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/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) deepchem/models/tensorgraph/tensor_graph.py +8 −12 Original line number Diff line number Diff line Loading @@ -13,6 +13,7 @@ from deepchem.data import NumpyDataset from deepchem.metrics import to_one_hot, from_one_hot from deepchem.models.models import Model from deepchem.models.tensorgraph.layers import InputFifoQueue, Label, Feature, Weights from deepchem.models.tensorgraph.optimizers import Adam from deepchem.trans import undo_transforms from deepchem.utils.evaluate import GeneratorEvaluator from deepchem.feat.graph_features import ConvMolFeaturizer Loading Loading @@ -54,6 +55,8 @@ class TensorGraph(Model): graph: tensorflow.Graph the Graph in which to create Tensorflow objects. If None, a new Graph is created. learning_rate: float or LearningRateSchedule the learning rate to use for optimization kwargs """ Loading @@ -67,12 +70,8 @@ class TensorGraph(Model): self.loss = None self.built = False self.queue_installed = False self.optimizer = TFWrapper( tf.train.AdamOptimizer, learning_rate=learning_rate, beta1=0.9, beta2=0.999, epsilon=1e-7) self.optimizer = Adam( learning_rate=learning_rate, beta1=0.9, beta2=0.999, epsilon=1e-7) # Singular place to hold Tensor objects which don't serialize # These have to be reconstructed on restoring from pickle Loading Loading @@ -425,11 +424,7 @@ class TensorGraph(Model): self.outputs.append(layer) def set_optimizer(self, optimizer): """Set the optimizer to use for fitting. The argument should be a callable object (most often a TFWrapper) that constructs a Tensorflow optimizer when called. """ """Set the optimizer to use for fitting.""" self.optimizer = optimizer def get_pickling_errors(self, obj, seen=None): Loading Loading @@ -562,7 +557,8 @@ class TensorGraph(Model): elif obj == "FileWriter": self.tensor_objects['FileWriter'] = tf.summary.FileWriter(self.model_dir) elif obj == 'Optimizer': self.tensor_objects['Optimizer'] = self.optimizer() self.tensor_objects['Optimizer'] = self.optimizer._create_optimizer( self._get_tf('GlobalStep')) elif obj == 'train_op': self.tensor_objects['train_op'] = self._get_tf('Optimizer').minimize( self.loss.out_tensor, global_step=self._get_tf('GlobalStep')) Loading deepchem/models/tensorgraph/tests/test_optimizers.py 0 → 100644 +41 −0 Original line number Diff line number Diff line import deepchem.models.tensorgraph.optimizers as optimizers import tensorflow as tf from tensorflow.python.framework import test_util class TestLayers(test_util.TensorFlowTestCase): """Test optimizers and related classes.""" def test_adam(self): """Test creating an Adam optimizer.""" opt = optimizers.Adam(learning_rate=0.01) with self.test_session() as sess: global_step = tf.Variable(0) tfopt = opt._create_optimizer(global_step) assert isinstance(tfopt, tf.train.AdamOptimizer) def test_gradient_descent(self): """Test creating a Gradient Descent optimizer.""" opt = optimizers.GradientDescent(learning_rate=0.01) with self.test_session() as sess: global_step = tf.Variable(0) tfopt = opt._create_optimizer(global_step) assert isinstance(tfopt, tf.train.GradientDescentOptimizer) def test_exponential_decay(self): """Test creating an optimizer with an exponentially decaying learning rate.""" rate = optimizers.ExponentialDecay( initial_rate=0.001, decay_rate=0.99, decay_steps=10000) opt = optimizers.Adam(learning_rate=rate) with self.test_session() as sess: global_step = tf.Variable(0) tfopt = opt._create_optimizer(global_step) def test_polynomial_decay(self): """Test creating an optimizer with a polynomially decaying learning rate.""" rate = optimizers.PolynomialDecay( initial_rate=0.001, final_rate=0.0001, decay_steps=10000) opt = optimizers.Adam(learning_rate=rate) with self.test_session() as sess: global_step = tf.Variable(0) tfopt = opt._create_optimizer(global_step) deepchem/models/tensorgraph/tests/test_tensor_graph.py +5 −9 Original line number Diff line number Diff line Loading @@ -12,7 +12,8 @@ from deepchem.data.datasets import Databag from deepchem.models.tensorgraph.layers import Dense, SoftMaxCrossEntropy, ReduceMean, SoftMax from deepchem.models.tensorgraph.layers import Feature, Label from deepchem.models.tensorgraph.layers import ReduceSquareDifference from deepchem.models.tensorgraph.tensor_graph import TensorGraph, TFWrapper from deepchem.models.tensorgraph.tensor_graph import TensorGraph from deepchem.models.tensorgraph.optimizers import GradientDescent, ExponentialDecay class TestTensorGraph(unittest.TestCase): Loading Loading @@ -179,14 +180,9 @@ class TestTensorGraph(unittest.TestCase): tg.add_output(output) tg.set_loss(loss) global_step = tg.get_global_step() def optimizer_function(): starter_learning_rate = 0.1 learning_rate = tf.train.exponential_decay( starter_learning_rate, global_step, 100000, 0.96, staircase=True) return tf.train.GradientDescentOptimizer(learning_rate) tg.set_optimizer(TFWrapper(optimizer_function)) learning_rate = ExponentialDecay( initial_rate=0.1, decay_rate=0.96, decay_steps=100000) tg.set_optimizer(GradientDescent(learning_rate=learning_rate)) tg.fit(dataset, nb_epoch=1000) prediction = np.squeeze(tg.predict_proba_on_batch(X)) tg.save() Loading 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/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)
deepchem/models/tensorgraph/tensor_graph.py +8 −12 Original line number Diff line number Diff line Loading @@ -13,6 +13,7 @@ from deepchem.data import NumpyDataset from deepchem.metrics import to_one_hot, from_one_hot from deepchem.models.models import Model from deepchem.models.tensorgraph.layers import InputFifoQueue, Label, Feature, Weights from deepchem.models.tensorgraph.optimizers import Adam from deepchem.trans import undo_transforms from deepchem.utils.evaluate import GeneratorEvaluator from deepchem.feat.graph_features import ConvMolFeaturizer Loading Loading @@ -54,6 +55,8 @@ class TensorGraph(Model): graph: tensorflow.Graph the Graph in which to create Tensorflow objects. If None, a new Graph is created. learning_rate: float or LearningRateSchedule the learning rate to use for optimization kwargs """ Loading @@ -67,12 +70,8 @@ class TensorGraph(Model): self.loss = None self.built = False self.queue_installed = False self.optimizer = TFWrapper( tf.train.AdamOptimizer, learning_rate=learning_rate, beta1=0.9, beta2=0.999, epsilon=1e-7) self.optimizer = Adam( learning_rate=learning_rate, beta1=0.9, beta2=0.999, epsilon=1e-7) # Singular place to hold Tensor objects which don't serialize # These have to be reconstructed on restoring from pickle Loading Loading @@ -425,11 +424,7 @@ class TensorGraph(Model): self.outputs.append(layer) def set_optimizer(self, optimizer): """Set the optimizer to use for fitting. The argument should be a callable object (most often a TFWrapper) that constructs a Tensorflow optimizer when called. """ """Set the optimizer to use for fitting.""" self.optimizer = optimizer def get_pickling_errors(self, obj, seen=None): Loading Loading @@ -562,7 +557,8 @@ class TensorGraph(Model): elif obj == "FileWriter": self.tensor_objects['FileWriter'] = tf.summary.FileWriter(self.model_dir) elif obj == 'Optimizer': self.tensor_objects['Optimizer'] = self.optimizer() self.tensor_objects['Optimizer'] = self.optimizer._create_optimizer( self._get_tf('GlobalStep')) elif obj == 'train_op': self.tensor_objects['train_op'] = self._get_tf('Optimizer').minimize( self.loss.out_tensor, global_step=self._get_tf('GlobalStep')) Loading
deepchem/models/tensorgraph/tests/test_optimizers.py 0 → 100644 +41 −0 Original line number Diff line number Diff line import deepchem.models.tensorgraph.optimizers as optimizers import tensorflow as tf from tensorflow.python.framework import test_util class TestLayers(test_util.TensorFlowTestCase): """Test optimizers and related classes.""" def test_adam(self): """Test creating an Adam optimizer.""" opt = optimizers.Adam(learning_rate=0.01) with self.test_session() as sess: global_step = tf.Variable(0) tfopt = opt._create_optimizer(global_step) assert isinstance(tfopt, tf.train.AdamOptimizer) def test_gradient_descent(self): """Test creating a Gradient Descent optimizer.""" opt = optimizers.GradientDescent(learning_rate=0.01) with self.test_session() as sess: global_step = tf.Variable(0) tfopt = opt._create_optimizer(global_step) assert isinstance(tfopt, tf.train.GradientDescentOptimizer) def test_exponential_decay(self): """Test creating an optimizer with an exponentially decaying learning rate.""" rate = optimizers.ExponentialDecay( initial_rate=0.001, decay_rate=0.99, decay_steps=10000) opt = optimizers.Adam(learning_rate=rate) with self.test_session() as sess: global_step = tf.Variable(0) tfopt = opt._create_optimizer(global_step) def test_polynomial_decay(self): """Test creating an optimizer with a polynomially decaying learning rate.""" rate = optimizers.PolynomialDecay( initial_rate=0.001, final_rate=0.0001, decay_steps=10000) opt = optimizers.Adam(learning_rate=rate) with self.test_session() as sess: global_step = tf.Variable(0) tfopt = opt._create_optimizer(global_step)
deepchem/models/tensorgraph/tests/test_tensor_graph.py +5 −9 Original line number Diff line number Diff line Loading @@ -12,7 +12,8 @@ from deepchem.data.datasets import Databag from deepchem.models.tensorgraph.layers import Dense, SoftMaxCrossEntropy, ReduceMean, SoftMax from deepchem.models.tensorgraph.layers import Feature, Label from deepchem.models.tensorgraph.layers import ReduceSquareDifference from deepchem.models.tensorgraph.tensor_graph import TensorGraph, TFWrapper from deepchem.models.tensorgraph.tensor_graph import TensorGraph from deepchem.models.tensorgraph.optimizers import GradientDescent, ExponentialDecay class TestTensorGraph(unittest.TestCase): Loading Loading @@ -179,14 +180,9 @@ class TestTensorGraph(unittest.TestCase): tg.add_output(output) tg.set_loss(loss) global_step = tg.get_global_step() def optimizer_function(): starter_learning_rate = 0.1 learning_rate = tf.train.exponential_decay( starter_learning_rate, global_step, 100000, 0.96, staircase=True) return tf.train.GradientDescentOptimizer(learning_rate) tg.set_optimizer(TFWrapper(optimizer_function)) learning_rate = ExponentialDecay( initial_rate=0.1, decay_rate=0.96, decay_steps=100000) tg.set_optimizer(GradientDescent(learning_rate=learning_rate)) tg.fit(dataset, nb_epoch=1000) prediction = np.squeeze(tg.predict_proba_on_batch(X)) tg.save() Loading
