6 +1000,9 @@ class Concat(Layer): try: s = list(self.in_layers[0].shape) for parent in self.in_layers[1:]: if s[axis] is None or parent.shape[axis] is None: s[axis] = None else: s[axis] += parent.shape[axis] self._shape = tuple(s) except: Loading Loading @@ -1419,7 +1450,7 @@ class ReduceSquareDifference(Layer): return out_tensor class Conv2D(Layer): class Conv2D(SharedVariableScope): """A 2D convolution on the input. This layer expects its input to be a four dimensional tensor of shape (batch size, height, width, # channels). Loading Loading @@ -1481,6 +1512,8 @@ class Conv2D(Layer): parent_tensor = inputs[0] if len(parent_tensor.get_shape()) == 3: parent_tensor = tf.expand_dims(parent_tensor, 3) for reuse in (self._reuse, False): try: out_tensor = tf.contrib.layers.conv2d( parent_tensor, num_outputs=self.num_outputs, Loading @@ -1489,15 +1522,22 @@ class Conv2D(Layer): padding=self.padding, activation_fn=self.activation_fn, normalizer_fn=self.normalizer_fn, scope=self.scope_name) out_tensor = out_tensor scope=self._get_scope_name(), reuse=reuse) break except ValueError: if reuse: # This probably means the variable hasn't been created yet, so try again # with reuse set to false. continue raise if set_tensors: self._record_variable_scope(self.scope_name) self.out_tensor = out_tensor return out_tensor class Conv3D(Layer): class Conv3D(SharedVariableScope): """A 3D convolution on the input. This layer expects its input to be a five dimensional tensor of shape Loading Loading @@ -1561,6 +1601,8 @@ class Conv3D(Layer): parent_tensor = inputs[0] if len(parent_tensor.get_shape()) == 4: parent_tensor = tf.expand_dims(parent_tensor, 4) for reuse in (self._reuse, False): try: out_tensor = tf.layers.conv3d( parent_tensor, filters=self.num_outputs, Loading @@ -1569,7 +1611,15 @@ class Conv3D(Layer): padding=self.padding, activation=self.activation_fn, activity_regularizer=self.normalizer_fn, name=self.scope_name) name=self._get_scope_name(), reuse=reuse) break except ValueError: if reuse: # This probably means the variable hasn't been created yet, so try again # with reuse set to false. continue raise out_tensor = out_tensor if set_tensors: self._record_variable_scope(self.scope_name) Loading @@ -1577,6 +1627,183 @@ class Conv3D(Layer): return out_tensor class Conv2DTranspose(SharedVariableScope): """A transposed 2D convolution on the input. This layer is typically used for upsampling in a deconvolutional network. It expects its input to be a four dimensional tensor of shape (batch size, height, width, # channels). If there is only one channel, the fourth dimension may optionally be omitted. """ def __init__(self, num_outputs, kernel_size=5, stride=1, padding='SAME', activation_fn=tf.nn.relu, normalizer_fn=None, scope_name=None, **kwargs): """Create a Conv2DTranspose layer. Parameters ---------- num_outputs: int the number of outputs produced by the convolutional kernel kernel_size: int or tuple the width of the convolutional kernel. This can be either a two element tuple, giving the kernel size along each dimension, or an integer to use the same size along both dimensions. stride: int or tuple the stride between applications of the convolutional kernel. This can be either a two element tuple, giving the stride along each dimension, or an integer to use the same stride along both dimensions. padding: str the padding method to use, either 'SAME' or 'VALID' activation_fn: object the Tensorflow activation function to apply to the output normalizer_fn: object the Tensorflow normalizer function to apply to the output """ self.num_outputs = num_outputs self.kernel_size = kernel_size self.stride = stride self.padding = padding self.activation_fn = activation_fn self.normalizer_fn = normalizer_fn super(Conv2DTranspose, self).__init__(**kwargs) if scope_name is None: scope_name = self.name self.scope_name = scope_name try: parent_shape = self.in_layers[0].shape strides = stride if isinstance(stride, int): strides = (stride, stride) self._shape = (parent_shape[0], parent_shape[1] * strides[0], parent_shape[2] * strides[1], num_outputs) except: pass def create_tensor(self, in_layers=None, set_tensors=True, **kwargs): inputs = self._get_input_tensors(in_layers) parent_tensor = inputs[0] if len(parent_tensor.get_shape()) == 3: parent_tensor = tf.expand_dims(parent_tensor, 3) for reuse in (self._reuse, False): try: out_tensor = tf.contrib.layers.conv2d_transpose( parent_tensor, num_outputs=self.num_outputs, kernel_size=self.kernel_size, stride=self.stride, padding=self.padding, activation_fn=self.activation_fn, normalizer_fn=self.normalizer_fn, scope=self._get_scope_name(), reuse=reuse) break except ValueError: if reuse: # This probably means the variable hasn't been created yet, so try again # with reuse set to false. continue raise if set_tensors: self._record_variable_scope(self.scope_name) self.out_tensor = out_tensor return out_tensor class Conv3DTranspose(SharedVariableScope): """A transposed 3D convolution on the input. This layer is typically used for upsampling in a deconvolutional network. It expects its input to be a five dimensional tensor of shape (batch size, height, width, depth, # channels). If there is only one channel, the fifth dimension may optionally be omitted. """ def __init__(self, num_outputs, kernel_size=5, stride=1, padding='SAME', activation_fn=tf.nn.relu, normalizer_fn=None, scope_name=None, **kwargs): """Create a Conv3DTranspose layer. Parameters ---------- num_outputs: int the number of outputs produced by the convolutional kernel kernel_size: int or tuple the width of the convolutional kernel. This can be either a three element tuple, giving the kernel size along each dimension, or an integer to use the same size along both dimensions. stride: int or tuple the stride between applications of the convolutional kernel. This can be either a three element tuple, giving the stride along each dimension, or an integer to use the same stride along both dimensions. padding: str the padding method to use, either 'SAME' or 'VALID' activation_fn: object the Tensorflow activation function to apply to the output normalizer_fn: object the Tensorflow normalizer function to apply to the output """ self.num_outputs = num_outputs self.kernel_size = kernel_size self.stride = stride self.padding = padding self.activation_fn = activation_fn self.normalizer_fn = normalizer_fn super(Conv3DTranspose, self).__init__(**kwargs) if scope_name is None: scope_name = self.name self.scope_name = scope_name try: parent_shape = self.in_layers[0].shape strides = stride if isinstance(stride, int): strides = (stride, stride, stride) self._shape = (parent_shape[0], parent_shape[1] * strides[0], parent_shape[2] * strides[1], parent_shape[3] * strides[2], num_outputs) except: pass def create_tensor(self, in_layers=None, set_tensors=True, **kwargs): inputs = self._get_input_tensors(in_layers) parent_tensor = inputs[0] if len(parent_tensor.get_shape()) == 4: parent_tensor = tf.expand_dims(parent_tensor, 4) for reuse in (self._reuse, False): try: out_tensor = tf.layers.conv3d_transpose( parent_tensor, filters=self.num_outputs, kernel_size=self.kernel_size, strides=self.stride, padding=self.padding, activation=self.activation_fn, activity_regularizer=self.normalizer_fn, name=self._get_scope_name(), reuse=reuse) break except ValueError: if reuse: # This probably means the variable hasn't been created yet, so try again # with reuse set to false. continue raise if set_tensors: self._record_variable_scope(self.scope_name) self.out_tensor = out_tensor return out_tensor class MaxPool1D(Layer): """A 1D max pooling on the input. Loading deepchem/models/tensorgraph/models/gan.py 0 → 100644 +495 −0 File added.Preview size limit exceeded, changes collapsed. Show changes deepchem/models/tensorgraph/tests/test_gan.py 0 → 100644 +100 −0 Original line number Diff line number Diff line import deepchem as dc import numpy as np import tensorflow as tf import unittest from deepchem.models.tensorgraph import layers def generate_batch(batch_size): """Draw training data from a Gaussian distribution, where the mean is a conditional input.""" means = 10 * np.random.random([batch_size, 1]) values = np.random.normal(means, scale=2.0) return means, values def generate_data(gan, batches, batch_size): for i in range(batches): means, values = generate_batch(batch_size) batch = {gan.data_inputs[0]: values, gan.conditional_inputs[0]: means} yield batch class TestGAN(unittest.TestCase): def test_cgan(self): """Test fitting a conditional GAN.""" class ExampleGAN(dc.models.GAN): def get_noise_input_shape(self): return (None, 2) def get_data_input_shapes(self): return [(None, 1)] def get_conditional_input_shapes(self): return [(None, 1)] def create_generator(self, noise_input, conditional_inputs): gen_in = layers.Concat([noise_input] + conditional_inputs) return [layers.Dense(1, in_layers=gen_in)] def create_discriminator(self, data_inputs, conditional_inputs): discrim_in = layers.Concat(data_inputs + conditional_inputs) dense = layers.Dense(10, in_layers=discrim_in, activation_fn=tf.nn.relu) return layers.Dense(1, in_layers=dense, activation_fn=tf.sigmoid) gan = ExampleGAN(learning_rate=0.003) gan.fit_gan( generate_data(gan, 5000, 100), generator_steps=0.5, checkpoint_interval=0) # See if it has done a plausible job of learning the distribution. means = 10 * np.random.random([1000, 1]) values = gan.predict_gan_generator(conditional_inputs=[means]) deltas = values - means assert abs(np.mean(deltas)) < 1.0 assert np.std(deltas) > 1.0 def test_wgan(self): """Test fitting a conditional WGAN.""" class ExampleWGAN(dc.models.WGAN): def get_noise_input_shape(self): return (None, 2) def get_data_input_shapes(self): return [(None, 1)] def get_conditional_input_shapes(self): return [(None, 1)] def create_generator(self, noise_input, conditional_inputs): gen_in = layers.Concat([noise_input] + conditional_inputs) return [layers.Dense(1, in_layers=gen_in)] def create_discriminator(self, data_inputs, conditional_inputs): discrim_in = layers.Concat(data_inputs + conditional_inputs) dense = layers.Dense(10, in_layers=discrim_in, activation_fn=tf.nn.relu) return layers.Dense(1, in_layers=dense) # We have to set the gradient penalty very small because the generator's # output is only a single number, so the default penalty would constrain # it far too much. gan = ExampleWGAN(learning_rate=0.003, gradient_penalty=0.1) gan.fit_gan( generate_data(gan, 10000, 100), generator_steps=0.1, checkpoint_interval=0) # See if it has done a plausible job of learning the distribution. means = 10 * np.random.random([1000, 1]) values = gan.predict_gan_generator(conditional_inputs=[means]) deltas = values - means assert abs(np.mean(deltas)) < 1.0 assert np.std(deltas) > 1.0 deepchem/models/tensorgraph/tests/test_layers.py +39 −1 Original line number Diff line number Diff line Loading @@ -5,7 +5,7 @@ from tensorflow.python.framework import test_util from deepchem.feat.graph_features import ConvMolFeaturizer from deepchem.feat.mol_graphs import ConvMol from deepchem.models.tensorgraph.layers import Add, Conv3D, MaxPool2D, MaxPool3D, GraphCNN, GraphEmbedPoolLayer from deepchem.models.tensorgraph.layers import Add, MaxPool2D, MaxPool3D, GraphCNN, GraphEmbedPoolLayer from deepchem.models.tensorgraph.layers import AlphaShareLayer from deepchem.models.tensorgraph.layers import AttnLSTMEmbedding from deepchem.models.tensorgraph.layers import BatchNorm Loading @@ -15,6 +15,9 @@ from deepchem.models.tensorgraph.layers import Concat from deepchem.models.tensorgraph.layers import Constant from deepchem.models.tensorgraph.layers import Conv1D, Squeeze from deepchem.models.tensorgraph.layers import Conv2D from deepchem.models.tensorgraph.layers import Conv2DTranspose from deepchem.models.tensorgraph.layers import Conv3D from deepchem.models.tensorgraph.layers import Conv3DTranspose from deepchem.models.tensorgraph.layers import Dense from deepchem.models.tensorgraph.layers import Exp from deepchem.models.tensorgraph.layers import Flatten Loading Loading @@ -395,6 +398,41 @@ class TestLayers(test_util.TensorFlowTestCase): assert out_tensor.shape == (batch_size, length, width, depth, out_channels) def test_conv_2D_transpose(self): """Test that Conv2DTranspose can be invoked.""" length = 4 width = 5 in_channels = 2 out_channels = 3 batch_size = 20 in_tensor = np.random.rand(batch_size, length, width, in_channels) with self.test_session() as sess: in_tensor = tf.convert_to_tensor(in_tensor, dtype=tf.float32) out_tensor = Conv2DTranspose( out_channels, kernel_size=1, stride=2)(in_tensor) sess.run(tf.global_variables_initializer()) out_tensor = out_tensor.eval() assert out_tensor.shape == (batch_size, 2 * length, 2 * width, out_channels) def test_conv_3D_transpose(self): """Test that Conv3DTranspose can be invoked.""" length = 4 width = 5 depth = 6 in_channels = 2 out_channels = 3 batch_size = 20 in_tensor = np.random.rand(batch_size, length, width, depth, in_channels) with self.test_session() as sess: in_tensor = tf.convert_to_tensor(in_tensor, dtype=tf.float32) out_tensor = Conv3DTranspose( out_channels, kernel_size=1, stride=(2, 3, 1))(in_tensor) sess.run(tf.global_variables_initializer()) out_tensor = out_tensor.eval() assert out_tensor.shape == (batch_size, 2 * length, 3 * width, depth, out_channels) def test_maxpool2D(self): """Test that MaxPool2D can be invoked.""" length = 2 Loading Loading
deepchem/models/__init__.py +1 −0 Original line number Diff line number Diff line Loading @@ -32,4 +32,5 @@ from deepchem.models.tensorgraph.models.graph_models import WeaveTensorGraph, DT from deepchem.models.tensorgraph.models.symmetry_function_regression import BPSymmetryFunctionRegression, ANIRegression from deepchem.models.tensorgraph.models.seqtoseq import SeqToSeq from deepchem.models.tensorgraph.models.gan import GAN, WGAN from deepchem.models.tensorgraph.models.text_cnn import TextCNNTensorGraph
deepchem/models/tensorgraph/layers.py +288 −61 Original line number Diff line number Diff line Loading @@ -27,6 +27,7 @@ class Layer(object): self.op_type = "gpu" self.variable_scope = '' self.variable_values = None self.out_tensor = None self.rnn_initial_states = [] self.rnn_final_states = [] self.rnn_zero_states = [] Loading @@ -51,9 +52,24 @@ class Layer(object): def create_tensor(self, in_layers=None, set_tensors=True, **kwargs): raise NotImplementedError("Subclasses must implement for themselves") def clone(self, in_layers): """Create a copy of this layer with different inputs.""" saved_inputs = self.in_layers self.in_layers = [] saved_tensors = self.none_tensors() copy = deepcopy(self) self.in_layers = saved_inputs self.set_tensors(saved_tensors) copy.in_layers = in_layers return copy def shared(self, in_layers): """ Share weights with different in tensors and a new out tensor Create a copy of this layer that shares variables with it. This is similar to clone(), but where clone() creates two independent layers, this causes the layers to share variables with each other. Parameters ---------- in_layers: list tensor Loading @@ -63,7 +79,9 @@ class Layer(object): ------- Layer """ raise NotImplementedError("Each Layer must implement shared for itself") if self.variable_scope == '': return self.clone(in_layers) raise ValueError('%s does not implement shared()' % self.__class__.__name__) def __call__(self, *in_layers): return self.create_tensor(in_layers=in_layers, set_tensors=False) Loading Loading @@ -173,14 +191,14 @@ class Layer(object): elif self.summary_op == 'histogram': tf.summary.histogram(self.name, self.tb_input, self.collections) def copy(self, replacements={}, variables_graph=None): def copy(self, replacements={}, variables_graph=None, shared=False): """Duplicate this Layer and all its inputs. This creates and returns a clone of this layer. It also recursively calls copy() on all of this layer's inputs to clone the entire hierarchy of layers. In the process, you can optionally tell it to replace particular layers with specific existing ones. For example, you can clone a stack of layers, while connecting the topmost ones to different inputs. This is similar to clone(), but instead of only cloning one layer, it also recursively calls copy() on all of this layer's inputs to clone the entire hierarchy of layers. In the process, you can optionally tell it to replace particular layers with specific existing ones. For example, you can clone a stack of layers, while connecting the topmost ones to different inputs. For example, consider a stack of dense layers that depend on an input: Loading Loading @@ -210,20 +228,24 @@ class Layer(object): the current value of each variable in each layer is recorded, and the copy has that value specified as its initial value. This allows a piece of a pre-trained model to be copied to another model. shared: bool if True, create new layers by calling shared() on the input layers. This means the newly created layers will share variables with the original ones. """ if self in replacements: return replacements[self] copied_inputs = [ layer.copy(replacements, variables_graph) for layer in self.in_layers layer.copy(replacements, variables_graph, shared) for layer in self.in_layers ] saved_inputs = self.in_layers self.in_layers = [] saved_tensors = self.none_tensors() copy = deepcopy(self) self.in_layers = saved_inputs self.set_tensors(saved_tensors) copy.in_layers = copied_inputs if shared: copy = self.shared(copied_inputs) else: copy = self.clone(copied_inputs) if variables_graph is not None: if shared: raise ValueError('Cannot specify variables_graph when shared==True') variables = variables_graph.get_layer_variables(self) if len(variables) > 0: with variables_graph._get_tf("Graph").as_default(): Loading Loading @@ -292,9 +314,9 @@ class TensorWrapper(Layer): """Used to wrap a tensorflow tensor.""" def __init__(self, out_tensor, **kwargs): super(TensorWrapper, self).__init__(**kwargs) self.out_tensor = out_tensor self._shape = out_tensor.get_shape().as_list() super(TensorWrapper, self).__init__(**kwargs) def create_tensor(self, in_layers=None, **kwargs): """Take no actions.""" Loading @@ -314,6 +336,33 @@ def convert_to_layers(in_layers): return layers class SharedVariableScope(Layer): """A Layer that can share variables with another layer via name scope. This abstract class can be used as a parent for any layer that implements shared() by means of the variable name scope. It exists to avoid duplicated code. """ def __init__(self, **kwargs): super(SharedVariableScope, self).__init__(**kwargs) self._reuse = False self._shared_with = None def shared(self, in_layers): copy = self.clone(in_layers) self._reuse = True copy._reuse = True copy._shared_with = self return copy def _get_scope_name(self): if self._shared_with is None: return self.name else: return self._shared_with._get_scope_name() class Conv1D(Layer): """A 1D convolution on the input. Loading Loading @@ -379,7 +428,7 @@ class Conv1D(Layer): return out_tensor class Dense(Layer): class Dense(SharedVariableScope): def __init__( self, Loading Loading @@ -422,8 +471,6 @@ class Dense(Layer): self._shape = tuple(parent_shape[:-1]) + (out_channels,) except: pass self._reuse = False self._shared_with = None def create_tensor(self, in_layers=None, set_tensors=True, **kwargs): inputs = self._get_input_tensors(in_layers) Loading Loading @@ -460,25 +507,6 @@ class Dense(Layer): self.out_tensor = out_tensor return out_tensor def shared(self, in_layers): copy = Dense( self.out_channels, self.activation_fn, self.biases_initializer, self.weights_initializer, time_series=self.time_series, in_layers=in_layers) self._reuse = True copy._reuse = True copy._shared_with = self return copy def _get_scope_name(self): if self._shared_with is None: return self.name else: return self._shared_with._get_scope_name() class Highway(Layer): """ Create a highway layer. y = H(x) * T(x) + x * (1 - T(x)) Loading Loading @@ -972,6 +1000,9 @@ class Concat(Layer): try: s = list(self.in_layers[0].shape) for parent in self.in_layers[1:]: if s[axis] is None or parent.shape[axis] is None: s[axis] = None else: s[axis] += parent.shape[axis] self._shape = tuple(s) except: Loading Loading @@ -1419,7 +1450,7 @@ class ReduceSquareDifference(Layer): return out_tensor class Conv2D(Layer): class Conv2D(SharedVariableScope): """A 2D convolution on the input. This layer expects its input to be a four dimensional tensor of shape (batch size, height, width, # channels). Loading Loading @@ -1481,6 +1512,8 @@ class Conv2D(Layer): parent_tensor = inputs[0] if len(parent_tensor.get_shape()) == 3: parent_tensor = tf.expand_dims(parent_tensor, 3) for reuse in (self._reuse, False): try: out_tensor = tf.contrib.layers.conv2d( parent_tensor, num_outputs=self.num_outputs, Loading @@ -1489,15 +1522,22 @@ class Conv2D(Layer): padding=self.padding, activation_fn=self.activation_fn, normalizer_fn=self.normalizer_fn, scope=self.scope_name) out_tensor = out_tensor scope=self._get_scope_name(), reuse=reuse) break except ValueError: if reuse: # This probably means the variable hasn't been created yet, so try again # with reuse set to false. continue raise if set_tensors: self._record_variable_scope(self.scope_name) self.out_tensor = out_tensor return out_tensor class Conv3D(Layer): class Conv3D(SharedVariableScope): """A 3D convolution on the input. This layer expects its input to be a five dimensional tensor of shape Loading Loading @@ -1561,6 +1601,8 @@ class Conv3D(Layer): parent_tensor = inputs[0] if len(parent_tensor.get_shape()) == 4: parent_tensor = tf.expand_dims(parent_tensor, 4) for reuse in (self._reuse, False): try: out_tensor = tf.layers.conv3d( parent_tensor, filters=self.num_outputs, Loading @@ -1569,7 +1611,15 @@ class Conv3D(Layer): padding=self.padding, activation=self.activation_fn, activity_regularizer=self.normalizer_fn, name=self.scope_name) name=self._get_scope_name(), reuse=reuse) break except ValueError: if reuse: # This probably means the variable hasn't been created yet, so try again # with reuse set to false. continue raise out_tensor = out_tensor if set_tensors: self._record_variable_scope(self.scope_name) Loading @@ -1577,6 +1627,183 @@ class Conv3D(Layer): return out_tensor class Conv2DTranspose(SharedVariableScope): """A transposed 2D convolution on the input. This layer is typically used for upsampling in a deconvolutional network. It expects its input to be a four dimensional tensor of shape (batch size, height, width, # channels). If there is only one channel, the fourth dimension may optionally be omitted. """ def __init__(self, num_outputs, kernel_size=5, stride=1, padding='SAME', activation_fn=tf.nn.relu, normalizer_fn=None, scope_name=None, **kwargs): """Create a Conv2DTranspose layer. Parameters ---------- num_outputs: int the number of outputs produced by the convolutional kernel kernel_size: int or tuple the width of the convolutional kernel. This can be either a two element tuple, giving the kernel size along each dimension, or an integer to use the same size along both dimensions. stride: int or tuple the stride between applications of the convolutional kernel. This can be either a two element tuple, giving the stride along each dimension, or an integer to use the same stride along both dimensions. padding: str the padding method to use, either 'SAME' or 'VALID' activation_fn: object the Tensorflow activation function to apply to the output normalizer_fn: object the Tensorflow normalizer function to apply to the output """ self.num_outputs = num_outputs self.kernel_size = kernel_size self.stride = stride self.padding = padding self.activation_fn = activation_fn self.normalizer_fn = normalizer_fn super(Conv2DTranspose, self).__init__(**kwargs) if scope_name is None: scope_name = self.name self.scope_name = scope_name try: parent_shape = self.in_layers[0].shape strides = stride if isinstance(stride, int): strides = (stride, stride) self._shape = (parent_shape[0], parent_shape[1] * strides[0], parent_shape[2] * strides[1], num_outputs) except: pass def create_tensor(self, in_layers=None, set_tensors=True, **kwargs): inputs = self._get_input_tensors(in_layers) parent_tensor = inputs[0] if len(parent_tensor.get_shape()) == 3: parent_tensor = tf.expand_dims(parent_tensor, 3) for reuse in (self._reuse, False): try: out_tensor = tf.contrib.layers.conv2d_transpose( parent_tensor, num_outputs=self.num_outputs, kernel_size=self.kernel_size, stride=self.stride, padding=self.padding, activation_fn=self.activation_fn, normalizer_fn=self.normalizer_fn, scope=self._get_scope_name(), reuse=reuse) break except ValueError: if reuse: # This probably means the variable hasn't been created yet, so try again # with reuse set to false. continue raise if set_tensors: self._record_variable_scope(self.scope_name) self.out_tensor = out_tensor return out_tensor class Conv3DTranspose(SharedVariableScope): """A transposed 3D convolution on the input. This layer is typically used for upsampling in a deconvolutional network. It expects its input to be a five dimensional tensor of shape (batch size, height, width, depth, # channels). If there is only one channel, the fifth dimension may optionally be omitted. """ def __init__(self, num_outputs, kernel_size=5, stride=1, padding='SAME', activation_fn=tf.nn.relu, normalizer_fn=None, scope_name=None, **kwargs): """Create a Conv3DTranspose layer. Parameters ---------- num_outputs: int the number of outputs produced by the convolutional kernel kernel_size: int or tuple the width of the convolutional kernel. This can be either a three element tuple, giving the kernel size along each dimension, or an integer to use the same size along both dimensions. stride: int or tuple the stride between applications of the convolutional kernel. This can be either a three element tuple, giving the stride along each dimension, or an integer to use the same stride along both dimensions. padding: str the padding method to use, either 'SAME' or 'VALID' activation_fn: object the Tensorflow activation function to apply to the output normalizer_fn: object the Tensorflow normalizer function to apply to the output """ self.num_outputs = num_outputs self.kernel_size = kernel_size self.stride = stride self.padding = padding self.activation_fn = activation_fn self.normalizer_fn = normalizer_fn super(Conv3DTranspose, self).__init__(**kwargs) if scope_name is None: scope_name = self.name self.scope_name = scope_name try: parent_shape = self.in_layers[0].shape strides = stride if isinstance(stride, int): strides = (stride, stride, stride) self._shape = (parent_shape[0], parent_shape[1] * strides[0], parent_shape[2] * strides[1], parent_shape[3] * strides[2], num_outputs) except: pass def create_tensor(self, in_layers=None, set_tensors=True, **kwargs): inputs = self._get_input_tensors(in_layers) parent_tensor = inputs[0] if len(parent_tensor.get_shape()) == 4: parent_tensor = tf.expand_dims(parent_tensor, 4) for reuse in (self._reuse, False): try: out_tensor = tf.layers.conv3d_transpose( parent_tensor, filters=self.num_outputs, kernel_size=self.kernel_size, strides=self.stride, padding=self.padding, activation=self.activation_fn, activity_regularizer=self.normalizer_fn, name=self._get_scope_name(), reuse=reuse) break except ValueError: if reuse: # This probably means the variable hasn't been created yet, so try again # with reuse set to false. continue raise if set_tensors: self._record_variable_scope(self.scope_name) self.out_tensor = out_tensor return out_tensor class MaxPool1D(Layer): """A 1D max pooling on the input. Loading
deepchem/models/tensorgraph/models/gan.py 0 → 100644 +495 −0 File added.Preview size limit exceeded, changes collapsed. Show changes
deepchem/models/tensorgraph/tests/test_gan.py 0 → 100644 +100 −0 Original line number Diff line number Diff line import deepchem as dc import numpy as np import tensorflow as tf import unittest from deepchem.models.tensorgraph import layers def generate_batch(batch_size): """Draw training data from a Gaussian distribution, where the mean is a conditional input.""" means = 10 * np.random.random([batch_size, 1]) values = np.random.normal(means, scale=2.0) return means, values def generate_data(gan, batches, batch_size): for i in range(batches): means, values = generate_batch(batch_size) batch = {gan.data_inputs[0]: values, gan.conditional_inputs[0]: means} yield batch class TestGAN(unittest.TestCase): def test_cgan(self): """Test fitting a conditional GAN.""" class ExampleGAN(dc.models.GAN): def get_noise_input_shape(self): return (None, 2) def get_data_input_shapes(self): return [(None, 1)] def get_conditional_input_shapes(self): return [(None, 1)] def create_generator(self, noise_input, conditional_inputs): gen_in = layers.Concat([noise_input] + conditional_inputs) return [layers.Dense(1, in_layers=gen_in)] def create_discriminator(self, data_inputs, conditional_inputs): discrim_in = layers.Concat(data_inputs + conditional_inputs) dense = layers.Dense(10, in_layers=discrim_in, activation_fn=tf.nn.relu) return layers.Dense(1, in_layers=dense, activation_fn=tf.sigmoid) gan = ExampleGAN(learning_rate=0.003) gan.fit_gan( generate_data(gan, 5000, 100), generator_steps=0.5, checkpoint_interval=0) # See if it has done a plausible job of learning the distribution. means = 10 * np.random.random([1000, 1]) values = gan.predict_gan_generator(conditional_inputs=[means]) deltas = values - means assert abs(np.mean(deltas)) < 1.0 assert np.std(deltas) > 1.0 def test_wgan(self): """Test fitting a conditional WGAN.""" class ExampleWGAN(dc.models.WGAN): def get_noise_input_shape(self): return (None, 2) def get_data_input_shapes(self): return [(None, 1)] def get_conditional_input_shapes(self): return [(None, 1)] def create_generator(self, noise_input, conditional_inputs): gen_in = layers.Concat([noise_input] + conditional_inputs) return [layers.Dense(1, in_layers=gen_in)] def create_discriminator(self, data_inputs, conditional_inputs): discrim_in = layers.Concat(data_inputs + conditional_inputs) dense = layers.Dense(10, in_layers=discrim_in, activation_fn=tf.nn.relu) return layers.Dense(1, in_layers=dense) # We have to set the gradient penalty very small because the generator's # output is only a single number, so the default penalty would constrain # it far too much. gan = ExampleWGAN(learning_rate=0.003, gradient_penalty=0.1) gan.fit_gan( generate_data(gan, 10000, 100), generator_steps=0.1, checkpoint_interval=0) # See if it has done a plausible job of learning the distribution. means = 10 * np.random.random([1000, 1]) values = gan.predict_gan_generator(conditional_inputs=[means]) deltas = values - means assert abs(np.mean(deltas)) < 1.0 assert np.std(deltas) > 1.0
deepchem/models/tensorgraph/tests/test_layers.py +39 −1 Original line number Diff line number Diff line Loading @@ -5,7 +5,7 @@ from tensorflow.python.framework import test_util from deepchem.feat.graph_features import ConvMolFeaturizer from deepchem.feat.mol_graphs import ConvMol from deepchem.models.tensorgraph.layers import Add, Conv3D, MaxPool2D, MaxPool3D, GraphCNN, GraphEmbedPoolLayer from deepchem.models.tensorgraph.layers import Add, MaxPool2D, MaxPool3D, GraphCNN, GraphEmbedPoolLayer from deepchem.models.tensorgraph.layers import AlphaShareLayer from deepchem.models.tensorgraph.layers import AttnLSTMEmbedding from deepchem.models.tensorgraph.layers import BatchNorm Loading @@ -15,6 +15,9 @@ from deepchem.models.tensorgraph.layers import Concat from deepchem.models.tensorgraph.layers import Constant from deepchem.models.tensorgraph.layers import Conv1D, Squeeze from deepchem.models.tensorgraph.layers import Conv2D from deepchem.models.tensorgraph.layers import Conv2DTranspose from deepchem.models.tensorgraph.layers import Conv3D from deepchem.models.tensorgraph.layers import Conv3DTranspose from deepchem.models.tensorgraph.layers import Dense from deepchem.models.tensorgraph.layers import Exp from deepchem.models.tensorgraph.layers import Flatten Loading Loading @@ -395,6 +398,41 @@ class TestLayers(test_util.TensorFlowTestCase): assert out_tensor.shape == (batch_size, length, width, depth, out_channels) def test_conv_2D_transpose(self): """Test that Conv2DTranspose can be invoked.""" length = 4 width = 5 in_channels = 2 out_channels = 3 batch_size = 20 in_tensor = np.random.rand(batch_size, length, width, in_channels) with self.test_session() as sess: in_tensor = tf.convert_to_tensor(in_tensor, dtype=tf.float32) out_tensor = Conv2DTranspose( out_channels, kernel_size=1, stride=2)(in_tensor) sess.run(tf.global_variables_initializer()) out_tensor = out_tensor.eval() assert out_tensor.shape == (batch_size, 2 * length, 2 * width, out_channels) def test_conv_3D_transpose(self): """Test that Conv3DTranspose can be invoked.""" length = 4 width = 5 depth = 6 in_channels = 2 out_channels = 3 batch_size = 20 in_tensor = np.random.rand(batch_size, length, width, depth, in_channels) with self.test_session() as sess: in_tensor = tf.convert_to_tensor(in_tensor, dtype=tf.float32) out_tensor = Conv3DTranspose( out_channels, kernel_size=1, stride=(2, 3, 1))(in_tensor) sess.run(tf.global_variables_initializer()) out_tensor = out_tensor.eval() assert out_tensor.shape == (batch_size, 2 * length, 3 * width, depth, out_channels) def test_maxpool2D(self): """Test that MaxPool2D can be invoked.""" length = 2 Loading
