Loading contrib/rl/tictactoe.py +2 −2 Original line number Diff line number Diff line Loading @@ -9,6 +9,7 @@ import deepchem as dc import deepchem.rl.envs.tictactoe from deepchem.models.tensorgraph.layers import Flatten, Dense, SoftMax, \ BatchNorm, Squeeze from deepchem.models.tensorgraph.optimizers import Adam class TicTacToePolicy(dc.rl.Policy): Loading Loading @@ -66,8 +67,7 @@ def eval_tic_tac_toe(value_weight, entropy_weight=0.01, value_weight=value_weight, model_dir=model_dir, optimizer=dc.models.tensorgraph.TFWrapper( tf.train.AdamOptimizer, learning_rate=0.001)) optimizer=Adam(learning_rate=0.001)) try: a3c.restore() except: Loading deepchem/models/tensorgraph/layers.py +14 −0 Original line number Diff line number Diff line Loading @@ -573,6 +573,20 @@ class Concat(Layer): return out_tensor class Stack(Layer): def __init__(self, axis=1, **kwargs): self.axis = axis super(Stack, self).__init__(**kwargs) def create_tensor(self, in_layers=None, set_tensors=True, **kwargs): inputs = self._get_input_tensors(in_layers) out_tensor = tf.stack(inputs, axis=self.axis) if set_tensors: self.out_tensor = out_tensor return out_tensor class Constant(Layer): """Output a constant value.""" Loading deepchem/models/tensorgraph/models/graph_models.py +12 −5 Original line number Diff line number Diff line Loading @@ -7,7 +7,7 @@ from deepchem.metrics import to_one_hot, from_one_hot from deepchem.models.tensorgraph.graph_layers import WeaveLayer, WeaveGather, \ Combine_AP, Separate_AP, DTNNEmbedding, DTNNStep, DTNNGather, DAGLayer, DAGGather, DTNNExtract from deepchem.models.tensorgraph.layers import Dense, Concat, SoftMax, SoftMaxCrossEntropy, GraphConv, BatchNorm, \ GraphPool, GraphGather, WeightedError, Dropout, BatchNormalization GraphPool, GraphGather, WeightedError, Dropout, BatchNormalization, Stack from deepchem.models.tensorgraph.layers import L2Loss, Label, Weights, Feature from deepchem.models.tensorgraph.tensor_graph import TensorGraph from deepchem.trans import undo_transforms Loading Loading @@ -106,7 +106,10 @@ class WeaveTensorGraph(TensorGraph): self.labels_fd.append(label) cost = L2Loss(in_layers=[label, regression]) costs.append(cost) if self.mode == "classification": all_cost = Concat(in_layers=costs, axis=1) elif self.mode == "regression": all_cost = Stack(in_layers=costs, axis=1) self.weights = Weights(shape=(None, self.n_tasks)) loss = WeightedError(in_layers=[all_cost, self.weights]) self.set_loss(loss) Loading Loading @@ -259,7 +262,7 @@ class DTNNTensorGraph(TensorGraph): cost = L2Loss(in_layers=[label, regression]) costs.append(cost) all_cost = Concat(in_layers=costs, axis=1) all_cost = Stack(in_layers=costs, axis=1) self.weights = Weights(shape=(None, self.n_tasks)) loss = WeightedError(in_layers=[all_cost, self.weights]) self.set_loss(loss) Loading Loading @@ -404,8 +407,10 @@ class DAGTensorGraph(TensorGraph): self.labels_fd.append(label) cost = L2Loss(in_layers=[label, regression]) costs.append(cost) if self.mode == "classification": all_cost = Concat(in_layers=costs, axis=1) elif self.mode == "regression": all_cost = Stack(in_layers=costs, axis=1) self.weights = Weights(shape=(None, self.n_tasks)) loss = WeightedError(in_layers=[all_cost, self.weights]) self.set_loss(loss) Loading Loading @@ -545,8 +550,10 @@ class GraphConvTensorGraph(TensorGraph): self.my_labels.append(label) cost = L2Loss(in_layers=[label, regression]) costs.append(cost) if self.mode == "classification": entropy = Concat(in_layers=costs, axis=-1) elif self.mode == "regression": entropy = Stack(in_layers=costs, axis=1) self.my_task_weights = Weights(shape=(None, self.n_tasks)) loss = WeightedError(in_layers=[entropy, self.my_task_weights]) self.set_loss(loss) Loading Loading @@ -604,7 +611,7 @@ class GraphConvTensorGraph(TensorGraph): if not self.built: self.build() return self.evaluate_generator( self.default_generator(dataset), self.default_generator(dataset, predict=True), metrics, labels=self.my_labels, weights=[self.my_task_weights]) Loading deepchem/models/tensorgraph/models/symmetry_function_regression.py +3 −3 Original line number Diff line number Diff line Loading @@ -8,7 +8,7 @@ Created on Thu Jul 6 20:31:47 2017 import numpy as np import tensorflow as tf from deepchem.models.tensorgraph.layers import Dense, Concat, WeightedError from deepchem.models.tensorgraph.layers import Dense, Concat, WeightedError, Stack from deepchem.models.tensorgraph.layers import L2Loss, Label, Weights, Feature from deepchem.models.tensorgraph.tensor_graph import TensorGraph from deepchem.models.tensorgraph.graph_layers import DTNNEmbedding Loading Loading @@ -69,7 +69,7 @@ class BPSymmetryFunctionRegression(TensorGraph): cost = L2Loss(in_layers=[label, output]) costs.append(cost) all_cost = Concat(in_layers=costs, axis=0) all_cost = Stack(in_layers=costs, axis=1) self.weights = Weights(shape=(None, self.n_tasks)) loss = WeightedError(in_layers=[all_cost, self.weights]) self.set_loss(loss) Loading Loading @@ -158,7 +158,7 @@ class ANIRegression(TensorGraph): cost = L2Loss(in_layers=[label, output]) costs.append(cost) all_cost = Concat(in_layers=costs, axis=0) all_cost = Stack(in_layers=costs, axis=1) self.weights = Weights(shape=(None, self.n_tasks)) loss = WeightedError(in_layers=[all_cost, self.weights]) self.set_loss(loss) Loading deepchem/models/tensorgraph/optimizers.py 0 → 100644 +170 −0 Original line number Diff line number Diff line """Optimizers and related classes for use with TensorGraph.""" import tensorflow as tf class Optimizer(object): """An algorithm for optimizing a TensorGraph based model. This is an abstract class. Subclasses represent specific optimization algorithms. """ def _create_optimizer(self, global_step): """Construct the TensorFlow optimizer. Parameters ---------- global_step: tensor a tensor containing the global step index during optimization, used for learning rate decay Returns ------- a new TensorFlow optimizer implementing the algorithm """ raise NotImplemented("Subclasses must implement this") class LearningRateSchedule(object): """A schedule for changing the learning rate over the course of optimization. This is an abstract class. Subclasses represent specific schedules. """ def _create_tensor(self, global_step): """Construct a tensor that equals the learning rate. Parameters ---------- global_step: tensor a tensor containing the global step index during optimization Returns ------- a tensor that equals the learning rate """ raise NotImplemented("Subclasses must implement this") class Adam(Optimizer): """The Adam optimization algorithm.""" def __init__(self, learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-08): """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 """ self.learning_rate = learning_rate self.beta1 = beta1 self.beta2 = beta2 self.epsilon = epsilon def _create_optimizer(self, global_step): if isinstance(self.learning_rate, LearningRateSchedule): learning_rate = self.learning_rate._create_tensor(global_step) else: learning_rate = self.learning_rate return tf.train.AdamOptimizer( learning_rate=learning_rate, beta1=self.beta1, beta2=self.beta2, epsilon=self.epsilon) class GradientDescent(Optimizer): """The gradient descent optimization algorithm.""" def __init__(self, learning_rate=0.001): """Construct a gradient descent optimizer. Parameters ---------- learning_rate: float or LearningRateSchedule the learning rate to use for optimization """ self.learning_rate = learning_rate def _create_optimizer(self, global_step): if isinstance(self.learning_rate, LearningRateSchedule): learning_rate = self.learning_rate._create_tensor(global_step) else: learning_rate = self.learning_rate return tf.train.GradientDescentOptimizer(learning_rate=learning_rate) class ExponentialDecay(LearningRateSchedule): """A learning rate that decreases exponentially with the number of training steps.""" def __init__(self, initial_rate, decay_rate, decay_steps, staircase=True): """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 """ self.initial_rate = initial_rate self.decay_rate = decay_rate self.decay_steps = decay_steps self.staircase = staircase def _create_tensor(self, global_step): return tf.train.exponential_decay( learning_rate=self.initial_rate, global_step=global_step, decay_rate=self.decay_rate, decay_steps=self.decay_steps, staircase=self.staircase) class PolynomialDecay(LearningRateSchedule): """A learning rate that decreases from an initial value to a final value over a fixed number of training steps.""" def __init__(self, initial_rate, final_rate, decay_steps, power=1.0): """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 """ self.initial_rate = initial_rate self.final_rate = final_rate self.decay_steps = decay_steps self.power = power def _create_tensor(self, global_step): return tf.train.polynomial_decay( learning_rate=self.initial_rate, end_learning_rate=self.final_rate, global_step=global_step, decay_steps=self.decay_steps, power=self.power) Loading
contrib/rl/tictactoe.py +2 −2 Original line number Diff line number Diff line Loading @@ -9,6 +9,7 @@ import deepchem as dc import deepchem.rl.envs.tictactoe from deepchem.models.tensorgraph.layers import Flatten, Dense, SoftMax, \ BatchNorm, Squeeze from deepchem.models.tensorgraph.optimizers import Adam class TicTacToePolicy(dc.rl.Policy): Loading Loading @@ -66,8 +67,7 @@ def eval_tic_tac_toe(value_weight, entropy_weight=0.01, value_weight=value_weight, model_dir=model_dir, optimizer=dc.models.tensorgraph.TFWrapper( tf.train.AdamOptimizer, learning_rate=0.001)) optimizer=Adam(learning_rate=0.001)) try: a3c.restore() except: Loading
deepchem/models/tensorgraph/layers.py +14 −0 Original line number Diff line number Diff line Loading @@ -573,6 +573,20 @@ class Concat(Layer): return out_tensor class Stack(Layer): def __init__(self, axis=1, **kwargs): self.axis = axis super(Stack, self).__init__(**kwargs) def create_tensor(self, in_layers=None, set_tensors=True, **kwargs): inputs = self._get_input_tensors(in_layers) out_tensor = tf.stack(inputs, axis=self.axis) if set_tensors: self.out_tensor = out_tensor return out_tensor class Constant(Layer): """Output a constant value.""" Loading
deepchem/models/tensorgraph/models/graph_models.py +12 −5 Original line number Diff line number Diff line Loading @@ -7,7 +7,7 @@ from deepchem.metrics import to_one_hot, from_one_hot from deepchem.models.tensorgraph.graph_layers import WeaveLayer, WeaveGather, \ Combine_AP, Separate_AP, DTNNEmbedding, DTNNStep, DTNNGather, DAGLayer, DAGGather, DTNNExtract from deepchem.models.tensorgraph.layers import Dense, Concat, SoftMax, SoftMaxCrossEntropy, GraphConv, BatchNorm, \ GraphPool, GraphGather, WeightedError, Dropout, BatchNormalization GraphPool, GraphGather, WeightedError, Dropout, BatchNormalization, Stack from deepchem.models.tensorgraph.layers import L2Loss, Label, Weights, Feature from deepchem.models.tensorgraph.tensor_graph import TensorGraph from deepchem.trans import undo_transforms Loading Loading @@ -106,7 +106,10 @@ class WeaveTensorGraph(TensorGraph): self.labels_fd.append(label) cost = L2Loss(in_layers=[label, regression]) costs.append(cost) if self.mode == "classification": all_cost = Concat(in_layers=costs, axis=1) elif self.mode == "regression": all_cost = Stack(in_layers=costs, axis=1) self.weights = Weights(shape=(None, self.n_tasks)) loss = WeightedError(in_layers=[all_cost, self.weights]) self.set_loss(loss) Loading Loading @@ -259,7 +262,7 @@ class DTNNTensorGraph(TensorGraph): cost = L2Loss(in_layers=[label, regression]) costs.append(cost) all_cost = Concat(in_layers=costs, axis=1) all_cost = Stack(in_layers=costs, axis=1) self.weights = Weights(shape=(None, self.n_tasks)) loss = WeightedError(in_layers=[all_cost, self.weights]) self.set_loss(loss) Loading Loading @@ -404,8 +407,10 @@ class DAGTensorGraph(TensorGraph): self.labels_fd.append(label) cost = L2Loss(in_layers=[label, regression]) costs.append(cost) if self.mode == "classification": all_cost = Concat(in_layers=costs, axis=1) elif self.mode == "regression": all_cost = Stack(in_layers=costs, axis=1) self.weights = Weights(shape=(None, self.n_tasks)) loss = WeightedError(in_layers=[all_cost, self.weights]) self.set_loss(loss) Loading Loading @@ -545,8 +550,10 @@ class GraphConvTensorGraph(TensorGraph): self.my_labels.append(label) cost = L2Loss(in_layers=[label, regression]) costs.append(cost) if self.mode == "classification": entropy = Concat(in_layers=costs, axis=-1) elif self.mode == "regression": entropy = Stack(in_layers=costs, axis=1) self.my_task_weights = Weights(shape=(None, self.n_tasks)) loss = WeightedError(in_layers=[entropy, self.my_task_weights]) self.set_loss(loss) Loading Loading @@ -604,7 +611,7 @@ class GraphConvTensorGraph(TensorGraph): if not self.built: self.build() return self.evaluate_generator( self.default_generator(dataset), self.default_generator(dataset, predict=True), metrics, labels=self.my_labels, weights=[self.my_task_weights]) Loading
deepchem/models/tensorgraph/models/symmetry_function_regression.py +3 −3 Original line number Diff line number Diff line Loading @@ -8,7 +8,7 @@ Created on Thu Jul 6 20:31:47 2017 import numpy as np import tensorflow as tf from deepchem.models.tensorgraph.layers import Dense, Concat, WeightedError from deepchem.models.tensorgraph.layers import Dense, Concat, WeightedError, Stack from deepchem.models.tensorgraph.layers import L2Loss, Label, Weights, Feature from deepchem.models.tensorgraph.tensor_graph import TensorGraph from deepchem.models.tensorgraph.graph_layers import DTNNEmbedding Loading Loading @@ -69,7 +69,7 @@ class BPSymmetryFunctionRegression(TensorGraph): cost = L2Loss(in_layers=[label, output]) costs.append(cost) all_cost = Concat(in_layers=costs, axis=0) all_cost = Stack(in_layers=costs, axis=1) self.weights = Weights(shape=(None, self.n_tasks)) loss = WeightedError(in_layers=[all_cost, self.weights]) self.set_loss(loss) Loading Loading @@ -158,7 +158,7 @@ class ANIRegression(TensorGraph): cost = L2Loss(in_layers=[label, output]) costs.append(cost) all_cost = Concat(in_layers=costs, axis=0) all_cost = Stack(in_layers=costs, axis=1) self.weights = Weights(shape=(None, self.n_tasks)) loss = WeightedError(in_layers=[all_cost, self.weights]) self.set_loss(loss) Loading
deepchem/models/tensorgraph/optimizers.py 0 → 100644 +170 −0 Original line number Diff line number Diff line """Optimizers and related classes for use with TensorGraph.""" import tensorflow as tf class Optimizer(object): """An algorithm for optimizing a TensorGraph based model. This is an abstract class. Subclasses represent specific optimization algorithms. """ def _create_optimizer(self, global_step): """Construct the TensorFlow optimizer. Parameters ---------- global_step: tensor a tensor containing the global step index during optimization, used for learning rate decay Returns ------- a new TensorFlow optimizer implementing the algorithm """ raise NotImplemented("Subclasses must implement this") class LearningRateSchedule(object): """A schedule for changing the learning rate over the course of optimization. This is an abstract class. Subclasses represent specific schedules. """ def _create_tensor(self, global_step): """Construct a tensor that equals the learning rate. Parameters ---------- global_step: tensor a tensor containing the global step index during optimization Returns ------- a tensor that equals the learning rate """ raise NotImplemented("Subclasses must implement this") class Adam(Optimizer): """The Adam optimization algorithm.""" def __init__(self, learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-08): """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 """ self.learning_rate = learning_rate self.beta1 = beta1 self.beta2 = beta2 self.epsilon = epsilon def _create_optimizer(self, global_step): if isinstance(self.learning_rate, LearningRateSchedule): learning_rate = self.learning_rate._create_tensor(global_step) else: learning_rate = self.learning_rate return tf.train.AdamOptimizer( learning_rate=learning_rate, beta1=self.beta1, beta2=self.beta2, epsilon=self.epsilon) class GradientDescent(Optimizer): """The gradient descent optimization algorithm.""" def __init__(self, learning_rate=0.001): """Construct a gradient descent optimizer. Parameters ---------- learning_rate: float or LearningRateSchedule the learning rate to use for optimization """ self.learning_rate = learning_rate def _create_optimizer(self, global_step): if isinstance(self.learning_rate, LearningRateSchedule): learning_rate = self.learning_rate._create_tensor(global_step) else: learning_rate = self.learning_rate return tf.train.GradientDescentOptimizer(learning_rate=learning_rate) class ExponentialDecay(LearningRateSchedule): """A learning rate that decreases exponentially with the number of training steps.""" def __init__(self, initial_rate, decay_rate, decay_steps, staircase=True): """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 """ self.initial_rate = initial_rate self.decay_rate = decay_rate self.decay_steps = decay_steps self.staircase = staircase def _create_tensor(self, global_step): return tf.train.exponential_decay( learning_rate=self.initial_rate, global_step=global_step, decay_rate=self.decay_rate, decay_steps=self.decay_steps, staircase=self.staircase) class PolynomialDecay(LearningRateSchedule): """A learning rate that decreases from an initial value to a final value over a fixed number of training steps.""" def __init__(self, initial_rate, final_rate, decay_steps, power=1.0): """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 """ self.initial_rate = initial_rate self.final_rate = final_rate self.decay_steps = decay_steps self.power = power def _create_tensor(self, global_step): return tf.train.polynomial_decay( learning_rate=self.initial_rate, end_learning_rate=self.final_rate, global_step=global_step, decay_steps=self.decay_steps, power=self.power)
