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/data/datasets.py +17 −0 Original line number Diff line number Diff line Loading @@ -396,6 +396,23 @@ class NumpyDataset(Dataset): ids = self.ids[indices] return NumpyDataset(X, y, w, ids) @staticmethod def from_DiskDataset(ds): """ Parameters ---------- ds : DiskDataset DiskDataset to transorm to NumpyDataset Returns ------- NumpyDataset Data of ds as NumpyDataset """ return NumpyDataset(ds.X, ds.y, ds.w, ds.ids) class DiskDataset(Dataset): """ Loading deepchem/hyper/__init__.py +16 −17 Original line number Diff line number Diff line Loading @@ -12,6 +12,7 @@ from operator import mul from deepchem.utils.evaluate import Evaluator from deepchem.utils.save import log class HyperparamOpt(object): """ Provides simple hyperparameter search capabilities. Loading @@ -23,8 +24,13 @@ class HyperparamOpt(object): # TODO(rbharath): This function is complicated and monolithic. Is there a nice # way to refactor this? def hyperparam_search(self, params_dict, train_dataset, valid_dataset, output_transformers, metric, use_max=True, def hyperparam_search(self, params_dict, train_dataset, valid_dataset, output_transformers, metric, use_max=True, logdir=None): """Perform hyperparams search according to params_dict. Loading @@ -40,8 +46,6 @@ class HyperparamOpt(object): number_combinations = reduce(mul, [len(vals) for vals in hyperparam_vals]) valid_csv_out = tempfile.NamedTemporaryFile() valid_stats_out = tempfile.NamedTemporaryFile() if use_max: best_validation_score = -np.inf else: Loading @@ -49,10 +53,10 @@ class HyperparamOpt(object): best_hyperparams = None best_model, best_model_dir = None, None all_scores = {} for ind, hyperparameter_tuple in enumerate(itertools.product(*hyperparam_vals)): for ind, hyperparameter_tuple in enumerate( itertools.product(*hyperparam_vals)): model_params = {} log("Fitting model %d/%d" % (ind+1, number_combinations), self.verbose) log("Fitting model %d/%d" % (ind + 1, number_combinations), self.verbose) for hyperparam, hyperparam_val in zip(hyperparams, hyperparameter_tuple): model_params[hyperparam] = hyperparam_val log("hyperparameters: %s" % str(model_params), self.verbose) Loading @@ -75,8 +79,7 @@ class HyperparamOpt(object): model.save() evaluator = Evaluator(model, valid_dataset, output_transformers) multitask_scores = evaluator.compute_model_performance( [metric], valid_csv_out.name, valid_stats_out.name) multitask_scores = evaluator.compute_model_performance([metric]) valid_score = multitask_scores[metric.name] all_scores[str(hyperparameter_tuple)] = valid_score Loading @@ -92,8 +95,8 @@ class HyperparamOpt(object): shutil.rmtree(model_dir) log("Model %d/%d, Metric %s, Validation set %s: %f" % (ind+1, number_combinations, metric.name, ind, valid_score), self.verbose) (ind + 1, number_combinations, metric.name, ind, valid_score), self.verbose) log("\tbest_validation_score so far: %f" % best_validation_score, self.verbose) if best_model is None: Loading @@ -101,14 +104,10 @@ class HyperparamOpt(object): # arbitrarily return last model best_model, best_hyperparams = model, hyperparameter_tuple return best_model, best_hyperparams, all_scores train_csv_out = tempfile.NamedTemporaryFile() train_stats_out = tempfile.NamedTemporaryFile() train_evaluator = Evaluator(best_model, train_dataset, output_transformers) multitask_scores = train_evaluator.compute_model_performance( [metric], train_csv_out.name, train_stats_out.name) multitask_scores = train_evaluator.compute_model_performance([metric]) train_score = multitask_scores[metric.name] log("Best hyperparameters: %s" % str(best_hyperparams), self.verbose) log("Best hyperparameters: %s" % str(best_hyperparams), self.verbose) log("train_score: %f" % train_score, self.verbose) log("validation_score: %f" % best_validation_score, self.verbose) return best_model, best_hyperparams, all_scores 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 +9 −13 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 @@ -234,7 +233,7 @@ class TensorGraph(Model): feed_dict[self.features[0]] = X_b if len(self.task_weights) == 1 and w_b is not None and not predict: feed_dict[self.task_weights[0]] = w_b for (inital_state, zero_state) in zip(self.rnn_initial_states, for (initial_state, zero_state) in zip(self.rnn_initial_states, self.rnn_zero_states): feed_dict[initial_state] = zero_state yield feed_dict 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 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/data/datasets.py +17 −0 Original line number Diff line number Diff line Loading @@ -396,6 +396,23 @@ class NumpyDataset(Dataset): ids = self.ids[indices] return NumpyDataset(X, y, w, ids) @staticmethod def from_DiskDataset(ds): """ Parameters ---------- ds : DiskDataset DiskDataset to transorm to NumpyDataset Returns ------- NumpyDataset Data of ds as NumpyDataset """ return NumpyDataset(ds.X, ds.y, ds.w, ds.ids) class DiskDataset(Dataset): """ Loading
deepchem/hyper/__init__.py +16 −17 Original line number Diff line number Diff line Loading @@ -12,6 +12,7 @@ from operator import mul from deepchem.utils.evaluate import Evaluator from deepchem.utils.save import log class HyperparamOpt(object): """ Provides simple hyperparameter search capabilities. Loading @@ -23,8 +24,13 @@ class HyperparamOpt(object): # TODO(rbharath): This function is complicated and monolithic. Is there a nice # way to refactor this? def hyperparam_search(self, params_dict, train_dataset, valid_dataset, output_transformers, metric, use_max=True, def hyperparam_search(self, params_dict, train_dataset, valid_dataset, output_transformers, metric, use_max=True, logdir=None): """Perform hyperparams search according to params_dict. Loading @@ -40,8 +46,6 @@ class HyperparamOpt(object): number_combinations = reduce(mul, [len(vals) for vals in hyperparam_vals]) valid_csv_out = tempfile.NamedTemporaryFile() valid_stats_out = tempfile.NamedTemporaryFile() if use_max: best_validation_score = -np.inf else: Loading @@ -49,10 +53,10 @@ class HyperparamOpt(object): best_hyperparams = None best_model, best_model_dir = None, None all_scores = {} for ind, hyperparameter_tuple in enumerate(itertools.product(*hyperparam_vals)): for ind, hyperparameter_tuple in enumerate( itertools.product(*hyperparam_vals)): model_params = {} log("Fitting model %d/%d" % (ind+1, number_combinations), self.verbose) log("Fitting model %d/%d" % (ind + 1, number_combinations), self.verbose) for hyperparam, hyperparam_val in zip(hyperparams, hyperparameter_tuple): model_params[hyperparam] = hyperparam_val log("hyperparameters: %s" % str(model_params), self.verbose) Loading @@ -75,8 +79,7 @@ class HyperparamOpt(object): model.save() evaluator = Evaluator(model, valid_dataset, output_transformers) multitask_scores = evaluator.compute_model_performance( [metric], valid_csv_out.name, valid_stats_out.name) multitask_scores = evaluator.compute_model_performance([metric]) valid_score = multitask_scores[metric.name] all_scores[str(hyperparameter_tuple)] = valid_score Loading @@ -92,8 +95,8 @@ class HyperparamOpt(object): shutil.rmtree(model_dir) log("Model %d/%d, Metric %s, Validation set %s: %f" % (ind+1, number_combinations, metric.name, ind, valid_score), self.verbose) (ind + 1, number_combinations, metric.name, ind, valid_score), self.verbose) log("\tbest_validation_score so far: %f" % best_validation_score, self.verbose) if best_model is None: Loading @@ -101,14 +104,10 @@ class HyperparamOpt(object): # arbitrarily return last model best_model, best_hyperparams = model, hyperparameter_tuple return best_model, best_hyperparams, all_scores train_csv_out = tempfile.NamedTemporaryFile() train_stats_out = tempfile.NamedTemporaryFile() train_evaluator = Evaluator(best_model, train_dataset, output_transformers) multitask_scores = train_evaluator.compute_model_performance( [metric], train_csv_out.name, train_stats_out.name) multitask_scores = train_evaluator.compute_model_performance([metric]) train_score = multitask_scores[metric.name] log("Best hyperparameters: %s" % str(best_hyperparams), self.verbose) log("Best hyperparameters: %s" % str(best_hyperparams), self.verbose) log("train_score: %f" % train_score, self.verbose) log("validation_score: %f" % best_validation_score, self.verbose) return best_model, best_hyperparams, all_scores
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 +9 −13 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 @@ -234,7 +233,7 @@ class TensorGraph(Model): feed_dict[self.features[0]] = X_b if len(self.task_weights) == 1 and w_b is not None and not predict: feed_dict[self.task_weights[0]] = w_b for (inital_state, zero_state) in zip(self.rnn_initial_states, for (initial_state, zero_state) in zip(self.rnn_initial_states, self.rnn_zero_states): feed_dict[initial_state] = zero_state yield feed_dict 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
