text
stringlengths
81
112k
Starts a thread for reading output from FFMPEG. The thread reads consecutive chunks from the stream and saves them in the given list. Args: stream: output stream of the FFMPEG process. chunks: list to save output chunks to. Returns: Thread def _start_reader_thread(self, stream, chunks): """Starts a thread for reading output from FFMPEG. The thread reads consecutive chunks from the stream and saves them in the given list. Args: stream: output stream of the FFMPEG process. chunks: list to save output chunks to. Returns: Thread """ import io # pylint: disable=g-import-not-at-top import threading # pylint: disable=g-import-not-at-top def target(): while True: chunk = stream.read(io.DEFAULT_BUFFER_SIZE) if not chunk: break chunks.append(chunk) thread = threading.Thread(target=target) thread.start() return thread
Finishes transconding and returns the video. Returns: bytes Raises: IOError: in case of transcoding error. def finish(self): """Finishes transconding and returns the video. Returns: bytes Raises: IOError: in case of transcoding error. """ if self.proc is None: return None self.proc.stdin.close() for thread in (self._out_thread, self._err_thread): thread.join() (out, err) = [ b"".join(chunks) for chunks in (self._out_chunks, self._err_chunks) ] self.proc.stdout.close() self.proc.stderr.close() if self.proc.returncode: err = "\n".join([" ".join(self.cmd), err.decode("utf8")]) raise IOError(err) del self.proc self.proc = None return out
Validates flags are set to acceptable values. def validate_flags(): """Validates flags are set to acceptable values.""" if FLAGS.cloud_mlengine_model_name: assert not FLAGS.server assert not FLAGS.servable_name else: assert FLAGS.server assert FLAGS.servable_name
Returns a request function. def make_request_fn(): """Returns a request function.""" if FLAGS.cloud_mlengine_model_name: request_fn = serving_utils.make_cloud_mlengine_request_fn( credentials=GoogleCredentials.get_application_default(), model_name=FLAGS.cloud_mlengine_model_name, version=FLAGS.cloud_mlengine_model_version) else: request_fn = serving_utils.make_grpc_request_fn( servable_name=FLAGS.servable_name, server=FLAGS.server, timeout_secs=FLAGS.timeout_secs) return request_fn
Convnet that encodes inputs into mean and std of a gaussian. Args: inputs: 5-D Tensor, shape (batch_size, num_frames, width, height, channels) n_layers: Number of layers. Returns: z_mu: Mean of the latent gaussians. z_log_var: log(var) of the latent gaussians. Raises: ValueError: If inputs is not a 5-D tensor or not float32. def encoder(self, inputs, n_layers=3): """Convnet that encodes inputs into mean and std of a gaussian. Args: inputs: 5-D Tensor, shape (batch_size, num_frames, width, height, channels) n_layers: Number of layers. Returns: z_mu: Mean of the latent gaussians. z_log_var: log(var) of the latent gaussians. Raises: ValueError: If inputs is not a 5-D tensor or not float32. """ latent_dims = self.hparams.z_dim shape_as_list = inputs.shape.as_list() if len(shape_as_list) != 5: raise ValueError("Expected inputs to be a 5-D, got %d" % len(shape_as_list)) if inputs.dtype != tf.float32: raise ValueError("Expected dtype tf.float32, got %s" % inputs.dtype) # Flatten (N,T,W,H,C) into (NT,W,H,C) batch_size, _ = shape_as_list[:2] inputs = tf.reshape(inputs, [-1] + list(inputs.shape)[2:]) n_filters = 64 rectified = None # Applies 3 layer conv-net with padding, instance normalization # and leaky relu as per the encoder in # https://github.com/alexlee-gk/video_prediction padding = [[0, 0], [1, 1], [1, 1], [0, 0]] for i in range(n_layers): with tf.variable_scope("layer_%d" % (i + 1)): n_filters *= 2**i if i: padded = tf.pad(rectified, padding) else: padded = tf.pad(inputs, padding) convolved = tf.layers.conv2d(padded, filters=n_filters, kernel_size=4, strides=2, padding="VALID") normalized = tf.contrib.layers.instance_norm(convolved) rectified = tf.nn.leaky_relu(normalized, alpha=0.2) # Mean pooling across all spatial dimensions. pooled = tf.nn.avg_pool( rectified, [1] + rectified.shape[1:3].as_list() + [1], strides=[1, 1, 1, 1], padding="VALID") squeezed = tf.squeeze(pooled, [1, 2]) # Down-project and output the mean and log of the standard deviation of # the latents. with tf.variable_scope("z_mu"): z_mu = tf.layers.dense(squeezed, latent_dims) with tf.variable_scope("z_log_sigma_sq"): z_log_var = tf.layers.dense(squeezed, latent_dims) z_log_var = tf.clip_by_value(z_log_var, -10, 10) # Reshape to (batch_size X num_frames X latent_dims) z_mu = tf.reshape(z_mu, (batch_size, -1, latent_dims)) z_log_var = tf.reshape( z_log_var, (batch_size, -1, latent_dims)) return z_mu, z_log_var
Get expected fully connected shape after a series of convolutions. def get_fc_dimensions(self, strides, kernel_sizes): """Get expected fully connected shape after a series of convolutions.""" output_height, output_width, _ = self.hparams.problem.frame_shape output_steps = self.hparams.video_num_target_frames output_shape = np.array([output_steps, output_height, output_width]) for curr_stride, kernel_size in zip(strides, kernel_sizes): output_shape = self.expected_output_shape( output_shape, np.array(curr_stride), 1, kernel_size) return np.prod(output_shape) * self.hparams.num_discriminator_filters * 8
3-D SNGAN discriminator. Args: frames: a list of batch-major tensors indexed by time. Returns: logits: 1-D Tensor with shape=batch_size. Positive logits imply that the discriminator thinks that it belongs to the true class. def discriminator(self, frames): """3-D SNGAN discriminator. Args: frames: a list of batch-major tensors indexed by time. Returns: logits: 1-D Tensor with shape=batch_size. Positive logits imply that the discriminator thinks that it belongs to the true class. """ ndf = self.hparams.num_discriminator_filters frames = tf.stack(frames) # Switch from time-major axis to batch-major axis. frames = common_video.swap_time_and_batch_axes(frames) # 3-D Conv-net mapping inputs to activations. num_outputs = [ndf, ndf*2, ndf*2, ndf*4, ndf*4, ndf*8, ndf*8] kernel_sizes = [3, 4, 3, 4, 3, 4, 3] strides = [[1, 1, 1], [1, 2, 2], [1, 1, 1], [1, 2, 2], [1, 1, 1], [2, 2, 2], [1, 1, 1]] names = ["video_sn_conv0_0", "video_sn_conv0_1", "video_sn_conv1_0", "video_sn_conv1_1", "video_sn_conv2_0", "video_sn_conv2_1", "video_sn_conv3_0"] iterable = zip(num_outputs, kernel_sizes, strides, names) activations = frames for num_filters, kernel_size, stride, name in iterable: activations = self.pad_conv3d_lrelu(activations, num_filters, kernel_size, stride, name) num_fc_dimensions = self.get_fc_dimensions(strides, kernel_sizes) activations = tf.reshape(activations, (-1, num_fc_dimensions)) return tf.squeeze(tf.layers.dense(activations, 1))
Performs the discriminator step in computing the GAN loss. Applies stop-gradient to the generated frames while computing the discriminator loss to make sure that the gradients are not back-propagated to the generator. This makes sure that only the discriminator is updated. Args: true_frames: True outputs gen_frames: Generated frames. Returns: d_loss: Loss component due to the discriminator. def d_step(self, true_frames, gen_frames): """Performs the discriminator step in computing the GAN loss. Applies stop-gradient to the generated frames while computing the discriminator loss to make sure that the gradients are not back-propagated to the generator. This makes sure that only the discriminator is updated. Args: true_frames: True outputs gen_frames: Generated frames. Returns: d_loss: Loss component due to the discriminator. """ hparam_to_disc_loss = { "least_squares": gan_losses.least_squares_discriminator_loss, "cross_entropy": gan_losses.modified_discriminator_loss, "wasserstein": gan_losses.wasserstein_discriminator_loss} # Concat across batch-axis. _, batch_size, _, _, _ = common_layers.shape_list(true_frames) all_frames = tf.concat( [true_frames, tf.stop_gradient(gen_frames)], axis=1) all_logits = self.discriminator(all_frames) true_logits, fake_logits_stop = \ all_logits[:batch_size], all_logits[batch_size:] mean_true_logits = tf.reduce_mean(true_logits) tf.summary.scalar("mean_true_logits", mean_true_logits) mean_fake_logits_stop = tf.reduce_mean(fake_logits_stop) tf.summary.scalar("mean_fake_logits_stop", mean_fake_logits_stop) discriminator_loss_func = hparam_to_disc_loss[self.hparams.gan_loss] gan_d_loss = discriminator_loss_func( discriminator_real_outputs=true_logits, discriminator_gen_outputs=fake_logits_stop, add_summaries=True) return gan_d_loss, true_logits, fake_logits_stop
Performs the generator step in computing the GAN loss. Args: gen_frames: Generated frames fake_logits_stop: Logits corresponding to the generated frames as per the discriminator. Assumed to have a stop-gradient term. Returns: gan_g_loss_pos_d: Loss. gan_g_loss_neg_d: -gan_g_loss_pos_d but with a stop gradient on generator. def g_step(self, gen_frames, fake_logits_stop): """Performs the generator step in computing the GAN loss. Args: gen_frames: Generated frames fake_logits_stop: Logits corresponding to the generated frames as per the discriminator. Assumed to have a stop-gradient term. Returns: gan_g_loss_pos_d: Loss. gan_g_loss_neg_d: -gan_g_loss_pos_d but with a stop gradient on generator. """ hparam_to_gen_loss = { "least_squares": gan_losses.least_squares_generator_loss, "cross_entropy": gan_losses.modified_generator_loss, "wasserstein": gan_losses.wasserstein_generator_loss } fake_logits = self.discriminator(gen_frames) mean_fake_logits = tf.reduce_mean(fake_logits) tf.summary.scalar("mean_fake_logits", mean_fake_logits) # Generator loss. # Using gan_g_loss_pos_d updates the discriminator as well. # To avoid this add gan_g_loss_neg_d = -gan_g_loss_pos_d # but with stop gradient on the generator. # This makes sure that the net gradient on the discriminator is zero and # net-gradient on the generator is just due to the gan_g_loss_pos_d. generator_loss_func = hparam_to_gen_loss[self.hparams.gan_loss] gan_g_loss_pos_d = generator_loss_func( discriminator_gen_outputs=fake_logits, add_summaries=True) gan_g_loss_neg_d = -generator_loss_func( discriminator_gen_outputs=fake_logits_stop, add_summaries=True) return gan_g_loss_pos_d, gan_g_loss_neg_d
Get the discriminator + generator loss at every step. This performs an 1:1 update of the discriminator and generator at every step. Args: true_frames: 5-D Tensor of shape (num_steps, batch_size, H, W, C) Assumed to be ground truth. gen_frames: 5-D Tensor of shape (num_steps, batch_size, H, W, C) Assumed to be fake. name: discriminator scope. Returns: loss: 0-D Tensor, with d_loss + g_loss def get_gan_loss(self, true_frames, gen_frames, name): """Get the discriminator + generator loss at every step. This performs an 1:1 update of the discriminator and generator at every step. Args: true_frames: 5-D Tensor of shape (num_steps, batch_size, H, W, C) Assumed to be ground truth. gen_frames: 5-D Tensor of shape (num_steps, batch_size, H, W, C) Assumed to be fake. name: discriminator scope. Returns: loss: 0-D Tensor, with d_loss + g_loss """ # D - STEP with tf.variable_scope("%s_discriminator" % name, reuse=tf.AUTO_REUSE): gan_d_loss, _, fake_logits_stop = self.d_step( true_frames, gen_frames) # G - STEP with tf.variable_scope("%s_discriminator" % name, reuse=True): gan_g_loss_pos_d, gan_g_loss_neg_d = self.g_step( gen_frames, fake_logits_stop) gan_g_loss = gan_g_loss_pos_d + gan_g_loss_neg_d tf.summary.scalar("gan_loss_%s" % name, gan_g_loss_pos_d + gan_d_loss) if self.hparams.gan_optimization == "joint": gan_loss = gan_g_loss + gan_d_loss else: curr_step = self.get_iteration_num() gan_loss = tf.cond( tf.logical_not(curr_step % 2 == 0), lambda: gan_g_loss, lambda: gan_d_loss) return gan_loss
Gets extra loss from VAE and GAN. def get_extra_loss(self, latent_means=None, latent_stds=None, true_frames=None, gen_frames=None): """Gets extra loss from VAE and GAN.""" if not self.is_training: return 0.0 vae_loss, d_vae_loss, d_gan_loss = 0.0, 0.0, 0.0 # Use sv2p's KL divergence computation. if self.hparams.use_vae: vae_loss = super(NextFrameSavpBase, self).get_extra_loss( latent_means=latent_means, latent_stds=latent_stds) if self.hparams.use_gan: # Strip out the first context_frames for the true_frames # Strip out the first context_frames - 1 for the gen_frames context_frames = self.hparams.video_num_input_frames true_frames = tf.stack( tf.unstack(true_frames, axis=0)[context_frames:]) # discriminator for VAE. if self.hparams.use_vae: gen_enc_frames = tf.stack( tf.unstack(gen_frames, axis=0)[context_frames-1:]) d_vae_loss = self.get_gan_loss(true_frames, gen_enc_frames, name="vae") # discriminator for GAN. gen_prior_frames = tf.stack( tf.unstack(self.gen_prior_video, axis=0)[context_frames-1:]) d_gan_loss = self.get_gan_loss(true_frames, gen_prior_frames, name="gan") return ( vae_loss + self.hparams.gan_loss_multiplier * d_gan_loss + self.hparams.gan_vae_loss_multiplier * d_vae_loss)
Pad, apply 3-D convolution and leaky relu. def pad_conv3d_lrelu(self, activations, n_filters, kernel_size, strides, scope): """Pad, apply 3-D convolution and leaky relu.""" padding = [[0, 0], [1, 1], [1, 1], [1, 1], [0, 0]] # tf.nn.conv3d accepts a list of 5 values for strides # with first and last value equal to 1 if isinstance(strides, numbers.Integral): strides = [strides] * 3 strides = [1] + strides + [1] # Filter_shape = [K, K, K, num_input, num_output] filter_shape = ( [kernel_size]*3 + activations.shape[-1:].as_list() + [n_filters]) with tf.variable_scope(scope, reuse=tf.AUTO_REUSE): conv_filter = tf.get_variable( "conv_filter", shape=filter_shape, initializer=tf.truncated_normal_initializer(stddev=0.02)) if self.hparams.use_spectral_norm: conv_filter, assign_op = common_layers.apply_spectral_norm(conv_filter) if self.is_training: tf.add_to_collection(tf.GraphKeys.UPDATE_OPS, assign_op) padded = tf.pad(activations, padding) convolved = tf.nn.conv3d( padded, conv_filter, strides=strides, padding="VALID") rectified = tf.nn.leaky_relu(convolved, alpha=0.2) return rectified
Weight-level magnitude pruning. def weight(w, sparsity): """Weight-level magnitude pruning.""" w_shape = common_layers.shape_list(w) k = int(np.prod(w_shape[:-1])) count = tf.to_int32(k * sparsity) mask = common_layers.weight_targeting(w, count) return (1 - mask) * w
Unit-level magnitude pruning. def unit(w, sparsity): """Unit-level magnitude pruning.""" w_shape = common_layers.shape_list(w) count = tf.to_int32(w_shape[-1] * sparsity) mask = common_layers.unit_targeting(w, count) return (1 - mask) * w
Prune the weights of a model and evaluate. def sparsify(sess, eval_model, pruning_strategy, pruning_params): """Prune the weights of a model and evaluate.""" weights = tf.trainable_variables() def should_prune(name): """Whether to prune a weight or not.""" in_whitelist = not pruning_params.white_list or any( e in name for e in pruning_params.white_list) in_blacklist = any(e in name for e in pruning_params.black_list) if pruning_params.white_list and not in_whitelist: return False elif in_blacklist: return False return True weights = [w for w in weights if should_prune(w.name)] tf.logging.info("Pruning weights: %s" % weights) unpruned_weights = sess.run(weights) reset_op = tf.no_op() for w, ow in zip(weights, unpruned_weights): op = tf.assign(w, ow) reset_op = tf.group(reset_op, op) for sparsity in pruning_params.sparsities: set_weights_op = tf.no_op() for w in weights: op = tf.assign(w, pruning_strategy(w, sparsity)) set_weights_op = tf.group(set_weights_op, op) sess.run(set_weights_op) acc = eval_model() tf.logging.info("\tPruning to sparsity = %f: acc = %f" % (sparsity, acc)) sess.run(reset_op)
Loads the configuration. def load_config(self): """Loads the configuration.""" config = dict([(key, value) for key, value in iteritems(self.options) if key in self.cfg.settings and value is not None]) for key, value in iteritems(config): self.cfg.set(key.lower(), value)
Set of hyperparameters. def ppo_base_v1(): """Set of hyperparameters.""" hparams = common_hparams.basic_params1() hparams.learning_rate_schedule = "constant" hparams.learning_rate_constant = 1e-4 hparams.clip_grad_norm = 0.5 hparams.weight_decay = 0 # If set, extends the LR warmup to all epochs except the final one. hparams.add_hparam("lr_decay_in_final_epoch", False) hparams.add_hparam("init_mean_factor", 0.1) hparams.add_hparam("init_logstd", 0.1) hparams.add_hparam("policy_layers", (100, 100)) hparams.add_hparam("value_layers", (100, 100)) hparams.add_hparam("clipping_coef", 0.2) hparams.add_hparam("gae_gamma", 0.99) hparams.add_hparam("gae_lambda", 0.95) hparams.add_hparam("entropy_loss_coef", 0.01) hparams.add_hparam("value_loss_coef", 1) hparams.add_hparam("optimization_epochs", 15) hparams.add_hparam("epoch_length", 200) hparams.add_hparam("epochs_num", 2000) hparams.add_hparam("eval_every_epochs", 10) hparams.add_hparam("save_models_every_epochs", 30) hparams.add_hparam("optimization_batch_size", 50) hparams.add_hparam("intrinsic_reward_scale", 0.) hparams.add_hparam("logits_clip", 0.0) hparams.add_hparam("dropout_ppo", 0.1) hparams.add_hparam("effective_num_agents", None) # TODO(afrozm): Clean this up, this is used in PPO learner to get modalities. hparams.add_hparam("policy_problem_name", "dummy_policy_problem") return hparams
Pong base parameters. def ppo_atari_base(): """Pong base parameters.""" hparams = ppo_discrete_action_base() hparams.learning_rate_constant = 1e-4 hparams.epoch_length = 200 hparams.gae_gamma = 0.985 hparams.gae_lambda = 0.985 hparams.entropy_loss_coef = 0.003 hparams.value_loss_coef = 1 hparams.optimization_epochs = 3 hparams.epochs_num = 1000 hparams.policy_network = "feed_forward_cnn_small_categorical_policy" hparams.clipping_coef = 0.2 hparams.optimization_batch_size = 20 hparams.clip_grad_norm = 0.5 return hparams
Parameters based on the original PPO paper. def ppo_original_params(): """Parameters based on the original PPO paper.""" hparams = ppo_atari_base() hparams.learning_rate_constant = 2.5e-4 hparams.gae_gamma = 0.99 hparams.gae_lambda = 0.95 hparams.clipping_coef = 0.1 hparams.value_loss_coef = 1 hparams.entropy_loss_coef = 0.01 hparams.eval_every_epochs = 200 hparams.dropout_ppo = 0.1 # The parameters below are modified to accommodate short epoch_length (which # is needed for model based rollouts). hparams.epoch_length = 50 hparams.optimization_batch_size = 20 return hparams
Atari parameters with world model as policy. def ppo_original_world_model(): """Atari parameters with world model as policy.""" hparams = ppo_original_params() hparams.policy_network = "next_frame_basic_deterministic" hparams_keys = hparams.values().keys() video_hparams = basic_deterministic_params.next_frame_basic_deterministic() for (name, value) in six.iteritems(video_hparams.values()): if name in hparams_keys: hparams.set_hparam(name, value) else: hparams.add_hparam(name, value) # Mostly to avoid decaying WM params when training the policy. hparams.weight_decay = 0 return hparams
Atari parameters with world model as policy. def ppo_tiny_world_model(): """Atari parameters with world model as policy.""" hparams = ppo_original_params() hparams.policy_network = "next_frame_basic_deterministic" hparams_keys = hparams.values().keys() video_hparams = basic_deterministic_params.next_frame_tiny() for (name, value) in six.iteritems(video_hparams.values()): if name in hparams_keys: hparams.set_hparam(name, value) else: hparams.add_hparam(name, value) hparams.weight_decay = 0 return hparams
Atari parameters with stochastic discrete world model as policy. def ppo_original_world_model_stochastic_discrete(): """Atari parameters with stochastic discrete world model as policy.""" hparams = ppo_original_params() hparams.policy_network = "next_frame_basic_stochastic_discrete" hparams_keys = hparams.values().keys() video_hparams = basic_stochastic.next_frame_basic_stochastic_discrete() for (name, value) in six.iteritems(video_hparams.values()): if name in hparams_keys: hparams.set_hparam(name, value) else: hparams.add_hparam(name, value) # To avoid OOM. Probably way to small. hparams.optimization_batch_size = 1 hparams.weight_decay = 0 return hparams
Returns a function creating a simulated env, in or out of graph. Args: **env_kwargs: kwargs to pass to the simulated env constructor. Returns: Function in_graph -> env. def make_simulated_env_fn(**env_kwargs): """Returns a function creating a simulated env, in or out of graph. Args: **env_kwargs: kwargs to pass to the simulated env constructor. Returns: Function in_graph -> env. """ def env_fn(in_graph): class_ = SimulatedBatchEnv if in_graph else SimulatedBatchGymEnv return class_(**env_kwargs) return env_fn
Extracts simulated env kwargs from real_env and loop hparams. def make_simulated_env_kwargs(real_env, hparams, **extra_kwargs): """Extracts simulated env kwargs from real_env and loop hparams.""" objs_and_attrs = [ (real_env, [ "reward_range", "observation_space", "action_space", "frame_height", "frame_width" ]), (hparams, ["frame_stack_size", "intrinsic_reward_scale"]) ] kwargs = { attr: getattr(obj, attr) # pylint: disable=g-complex-comprehension for (obj, attrs) in objs_and_attrs for attr in attrs } kwargs["model_name"] = hparams.generative_model kwargs["model_hparams"] = trainer_lib.create_hparams( hparams.generative_model_params ) if hparams.wm_policy_param_sharing: kwargs["model_hparams"].optimizer_zero_grads = True kwargs.update(extra_kwargs) return kwargs
Get a policy network. Args: observations: observations hparams: parameters action_space: action space Returns: Tuple (action logits, value). def get_policy(observations, hparams, action_space): """Get a policy network. Args: observations: observations hparams: parameters action_space: action space Returns: Tuple (action logits, value). """ if not isinstance(action_space, gym.spaces.Discrete): raise ValueError("Expecting discrete action space.") obs_shape = common_layers.shape_list(observations) (frame_height, frame_width) = obs_shape[2:4] # TODO(afrozm): We have these dummy problems mainly for hparams, so cleanup # when possible and do this properly. if hparams.policy_problem_name == "dummy_policy_problem_ttt": tf.logging.info("Using DummyPolicyProblemTTT for the policy.") policy_problem = tic_tac_toe_env.DummyPolicyProblemTTT() else: tf.logging.info("Using DummyPolicyProblem for the policy.") policy_problem = DummyPolicyProblem(action_space, frame_height, frame_width) trainer_lib.add_problem_hparams(hparams, policy_problem) hparams.force_full_predict = True model = registry.model(hparams.policy_network)( hparams, tf.estimator.ModeKeys.TRAIN ) try: num_target_frames = hparams.video_num_target_frames except AttributeError: num_target_frames = 1 features = { "inputs": observations, "input_action": tf.zeros(obs_shape[:2] + [1], dtype=tf.int32), "input_reward": tf.zeros(obs_shape[:2] + [1], dtype=tf.int32), "targets": tf.zeros(obs_shape[:1] + [num_target_frames] + obs_shape[2:]), "target_action": tf.zeros( obs_shape[:1] + [num_target_frames, 1], dtype=tf.int32), "target_reward": tf.zeros( obs_shape[:1] + [num_target_frames, 1], dtype=tf.int32), "target_policy": tf.zeros( obs_shape[:1] + [num_target_frames] + [action_space.n]), "target_value": tf.zeros( obs_shape[:1] + [num_target_frames]) } with tf.variable_scope(tf.get_variable_scope(), reuse=tf.AUTO_REUSE): t2t_model.create_dummy_vars() (targets, _) = model(features) return (targets["target_policy"][:, 0, :], targets["target_value"][:, 0])
Base set of hparams for model-free PPO. def rlmf_tictactoe(): """Base set of hparams for model-free PPO.""" hparams = rlmf_original() hparams.game = "tictactoe" hparams.rl_env_name = "T2TEnv-TicTacToeEnv-v0" # Since we don't have any no-op actions, otherwise we have to have an # attribute called `get_action_meanings`. hparams.eval_max_num_noops = 0 hparams.max_num_noops = 0 hparams.rl_should_derive_observation_space = False hparams.policy_network = "feed_forward_categorical_policy" hparams.base_algo_params = "ppo_ttt_params" # Number of last observations to feed to the agent hparams.frame_stack_size = 1 return hparams
Tiny set of hparams for model-free PPO. def rlmf_tiny(): """Tiny set of hparams for model-free PPO.""" hparams = rlmf_original() hparams = hparams.override_from_dict(rlmf_tiny_overrides()) hparams.batch_size = 2 hparams.base_algo_params = "ppo_original_tiny" hparams.add_hparam("ppo_epochs_num", 3) hparams.add_hparam("ppo_epoch_length", 2) return hparams
Tiny DQN params. def rlmf_dqn_tiny(): """Tiny DQN params.""" hparams = rlmf_original() hparams = hparams.override_from_dict(rlmf_tiny_overrides()) hparams.batch_size = 1 hparams.base_algo = "dqn" hparams.base_algo_params = "dqn_original_params" hparams.add_hparam("dqn_num_frames", 128) hparams.add_hparam("dqn_save_every_steps", 128) hparams.add_hparam("dqn_replay_buffer_replay_capacity", 100) hparams.add_hparam("dqn_agent_min_replay_history", 10) return hparams
Eval set of hparams for model-free PPO. def rlmf_eval(): """Eval set of hparams for model-free PPO.""" hparams = rlmf_original() hparams.batch_size = 8 hparams.eval_sampling_temps = [0.0, 0.5, 1.0] hparams.eval_rl_env_max_episode_steps = -1 hparams.add_hparam("ppo_epoch_length", 128) hparams.add_hparam("ppo_optimization_batch_size", 32) hparams.add_hparam("ppo_epochs_num", 10000) hparams.add_hparam("ppo_eval_every_epochs", 500) hparams.add_hparam("attempt", 0) hparams.add_hparam("moe_loss_coef", 0) return hparams
Feed-forward Gaussian. def feed_forward_gaussian_fun(action_space, config, observations): """Feed-forward Gaussian.""" if not isinstance(action_space, gym.spaces.box.Box): raise ValueError("Expecting continuous action space.") mean_weights_initializer = tf.initializers.variance_scaling( scale=config.init_mean_factor) logstd_initializer = tf.random_normal_initializer(config.init_logstd, 1e-10) flat_observations = tf.reshape(observations, [ tf.shape(observations)[0], tf.shape(observations)[1], functools.reduce(operator.mul, observations.shape.as_list()[2:], 1)]) with tf.variable_scope("network_parameters"): with tf.variable_scope("policy"): x = flat_observations for size in config.policy_layers: x = tf.layers.dense(x, size, activation=tf.nn.relu) mean = tf.layers.dense( x, action_space.shape[0], activation=tf.tanh, kernel_initializer=mean_weights_initializer) logstd = tf.get_variable( "logstd", mean.shape[2:], tf.float32, logstd_initializer) logstd = tf.tile( logstd[None, None], [tf.shape(mean)[0], tf.shape(mean)[1]] + [1] * (mean.shape.ndims - 2)) with tf.variable_scope("value"): x = flat_observations for size in config.value_layers: x = tf.layers.dense(x, size, activation=tf.nn.relu) value = tf.layers.dense(x, 1)[..., 0] mean = tf.check_numerics(mean, "mean") logstd = tf.check_numerics(logstd, "logstd") value = tf.check_numerics(value, "value") policy = tfp.distributions.MultivariateNormalDiag(mean, tf.exp(logstd)) return NetworkOutput(policy, value, lambda a: tf.clip_by_value(a, -2., 2))
Curvature range. Returns: h_max_t, h_min_t ops def _curvature_range(self): """Curvature range. Returns: h_max_t, h_min_t ops """ self._curv_win = tf.get_variable("curv_win", dtype=tf.float32, trainable=False, shape=[self.curvature_window_width,], initializer=tf.zeros_initializer) # We use log smoothing for curvature range self._curv_win = tf.scatter_update(self._curv_win, self._step % self.curvature_window_width, tf.log(self._grad_norm_squared)) # Note here the iterations start from iteration 0 valid_window = tf.slice(self._curv_win, tf.constant([0,]), tf.expand_dims( tf.minimum( tf.constant(self.curvature_window_width), self._step + 1), dim=0)) self._h_min_t = tf.reduce_min(valid_window) self._h_max_t = tf.reduce_max(valid_window) curv_range_ops = [] with tf.control_dependencies([self._h_min_t, self._h_max_t]): avg_op = self._moving_averager.apply([self._h_min_t, self._h_max_t]) with tf.control_dependencies([avg_op]): self._h_min = tf.exp( tf.identity(self._moving_averager.average(self._h_min_t))) self._h_max = tf.exp( tf.identity(self._moving_averager.average(self._h_max_t))) if self._sparsity_debias: self._h_min *= self._sparsity_avg self._h_max *= self._sparsity_avg curv_range_ops.append(avg_op) return curv_range_ops
Estimate of gradient Variance. Returns: C_t ops. def _grad_variance(self): """Estimate of gradient Variance. Returns: C_t ops. """ grad_var_ops = [] tensor_to_avg = [] for t, g in zip(self._vars, self._grad): if isinstance(g, tf.IndexedSlices): tensor_to_avg.append( tf.reshape(tf.unsorted_segment_sum(g.values, g.indices, g.dense_shape[0]), shape=t.get_shape())) else: tensor_to_avg.append(g) avg_op = self._moving_averager.apply(tensor_to_avg) grad_var_ops.append(avg_op) with tf.control_dependencies([avg_op]): self._grad_avg = [self._moving_averager.average(val) for val in tensor_to_avg] self._grad_avg_squared = [tf.square(val) for val in self._grad_avg] # Compute Variance self._grad_var = tf.maximum( tf.constant(1e-6, dtype=self._grad_norm_squared_avg.dtype), self._grad_norm_squared_avg - tf.add_n([tf.reduce_sum(val) for val in self._grad_avg_squared])) if self._sparsity_debias: self._grad_var *= self._sparsity_avg return grad_var_ops
Distance to optimum. Returns: D_t ops def _dist_to_opt(self): """Distance to optimum. Returns: D_t ops """ dist_to_opt_ops = [] # Running average of the norm of gradient self._grad_norm = tf.sqrt(self._grad_norm_squared) avg_op = self._moving_averager.apply([self._grad_norm,]) dist_to_opt_ops.append(avg_op) with tf.control_dependencies([avg_op]): self._grad_norm_avg = self._moving_averager.average(self._grad_norm) # Single iteration distance estimation, note here # self._grad_norm_avg is per variable self._d_t = self._grad_norm_avg / self._grad_norm_squared_avg # Running average of distance avg_op = self._moving_averager.apply([self._d_t]) dist_to_opt_ops.append(avg_op) with tf.control_dependencies([avg_op]): self._dist_to_opt_avg = tf.identity( self._moving_averager.average(self._d_t)) if self._sparsity_debias: self._dist_to_opt_avg /= tf.sqrt(self._sparsity_avg) return dist_to_opt_ops
Gradient sparsity. def _grad_sparsity(self): """Gradient sparsity.""" # If the sparse minibatch gradient has 10 percent of its entries # non-zero, its sparsity is 0.1. # The norm of dense gradient averaged from full dataset # are roughly estimated norm of minibatch # sparse gradient norm * sqrt(sparsity) # An extension maybe only correct the sparse blob. non_zero_cnt = tf.add_n([tf.count_nonzero(g) for g in self._grad]) all_entry_cnt = tf.add_n([tf.size(g) for g in self._grad]) self._sparsity = tf.cast(non_zero_cnt, self._grad[0].dtype) self._sparsity /= tf.cast(all_entry_cnt, self._grad[0].dtype) avg_op = self._moving_averager.apply([self._sparsity,]) with tf.control_dependencies([avg_op]): self._sparsity_avg = self._moving_averager.average(self._sparsity) return avg_op
Prepare Variables for YellowFin. Returns: Grad**2, Norm, Norm**2, Mean(Norm**2) ops def _prepare_variables(self): """Prepare Variables for YellowFin. Returns: Grad**2, Norm, Norm**2, Mean(Norm**2) ops """ self._moving_averager = tf.train.ExponentialMovingAverage( decay=self._beta, zero_debias=self._zero_debias) # assert self._grad is not None and len(self._grad) > 0 # List for the returned Operations prepare_variables_op = [] # Get per var g**2 and norm**2 self._grad_squared = [] self._grad_norm_squared = [] # Gradient squared for v, g in zip(self._vars, self._grad): if g is None: continue with tf.colocate_with(v): self._grad_squared.append(tf.square(g)) # Norm squared. self._grad_norm_squared = [tf.reduce_sum(g_sq) for g_sq in self._grad_squared] if self._sparsity_debias: avg_op_sparsity = self._grad_sparsity() prepare_variables_op.append(avg_op_sparsity) # The following running average on squared norm of gradient # is shared by grad_var and dist_to_opt avg_op = self._moving_averager.apply(self._grad_norm_squared) with tf.control_dependencies([avg_op]): self._grad_norm_squared_avg = [self._moving_averager.average(val) for val in self._grad_norm_squared] self._grad_norm_squared = tf.add_n(self._grad_norm_squared) self._grad_norm_squared_avg = tf.add_n(self._grad_norm_squared_avg) prepare_variables_op.append(avg_op) return tf.group(*prepare_variables_op)
Get the cubic root. def _get_cubic_root(self): """Get the cubic root.""" # We have the equation x^2 D^2 + (1-x)^4 * C / h_min^2 # where x = sqrt(mu). # We substitute x, which is sqrt(mu), with x = y + 1. # It gives y^3 + py = q # where p = (D^2 h_min^2)/(2*C) and q = -p. # We use the Vieta's substitution to compute the root. # There is only one real solution y (which is in [0, 1] ). # http://mathworld.wolfram.com/VietasSubstitution.html assert_array = [ tf.Assert( tf.logical_not(tf.is_nan(self._dist_to_opt_avg)), [self._dist_to_opt_avg,]), tf.Assert( tf.logical_not(tf.is_nan(self._h_min)), [self._h_min,]), tf.Assert( tf.logical_not(tf.is_nan(self._grad_var)), [self._grad_var,]), tf.Assert( tf.logical_not(tf.is_inf(self._dist_to_opt_avg)), [self._dist_to_opt_avg,]), tf.Assert( tf.logical_not(tf.is_inf(self._h_min)), [self._h_min,]), tf.Assert( tf.logical_not(tf.is_inf(self._grad_var)), [self._grad_var,]) ] with tf.control_dependencies(assert_array): p = self._dist_to_opt_avg**2 * self._h_min**2 / 2 / self._grad_var w3 = (-tf.sqrt(p**2 + 4.0 / 27.0 * p**3) - p) / 2.0 w = tf.sign(w3) * tf.pow(tf.abs(w3), 1.0/3.0) y = w - p / 3.0 / w x = y + 1 return x
Get lr minimizing the surrogate. Returns: The lr_t. def _get_lr_tensor(self): """Get lr minimizing the surrogate. Returns: The lr_t. """ lr = tf.squared_difference(1.0, tf.sqrt(self._mu)) / self._h_min return lr
Get the min mu which minimize the surrogate. Returns: The mu_t. def _get_mu_tensor(self): """Get the min mu which minimize the surrogate. Returns: The mu_t. """ root = self._get_cubic_root() dr = self._h_max / self._h_min mu = tf.maximum( root**2, ((tf.sqrt(dr) - 1) / (tf.sqrt(dr) + 1))**2) return mu
YellowFin auto-tuning optimizer based on momentum SGD. Returns: YF ops (Curvature range, Grad_variance, Dist_to_opt, Single-Step, Auto-Tuning) def _yellowfin(self): """YellowFin auto-tuning optimizer based on momentum SGD. Returns: YF ops (Curvature range, Grad_variance, Dist_to_opt, Single-Step, Auto-Tuning) """ # List for the returned Operations. yellowfin_ops = [] # Curvature range ops. curv_range_ops = self._curvature_range() yellowfin_ops += curv_range_ops # Estimate of gradient Variance ops. grad_var_ops = self._grad_variance() yellowfin_ops += grad_var_ops # Distance to optimum ops. dist_to_opt_ops = self._dist_to_opt() yellowfin_ops += dist_to_opt_ops # Single-Step: minimizes the surrogate for the expected # squared distance from the optimum of a local quadratic # approximation after a single step while keeping all directions in the # robust region. self._mu = tf.identity(tf.cond(self._do_tune, self._get_mu_tensor, lambda: self._mu_var)) with tf.control_dependencies([self._mu]): self._lr = tf.identity(tf.cond(self._do_tune, self._get_lr_tensor, lambda: self._lr_var)) # Tune learning rate and momentum. with tf.control_dependencies([self._mu, self._lr]): self._mu = self._beta * self._mu_var + (1 - self._beta) * self._mu self._lr = self._beta * self._lr_var + (1 - self._beta) * self._lr yellowfin_ops.append(tf.assign(self._mu_var, self._mu)) yellowfin_ops.append(tf.assign(self._lr_var, self._lr)) yellowfin_ops = tf.group(*yellowfin_ops) return yellowfin_ops
Applying gradients and tune hyperparams with YellowFin. Args: grads_and_vars: List of (gradient, variable) pairs as returned by compute_gradients(). global_step: Optional Variable to increment by one after the variables have been updated. name: Optional name for the returned operation. Default to the name passed to the Optimizer constructor. Returns: (A group of operations) Variable Update with Momentum ops, YellowFin ops(Curvature, Variance, Distance) ops, SingleStep and lr_mu tuning ops, Step increment ops. def apply_gradients(self, grads_and_vars, global_step=None, name=None): """Applying gradients and tune hyperparams with YellowFin. Args: grads_and_vars: List of (gradient, variable) pairs as returned by compute_gradients(). global_step: Optional Variable to increment by one after the variables have been updated. name: Optional name for the returned operation. Default to the name passed to the Optimizer constructor. Returns: (A group of operations) Variable Update with Momentum ops, YellowFin ops(Curvature, Variance, Distance) ops, SingleStep and lr_mu tuning ops, Step increment ops. """ self._grad, self._vars = zip(*[(g, t) for g, t in grads_and_vars if g is not None]) # Var update with Momentum. with tf.variable_scope("apply_updates"): # Gradient Clipping? if self._clip_thresh_var is not None: self._grad, _ = tf.clip_by_global_norm( self._grad, self._clip_thresh_var) apply_grad_op = self._momentum_optimizer.apply_gradients( zip(self._grad, self._vars), global_step=global_step, name=name) else: apply_grad_op = self._momentum_optimizer.apply_gradients( zip(self._grad, self._vars), global_step=global_step, name=name) # Begin lr and mu tuning. with tf.variable_scope("prepare_yellowFin_variables"): # the dependencies ideally only need to be after clip is done, # i.e. depends on self._grads. However, the control_dependencies # does not support indexed slice for sparse gradients. # The alternative dependencies here might be slightly slower due # to less parallelization. with tf.control_dependencies([apply_grad_op,]): prepare_variables_op = self._prepare_variables() with tf.variable_scope("yellowfin"): with tf.control_dependencies([prepare_variables_op]): yellowfin_op = self._yellowfin() # Update YellowFin step variable. with tf.control_dependencies([yellowfin_op]): self._increment_step_op = tf.assign_add(self._step, 1).op return tf.group(apply_grad_op, prepare_variables_op, yellowfin_op, self._increment_step_op)
Compute gradients through momentum optimizer. Args: loss: A Tensor containing the value to minimize. var_list: Optional list or tuple of tf.Variable to update to minimize loss. Defaults to the list of variables collected in the graph under the key GraphKey.TRAINABLE_VARIABLES. global_step: Optional Variable to increment by one after the variables have been updated. gate_gradients: How to gate the computation of gradients. Can be GATE_NONE, GATE_OP, or GATE_GRAPH. aggregation_method: Specifies the method used to combine gradient terms. Valid values are defined in the class AggregationMethod. colocate_gradients_with_ops: If True, try collocating gradients with the corresponding op. name: Optional name for the returned operation. Default to the name passed to the Optimizer constructor. grad_loss: Optional. A Tensor holding the gradient computed for loss. Returns: A list of (gradient, variable) pairs. Variable is always present, but gradient can be None. def compute_gradients(self, loss, var_list, global_step=None, gate_gradients=GATE_OP, aggregation_method=None, colocate_gradients_with_ops=False, name=None, grad_loss=None): """Compute gradients through momentum optimizer. Args: loss: A Tensor containing the value to minimize. var_list: Optional list or tuple of tf.Variable to update to minimize loss. Defaults to the list of variables collected in the graph under the key GraphKey.TRAINABLE_VARIABLES. global_step: Optional Variable to increment by one after the variables have been updated. gate_gradients: How to gate the computation of gradients. Can be GATE_NONE, GATE_OP, or GATE_GRAPH. aggregation_method: Specifies the method used to combine gradient terms. Valid values are defined in the class AggregationMethod. colocate_gradients_with_ops: If True, try collocating gradients with the corresponding op. name: Optional name for the returned operation. Default to the name passed to the Optimizer constructor. grad_loss: Optional. A Tensor holding the gradient computed for loss. Returns: A list of (gradient, variable) pairs. Variable is always present, but gradient can be None. """ del global_step, name # Unused for now. return self._momentum_optimizer.compute_gradients( loss, var_list=var_list, gate_gradients=gate_gradients, aggregation_method=aggregation_method, colocate_gradients_with_ops=colocate_gradients_with_ops, grad_loss=grad_loss)
Adapted from TensorFlow Optimizer base class member function. Add operations to minimize `loss` by updating `var_list`. This method simply combines calls `compute_gradients()` and `apply_gradients()`. If you want to process the gradient before applying them call `tf.gradients()` and `self.apply_gradients()` explicitly instead of using this function. Args: loss: A Tensor containing the value to minimize. global_step: Optional Variable to increment by one after the variables have been updated. var_list: Optional list or tuple of Variable objects to update to minimize loss. Defaults to the list of variables collected in the graph under the key GraphKeys.TRAINABLE_VARIABLES. gate_gradients: How to gate the computation of gradients. Can be GATE_NONE, GATE_OP, or GATE_GRAPH. aggregation_method: Specifies the method used to combine gradient terms. Valid values are defined in the class AggregationMethod. colocate_gradients_with_ops: If True, try collocating gradients with the corresponding op. name: Optional name for the returned operation. grad_loss: Optional. A Tensor holding the gradient computed for loss. Returns: An Operation that updates the variables in var_list. If global_step was not None, that operation also increments global_step. Raises: ValueError: if no gradients are provided for any variable. def minimize(self, loss, global_step=None, var_list=None, gate_gradients=GATE_OP, aggregation_method=None, colocate_gradients_with_ops=False, name=None, grad_loss=None): """Adapted from TensorFlow Optimizer base class member function. Add operations to minimize `loss` by updating `var_list`. This method simply combines calls `compute_gradients()` and `apply_gradients()`. If you want to process the gradient before applying them call `tf.gradients()` and `self.apply_gradients()` explicitly instead of using this function. Args: loss: A Tensor containing the value to minimize. global_step: Optional Variable to increment by one after the variables have been updated. var_list: Optional list or tuple of Variable objects to update to minimize loss. Defaults to the list of variables collected in the graph under the key GraphKeys.TRAINABLE_VARIABLES. gate_gradients: How to gate the computation of gradients. Can be GATE_NONE, GATE_OP, or GATE_GRAPH. aggregation_method: Specifies the method used to combine gradient terms. Valid values are defined in the class AggregationMethod. colocate_gradients_with_ops: If True, try collocating gradients with the corresponding op. name: Optional name for the returned operation. grad_loss: Optional. A Tensor holding the gradient computed for loss. Returns: An Operation that updates the variables in var_list. If global_step was not None, that operation also increments global_step. Raises: ValueError: if no gradients are provided for any variable. """ grads_and_vars = self._momentum_optimizer.compute_gradients( loss, var_list=var_list, gate_gradients=gate_gradients, aggregation_method=aggregation_method, colocate_gradients_with_ops=colocate_gradients_with_ops, grad_loss=grad_loss) vars_with_grad = [v for g, v in grads_and_vars if g is not None] if not vars_with_grad: raise ValueError( "No gradients provided for any variable, check your graph for ops" " that do not support gradients, between variables %s and loss %s." % ([str(v) for _, v in grads_and_vars], loss)) for g, v in grads_and_vars: print("g ", g) print("v ", v) return self.apply_gradients(grads_and_vars, global_step=global_step, name=name)
A stack of convolution blocks with residual connections. def residual_dilated_conv(x, repeat, padding, name, hparams): """A stack of convolution blocks with residual connections.""" with tf.variable_scope(name): k = (hparams.kernel_height, hparams.kernel_width) dilations_and_kernels = [((2**i, 1), k) for i in range(hparams.num_hidden_layers)] for i in range(repeat): with tf.variable_scope("repeat_%d" % i): y = common_layers.conv_block( common_layers.layer_norm(x, hparams.hidden_size, name="lnorm"), hparams.hidden_size, dilations_and_kernels, padding=padding, name="residual_conv") y = tf.nn.dropout(y, 1.0 - hparams.dropout) x += y return x
ByteNet, main step used for training. def bytenet_internal(inputs, targets, hparams): """ByteNet, main step used for training.""" with tf.variable_scope("bytenet"): # Flatten inputs and extend length by 50%. inputs = tf.expand_dims(common_layers.flatten4d3d(inputs), axis=2) extend_length = tf.to_int32(0.5 * tf.to_float(tf.shape(inputs)[1])) inputs_shape = inputs.shape.as_list() inputs = tf.pad(inputs, [[0, 0], [0, extend_length], [0, 0], [0, 0]]) inputs_shape[1] = None inputs.set_shape(inputs_shape) # Don't lose the other shapes when padding. # Pad inputs and targets to be the same length, divisible by 50. inputs, targets = common_layers.pad_to_same_length( inputs, targets, final_length_divisible_by=50) final_encoder = residual_dilated_conv(inputs, hparams.num_block_repeat, "SAME", "encoder", hparams) shifted_targets = common_layers.shift_right(targets) kernel = (hparams.kernel_height, hparams.kernel_width) decoder_start = common_layers.conv_block( tf.concat([final_encoder, shifted_targets], axis=3), hparams.hidden_size, [((1, 1), kernel)], padding="LEFT") return residual_dilated_conv(decoder_start, hparams.num_block_repeat, "LEFT", "decoder", hparams)
Set of hyperparameters. def bytenet_base(): """Set of hyperparameters.""" hparams = common_hparams.basic_params1() hparams.batch_size = 2048 hparams.hidden_size = 768 hparams.dropout = 0.2 hparams.symbol_dropout = 0.2 hparams.label_smoothing = 0.1 hparams.clip_grad_norm = 2.0 hparams.num_hidden_layers = 4 hparams.kernel_height = 3 hparams.kernel_width = 1 hparams.learning_rate_decay_scheme = "exp" hparams.learning_rate = 0.05 hparams.learning_rate_warmup_steps = 3000 hparams.initializer_gain = 1.0 hparams.weight_decay = 3.0 hparams.num_sampled_classes = 0 hparams.sampling_method = "argmax" hparams.optimizer_adam_epsilon = 1e-6 hparams.optimizer_adam_beta1 = 0.85 hparams.optimizer_adam_beta2 = 0.997 hparams.add_hparam("num_block_repeat", 4) return hparams
Downloads and prepairs the dataset to be parsed by the data_generator. def _download_and_parse_dataset(tmp_dir, train): """Downloads and prepairs the dataset to be parsed by the data_generator.""" file_path = generator_utils.maybe_download(tmp_dir, _SNLI_ZIP, _SNLI_URL) zip_ref = zipfile.ZipFile(file_path, 'r') zip_ref.extractall(tmp_dir) zip_ref.close() file_name = 'train' if train else 'dev' dataset_file_path = os.path.join(tmp_dir, _SNLI_DATA_PATH % file_name) _parse_dataset(dataset_file_path, tmp_dir, train)
Parse str to tokens and pos tags. def _get_tokens_and_tags(parse_str): """Parse str to tokens and pos tags.""" tokens = [] parse_split = parse_str.split(' ') for p in parse_split: assert p.startswith('(') or p.endswith(')') if p.endswith(')'): token = p.replace(')', '') tokens.append(token) return tokens
Convert the dataset in to a simpler format. This function creates two files. One for being processed to produce a vocab and another to generate the data. Args: file_path: string, path to the file to parse. tmp_dir: string, path to the directory to output the files. train: bool, indicating if we are parsing the training set. def _parse_dataset(file_path, tmp_dir, train): """Convert the dataset in to a simpler format. This function creates two files. One for being processed to produce a vocab and another to generate the data. Args: file_path: string, path to the file to parse. tmp_dir: string, path to the directory to output the files. train: bool, indicating if we are parsing the training set. """ input_path = file_path file_name = 'train' if train else 'dev' gen_output_path = os.path.join(tmp_dir, file_name + '.txt') example_output_path = os.path.join(tmp_dir, _EXAMPLES_FILE) print('input path: ' + input_path) print('gen_output_path: ' + gen_output_path) print('example_output_path: ' + example_output_path) input_file = tf.gfile.Open(input_path, mode='r') examples = [] for counter, line in enumerate(input_file): if counter == 0: # Ignore first line since its a header. continue # Get the token and embedding vector. line_split = line.split('\t') parse1 = line_split[_PARSE1_INDEX] parse2 = line_split[_PARSE2_INDEX] consensus_label = line_split[_LABEL_INDEX] tokens1 = _get_tokens_and_tags(parse1) tokens2 = _get_tokens_and_tags(parse2) tokens1_str = ' '.join(tokens1) tokens2_str = ' '.join(tokens2) if consensus_label != '-': examples.append([tokens1_str, tokens2_str, consensus_label]) input_file.close() # Output tab delimited file of lines of examples (sentence1, sentence2, label) with tf.gfile.GFile(gen_output_path, 'w') as f: for tokens1_str, tokens2_str, consensus_label in examples: f.write('%s\t%s\t%s\n' % (tokens1_str, tokens2_str, consensus_label)) if train: # Output file containing all the sentences for generating the vocab from. with tf.gfile.GFile(example_output_path, 'w') as f: for tokens1_str, tokens2_str, consensus_label in examples: f.write('%s %s\n' % (tokens1_str, tokens2_str))
Read or create vocabulary. def _get_or_generate_vocab(tmp_dir, vocab_filename, vocab_size): """Read or create vocabulary.""" vocab_filepath = os.path.join(tmp_dir, vocab_filename) print('Vocab file written to: ' + vocab_filepath) if tf.gfile.Exists(vocab_filepath): gs = text_encoder.SubwordTextEncoder(vocab_filepath) return gs example_file = os.path.join(tmp_dir, _EXAMPLES_FILE) gs = text_encoder.SubwordTextEncoder() token_counts = tokenizer.corpus_token_counts( example_file, corpus_max_lines=1000000) gs = gs.build_to_target_size( vocab_size, token_counts, min_val=1, max_val=1e3) gs.store_to_file(vocab_filepath) return gs
Generate example dicts. def snli_token_generator(tmp_dir, train, vocab_size): """Generate example dicts.""" _download_and_parse_dataset(tmp_dir, train) symbolizer_vocab = _get_or_generate_vocab( tmp_dir, 'vocab.subword_text_encoder', vocab_size) file_name = 'train' if train else 'dev' data_file = os.path.join(tmp_dir, file_name + '.txt') with tf.gfile.GFile(data_file, mode='r') as f: for line in f: sent1, sent2, label = line.strip().split('\t') sent1_enc = symbolizer_vocab.encode(sent1) sent2_enc = symbolizer_vocab.encode(sent2) inputs = sent1_enc + [_SEP] + sent2_enc + [_EOS] yield { 'inputs': inputs, 'targets': [_LABEL_TO_ID[label]], }
Split items into num_shards groups. def shard(items, num_shards): """Split items into num_shards groups.""" sharded = [] num_per_shard = len(items) // num_shards start = 0 for _ in range(num_shards): sharded.append(items[start:start + num_per_shard]) start += num_per_shard remainder = len(items) % num_shards start = len(items) - remainder for i in range(remainder): sharded[i].append(items[start + i]) assert sum([len(fs) for fs in sharded]) == len(items) return sharded
An initializer function for random normal coefficients. def RandomNormalInitializer(stddev=1e-2): """An initializer function for random normal coefficients.""" def init(shape, rng): return (stddev * backend.random.normal(rng, shape)).astype('float32') return init
An initializer function for random Glorot-scaled coefficients. def GlorotNormalInitializer(out_dim=0, in_dim=1, scale=onp.sqrt(2)): """An initializer function for random Glorot-scaled coefficients.""" def init(shape, rng): fan_in, fan_out = shape[in_dim], shape[out_dim] size = onp.prod(onp.delete(shape, [in_dim, out_dim])) std = scale / np.sqrt((fan_in + fan_out) / 2. * size) return (std * backend.random.normal(rng, shape)).astype('float32') return init
An initializer function for random uniform Glorot-scaled coefficients. def GlorotUniformInitializer(out_dim=0, in_dim=1): """An initializer function for random uniform Glorot-scaled coefficients.""" def init(shape, rng): fan_in, fan_out = shape[in_dim], shape[out_dim] std = np.sqrt(2.0 / (fan_in + fan_out)) a = np.sqrt(3.0) * std return backend.random.uniform(rng, shape, minval=-a, maxval=a) return init
Make a n+1 dim one-hot array from n dim int-categorical array. def one_hot(x, size, dtype=np.float32): """Make a n+1 dim one-hot array from n dim int-categorical array.""" return np.array(x[..., np.newaxis] == np.arange(size), dtype)
Apply log softmax to x: log-normalize along the given axis. def LogSoftmax(x, params, axis=-1, **kwargs): """Apply log softmax to x: log-normalize along the given axis.""" del params, kwargs return x - backend.logsumexp(x, axis, keepdims=True)
Apply softmax to x: exponentiate and normalize along the given axis. def Softmax(x, params, axis=-1, **kwargs): """Apply softmax to x: exponentiate and normalize along the given axis.""" del params, kwargs return np.exp(x - backend.logsumexp(x, axis, keepdims=True))
Convert padding string to list of pairs of pad values. def padtype_to_pads(in_shape, window_shape, window_strides, padding): """Convert padding string to list of pairs of pad values.""" padding = padding.upper() if padding == 'SAME': out_shape = onp.ceil( onp.true_divide(in_shape, window_strides)).astype(int) pad_sizes = [max((out_size - 1) * stride + window_shape - in_size, 0) for out_size, stride, window_shape, in_size in zip(out_shape, window_strides, window_shape, in_shape)] return [(pad_size // 2, pad_size - pad_size // 2) for pad_size in pad_sizes] elif padding == 'VALID': return [(0, 0)] * len(in_shape) else: msg = 'Unknown padding type: {}.' raise TypeError(msg.format(padding))
Output shape of a flatten layer. def _flatten_output_shape(input_shape, num_axis_to_keep=1): """Output shape of a flatten layer.""" if num_axis_to_keep >= len(input_shape): raise ValueError( "num_axis_to_keep[%d] should be less than input's rank[%d]" % (num_axis_to_keep, len(input_shape))) return tuple(input_shape[:num_axis_to_keep]) + ( reduce(op.mul, input_shape[num_axis_to_keep:], 1),)
Helper to initialize batch norm params. def _batch_norm_new_params(input_shape, rng, axis=(0, 1, 2), center=True, scale=True, **kwargs): """Helper to initialize batch norm params.""" del rng, kwargs axis = (axis,) if np.isscalar(axis) else axis shape = tuple(d for i, d in enumerate(input_shape) if i not in axis) beta = np.zeros(shape, dtype='float32') if center else () gamma = np.ones(shape, dtype='float32') if scale else () return (beta, gamma)
Layer construction function for a batch normalization layer. def BatchNorm(x, params, axis=(0, 1, 2), epsilon=1e-5, center=True, scale=True, **unused_kwargs): """Layer construction function for a batch normalization layer.""" mean = np.mean(x, axis, keepdims=True) # Fast but less numerically-stable variance calculation than np.var. m1 = np.mean(x**2, axis, keepdims=True) var = m1 - mean**2 z = (x - mean) / np.sqrt(var + epsilon) # Expand the parameters to have the right axes. beta, gamma = params # TODO(phawkins): np.expand_dims should accept an axis tuple. # (https://github.com/numpy/numpy/issues/12290) ed = tuple(None if i in axis else slice(None) for i in range(np.ndim(x))) beta = beta[ed] gamma = gamma[ed] # Return the z rescaled by the parameters if requested. if center and scale: return gamma * z + beta if center: return z + beta if scale: return gamma * z return z
Helper: compute the output shape for the pooling layer. def _pooling_output_shape(input_shape, pool_size=(2, 2), strides=None, padding='VALID'): """Helper: compute the output shape for the pooling layer.""" dims = (1,) + pool_size + (1,) # NHWC spatial_strides = strides or (1,) * len(pool_size) strides = (1,) + spatial_strides + (1,) pads = padtype_to_pads(input_shape, dims, strides, padding) operand_padded = onp.add(input_shape, onp.add(*zip(*pads))) t = onp.floor_divide(onp.subtract(operand_padded, dims), strides) + 1 return tuple(t)
Helper: general pooling computation used in pooling layers later. def _pooling_general(inputs, reducer, init_val, rescaler=None, pool_size=(2, 2), strides=None, padding='VALID'): """Helper: general pooling computation used in pooling layers later.""" spatial_strides = strides or (1,) * len(pool_size) rescale = rescaler(pool_size, spatial_strides, padding) if rescaler else None dims = (1,) + pool_size + (1,) # NHWC strides = (1,) + spatial_strides + (1,) out = lax.reduce_window(inputs, init_val, reducer, dims, strides, padding) return rescale(out, inputs) if rescale else out
Layer construction function for a dropout layer with given rate. def Dropout(x, params, rate=0.0, mode='train', rng=None, **kwargs): """Layer construction function for a dropout layer with given rate.""" del params, kwargs if rng is None: msg = ('Dropout layer requires apply_fun to be called with a rng keyword ' 'argument. That is, instead of `Dropout(params, inputs)`, call ' 'it like `Dropout(params, inputs, rng=key)`.') raise ValueError(msg) if rate >= 1.0: raise ValueError('Dropout rate (%f) must be lower than 1.' % rate) if mode == 'train' and rate > 0.0: keep = backend.random.bernoulli(rng, 1.0 - rate, x.shape) return np.where(keep, x / (1.0 - rate), 0) else: return x
Helper to calculate the kernel shape. def _kernel_shape(self, input_shape): """Helper to calculate the kernel shape.""" kernel_size_iter = iter(self._kernel_size) return [self._filters if c == 'O' else input_shape[self._lhs_spec.index('C')] if c == 'I' else next(kernel_size_iter) for c in self._rhs_spec]
Compute the shape of a conv given input shapes in canonical order. def _conv_shape_tuple(self, lhs_shape, rhs_shape, strides, pads): """Compute the shape of a conv given input shapes in canonical order.""" if isinstance(pads, str): pads = padtype_to_pads(lhs_shape[2:], rhs_shape[2:], strides, pads) if len(pads) != len(lhs_shape) - 2: msg = 'Wrong number of explicit pads for conv: expected {}, got {}.' raise TypeError(msg.format(len(lhs_shape) - 2, len(pads))) lhs_padded = onp.add(lhs_shape[2:], onp.add(*zip(*pads))) out_space = onp.floor_divide( onp.subtract(lhs_padded, rhs_shape[2:]), strides) + 1 out_space = onp.maximum(0, out_space) out_shape = (lhs_shape[0], rhs_shape[0]) + tuple(out_space) return tuple(out_shape)
Utility for convolution dimension permutations relative to Conv HLO. def _conv_general_permutations(self, dimension_numbers): """Utility for convolution dimension permutations relative to Conv HLO.""" lhs_spec, rhs_spec, out_spec = dimension_numbers lhs_char, rhs_char, out_char = ('N', 'C'), ('O', 'I'), ('N', 'C') charpairs = (lhs_char, rhs_char, out_char) for i, (a, b) in enumerate(charpairs): if not (dimension_numbers[i].count(a) == 1 and dimension_numbers[i].count(b) == 1): msg = ('convolution dimension_numbers[{}] must contain the characters ' '"{}" and "{}" exatly once, got {}.') raise TypeError(msg.format(i, a, b, dimension_numbers[i])) if len(dimension_numbers[i]) != len(set(dimension_numbers[i])): msg = ('convolution dimension_numbers[{}] cannot have duplicate ' 'characters, got {}.') raise TypeError(msg.format(i, dimension_numbers[i])) if not (set(lhs_spec) - set(lhs_char) == set(rhs_spec) - set(rhs_char) == set(out_spec) - set(out_char)): msg = ('convolution dimension_numbers elements must each have the same ' 'set of spatial characters, got {}.') raise TypeError(msg.format(dimension_numbers)) def getperm(spec, charpair): spatial = (i for i, c in enumerate(spec) if c not in charpair) if spec is not rhs_spec: spatial = sorted(spatial, key=lambda i: rhs_spec.index(spec[i])) return (spec.index(charpair[0]), spec.index(charpair[1])) + tuple(spatial) lhs_perm, rhs_perm, out_perm = map(getperm, dimension_numbers, charpairs) return lhs_perm, rhs_perm, out_perm
Generalized computation of conv shape. def _conv_general_shape_tuple(self, lhs_shape, rhs_shape, window_strides, padding, dimension_numbers): """Generalized computation of conv shape.""" lhs_perm, rhs_perm, out_perm = self._conv_general_permutations( dimension_numbers) lhs_trans = onp.take(lhs_shape, lhs_perm) rhs_trans = onp.take(rhs_shape, rhs_perm) out_trans = self._conv_shape_tuple( lhs_trans, rhs_trans, window_strides, padding) return tuple(onp.take(out_trans, onp.argsort(out_perm)))
Factory for dopamine agent initialization. Args: agent_kwargs: dict of BatchDQNAgent parameters Returns: Function(sess, environment, summary_writer) -> BatchDQNAgent instance. def get_create_agent(agent_kwargs): """Factory for dopamine agent initialization. Args: agent_kwargs: dict of BatchDQNAgent parameters Returns: Function(sess, environment, summary_writer) -> BatchDQNAgent instance. """ def create_agent(sess, environment, summary_writer=None): """Creates a DQN agent. Simplified version of `dopamine.discrete_domains.train.create_agent` Args: sess: a session environment: an environment summary_writer: a summary writer. Returns: a DQN agent. """ return BatchDQNAgent( env_batch_size=environment.batch_size, sess=sess, num_actions=environment.action_space.n, summary_writer=summary_writer, tf_device="/gpu:*", **agent_kwargs) return create_agent
Factory for dopamine environment initialization function. Args: batch_env_fn: function(in_graph: bool) -> batch environment. time_limit: time steps limit for environment. Returns: function (with optional, unused parameters) initializing environment. def get_create_batch_env_fun(batch_env_fn, time_limit): """Factory for dopamine environment initialization function. Args: batch_env_fn: function(in_graph: bool) -> batch environment. time_limit: time steps limit for environment. Returns: function (with optional, unused parameters) initializing environment. """ def create_env_fun(game_name=None, sticky_actions=None): del game_name, sticky_actions batch_env = batch_env_fn(in_graph=False) batch_env = ResizeBatchObservation(batch_env) # pylint: disable=redefined-variable-type batch_env = DopamineBatchEnv(batch_env, max_episode_steps=time_limit) return batch_env return create_env_fun
Split hparams, based on key prefixes. Args: hparams: hyperparameters Returns: Tuple of hparams for respectably: agent, optimizer, runner, replay_buffer. def _parse_hparams(hparams): """Split hparams, based on key prefixes. Args: hparams: hyperparameters Returns: Tuple of hparams for respectably: agent, optimizer, runner, replay_buffer. """ prefixes = ["agent_", "optimizer_", "runner_", "replay_buffer_"] ret = [] for prefix in prefixes: ret_dict = {} for key in hparams.values(): if prefix in key: par_name = key[len(prefix):] ret_dict[par_name] = hparams.get(key) ret.append(ret_dict) return ret
Build WrappedReplayBuffer with custom OutOfGraphReplayBuffer. def _build_replay_buffer(self, use_staging): """Build WrappedReplayBuffer with custom OutOfGraphReplayBuffer.""" replay_buffer_kwargs = dict( observation_shape=dqn_agent.NATURE_DQN_OBSERVATION_SHAPE, stack_size=dqn_agent.NATURE_DQN_STACK_SIZE, replay_capacity=self._replay_capacity, batch_size=self._buffer_batch_size, update_horizon=self.update_horizon, gamma=self.gamma, extra_storage_types=None, observation_dtype=np.uint8, ) replay_memory = _OutOfGraphReplayBuffer( artificial_done=not self._generates_trainable_dones, **replay_buffer_kwargs) return circular_replay_buffer.WrappedReplayBuffer( wrapped_memory=replay_memory, use_staging=use_staging, **replay_buffer_kwargs)
Append artificial_done to *args and run parent method. def add(self, observation, action, reward, terminal, *args): """Append artificial_done to *args and run parent method.""" # If this will be a problem for maintenance, we could probably override # DQNAgent.add() method instead. artificial_done = self._artificial_done and terminal args = list(args) args.append(artificial_done) return super(_OutOfGraphReplayBuffer, self).add(observation, action, reward, terminal, *args)
Step. def step(self, actions): """Step.""" self._elapsed_steps += 1 obs, rewards, dones = \ [np.array(r) for r in self.batch_env.step(actions)] if self._elapsed_steps > self._max_episode_steps: done = True if self._elapsed_steps > self._max_episode_steps + 1: rewards.fill(0) else: done = dones[0] assert np.all(done == dones), ("Current modifications of Dopamine " "require same number of steps for each " "environment in batch") del dones self.game_over = done return obs, rewards, done, {}
Set of hyperparameters. def text_cnn_base(): """Set of hyperparameters.""" hparams = common_hparams.basic_params1() hparams.batch_size = 4096 hparams.max_length = 256 hparams.clip_grad_norm = 0. # i.e. no gradient clipping hparams.optimizer_adam_epsilon = 1e-9 hparams.learning_rate_schedule = "legacy" hparams.learning_rate_decay_scheme = "noam" hparams.learning_rate = 0.1 hparams.learning_rate_warmup_steps = 4000 hparams.initializer_gain = 1.0 hparams.num_hidden_layers = 6 hparams.initializer = "uniform_unit_scaling" hparams.weight_decay = 0.0 hparams.optimizer_adam_beta1 = 0.9 hparams.optimizer_adam_beta2 = 0.98 hparams.num_sampled_classes = 0 hparams.label_smoothing = 0.1 hparams.shared_embedding_and_softmax_weights = True hparams.symbol_modality_num_shards = 16 # Add new ones like this. hparams.add_hparam("filter_sizes", [2, 3, 4, 5]) hparams.add_hparam("num_filters", 128) hparams.add_hparam("output_dropout", 0.4) return hparams
Hparams for next_frame_glow. def next_frame_glow_hparams(): """Hparams for next_frame_glow.""" hparams = glow.glow_hparams() # Possible modes are conditional and unconditional hparams.add_hparam("gen_mode", "conditional") hparams.add_hparam("learn_top_scale", False) hparams.add_hparam("condition_all_levels", True) # For each video, substitutes "num_input_frames + num_output_frames" with a # randomly sampled patch of length "num_train_frames" during training. # -1 indicates that the entire video is used for training. hparams.add_hparam("num_train_frames", -1) # The following are hparams that model the latent transitions. # Encoder that maps the latents to a Gaussian distribution. # This function is used to model the prior over z_{t}. Can be, # Pointwise -> point-wise multiplication of z_{t-1}. # conv_net -> one-layer convolution over z_{t-1} .. z_{t - num_cond_latents} # conv3d_net or conv_lstm hparams.add_hparam("latent_dist_encoder", "conv_net") # Number of latents used in the encoder above. hparams.add_hparam("num_cond_latents", 1) hparams.add_hparam("latent_architecture", "glow_resnet") hparams.add_hparam("latent_apply_dilations", False) hparams.add_hparam("latent_dilation_rates", [1, 3]) # Use latent skip connections hparams.add_hparam("model_input", False) hparams.add_hparam("cond_first_frame", False) hparams.add_hparam("latent_skip", True) hparams.add_hparam("latent_encoder_depth", 2) hparams.add_hparam("latent_encoder_width", 512) hparams.add_hparam("latent_dropout", 0.0) hparams.add_hparam("latent_pre_output_channels", 512) hparams.add_hparam("latent_activation", "relu") hparams.add_hparam("latent_noise", 0.0) # Pretrains the glow encoder for "pretrain_steps" number of steps. # By default, don't pretrain and learn end-to-end hparams.add_hparam("pretrain_steps", -1) hparams.bottom = { "inputs": modalities.video_raw_bottom, "targets": modalities.video_raw_targets_bottom, } hparams.loss = { "targets": modalities.video_l1_raw_loss, } hparams.top = { "targets": modalities.video_raw_top, } hparams.init_batch_size = 256 hparams.batch_size = 32 # Possible options: are prev_frame, single_conv and normal hparams.top_prior = "single_conv" return hparams
Hparams to reproduce bits-per-pixel results on BAIR action-free dataset. def next_frame_glow_bair_quant(): """Hparams to reproduce bits-per-pixel results on BAIR action-free dataset.""" hparams = next_frame_glow_hparams() hparams.video_num_input_frames = 3 hparams.video_num_target_frames = 10 hparams.num_train_frames = 4 hparams.num_cond_latents = 3 hparams.depth = 24 hparams.latent_dist_encoder = "conv3d_net" hparams.latent_encoder_width = 256 hparams.latent_architecture = "glow_resnet" hparams.latent_encoder_depth = 5 hparams.latent_apply_dilations = True hparams.latent_activation = "gatu" hparams.activation = "gatu" hparams.learning_rate_constant = 3e-4 hparams.learning_rate_schedule = "constant*linear_warmup" hparams.learning_rate_warmup_steps = 10000 hparams.init_batch_size = 128 hparams.batch_size = 5 return hparams
Hparams for qualitative video generation results. def next_frame_glow_bair_qual(): """Hparams for qualitative video generation results.""" hparams = next_frame_glow_bair_quant() hparams.coupling = "additive" hparams.temperature = 0.5 hparams.coupling_width = 392 return hparams
Hparams for qualitative and quantitative results on shapes dataset. def next_frame_glow_shapes(): """Hparams for qualitative and quantitative results on shapes dataset.""" hparams = next_frame_glow_bair_quant() hparams.video_num_input_frames = 1 hparams.video_num_target_frames = 2 hparams.num_train_frames = 2 hparams.num_cond_latents = 1 hparams.coupling = "additive" hparams.coupling_width = 512 hparams.latent_encoder_depth = 10 hparams.latent_skip = False hparams.learning_rate_constant = 1e-4 hparams.batch_size = 10 return hparams
Get z^{cond}_{t} given z^{1..t-1}. Args: all_latents: list of list of tensors, outer-size equals no.of time_steps-1 inner-size equals hparams.n_levels. hparams: See next_frame_glow_hparams. Returns: cond_latents: conditional latents at time-step t. def get_cond_latents(all_latents=None, hparams=None): """Get z^{cond}_{t} given z^{1..t-1}. Args: all_latents: list of list of tensors, outer-size equals no.of time_steps-1 inner-size equals hparams.n_levels. hparams: See next_frame_glow_hparams. Returns: cond_latents: conditional latents at time-step t. """ cond_latents = None if hparams.gen_mode == "conditional": if hparams.latent_dist_encoder in ["conv_net", "conv3d_net"]: num_cond_latents = (hparams.num_cond_latents + int(hparams.cond_first_frame)) if len(all_latents) >= num_cond_latents: cond_latents = all_latents[-hparams.num_cond_latents:] if hparams.cond_first_frame: cond_latents = [all_latents[0]] + cond_latents elif hparams.latent_dist_encoder in ["pointwise", "conv_lstm"]: if all_latents: cond_latents = all_latents[-1] if hparams.gen_mode == "conditional": global_step = tf.train.get_or_create_global_step() condition = tf.greater(global_step, hparams.pretrain_steps) else: condition = tf.constant(False, dtype=tf.bool) return condition, cond_latents
Small fully connected model. def basic_fc_small(): """Small fully connected model.""" hparams = common_hparams.basic_params1() hparams.learning_rate = 0.1 hparams.batch_size = 128 hparams.hidden_size = 256 hparams.num_hidden_layers = 2 hparams.initializer = "uniform_unit_scaling" hparams.initializer_gain = 1.0 hparams.weight_decay = 0.0 hparams.dropout = 0.0 return hparams
A stack of layers. Args: mp: a Parallelism object inputs: a list of Tensors self_attention_bias: list of bias Tensor for self-attention (see common_attention.attention_bias()) layers: a string hparams: hyperparameters for model encoder_output: optional list of tensors encoder_decoder_attention_bias: optional list of tensors Returns: y: a list of Tensors def _layer_stack(mp, inputs, self_attention_bias, layers, hparams, encoder_output=None, encoder_decoder_attention_bias=None): """A stack of layers. Args: mp: a Parallelism object inputs: a list of Tensors self_attention_bias: list of bias Tensor for self-attention (see common_attention.attention_bias()) layers: a string hparams: hyperparameters for model encoder_output: optional list of tensors encoder_decoder_attention_bias: optional list of tensors Returns: y: a list of Tensors """ layers = layers.strip(",").split(",") # scaled_dot_product_attention_with_projections uses a 3d attention bias # (no heads), where multihead_attention uses 4d attention bias. self_attention_bias_3d = mp(tf.squeeze, self_attention_bias, 1) if encoder_decoder_attention_bias is not None: encoder_decoder_attention_bias_3d = mp( tf.squeeze, encoder_decoder_attention_bias, 1) relu_dropout_broadcast_dims = ( common_layers.comma_separated_string_to_integer_list( getattr(hparams, "relu_dropout_broadcast_dims", ""))) mix_size = int(hparams.mix_fraction * hparams.hidden_size) accumulator = inputs x = inputs for layer_num, layer_type in enumerate(layers): with tf.variable_scope("%s_%d" % (layer_type, layer_num)): tf.logging.info("%s_%d" % (layer_type, layer_num)) if layer_type == "a": # accumulate accumulator = mp(tf.add, x, accumulator) x = accumulator elif layer_type == "n": # normalize x = mp(common_layers.apply_norm, x, hparams.norm_type, hparams.hidden_size, hparams.norm_epsilon) elif layer_type == "d": # dropout x = mp(tf.nn.dropout, x, 1.0 - hparams.layer_prepostprocess_dropout) elif layer_type == "m": if mix_size > 0: # mix across shards def _split(t): return tuple(tf.split( t, [mix_size, hparams.hidden_size - mix_size], 2)) to_mix, to_keep = mp(_split, x) mixed = expert_utils.all_reduce_ring(to_mix, mp) mixed = mp(tf.multiply, mixed, mp.n ** -0.5) x = mp(lambda a, b: tf.concat([a, b], 2), mixed, to_keep) elif layer_type == "att": # single-head attention q = mp(tf.layers.dense, x, hparams.hidden_size, use_bias=False, name="q_transform") x = mp( common_attention.scaled_dot_product_attention_simple, q, x, x, self_attention_bias_3d) x = mp(tf.layers.dense, x, hparams.hidden_size, use_bias=False, name="o_transform") elif layer_type == "enc-att": # single-head attention over encoder q = mp(tf.layers.dense, x, hparams.hidden_size, use_bias=False, name="q_transform") assert encoder_output is not None x = mp( common_attention.scaled_dot_product_attention_simple, q, encoder_output, encoder_output, encoder_decoder_attention_bias_3d) x = mp(tf.layers.dense, x, hparams.hidden_size, use_bias=False, name="o_transform") elif layer_type == "multihead-att": # multi-head attention x = mp( common_attention.multihead_attention, x, None, self_attention_bias, # bias hparams.multihead_attention_key_channels or hparams.hidden_size, hparams.multihead_attention_value_channels or hparams.hidden_size, hparams.hidden_size, hparams.multihead_attention_num_heads, hparams.attention_dropout) elif layer_type == "enc-multihead-att": # multi-head attention x = mp( common_attention.multihead_attention, x, encoder_output, encoder_decoder_attention_bias, # bias hparams.multihead_attention_key_channels or hparams.hidden_size, hparams.multihead_attention_value_channels or hparams.hidden_size, hparams.hidden_size, hparams.multihead_attention_num_heads, hparams.attention_dropout) elif layer_type == "ffn": x = mp( common_layers.dense_relu_dense, x, hparams.filter_size, hparams.hidden_size, dropout=hparams.relu_dropout, dropout_broadcast_dims=[relu_dropout_broadcast_dims] * mp.n) else: assert False, "unknown sublayer %s" % layer_type return x
Set of hyperparameters. def transformer_symshard_base(): """Set of hyperparameters.""" hparams = common_hparams.basic_params1() hparams.hidden_size = 256 hparams.batch_size = 2048 hparams.max_length = 0 # All hyperparameters ending in "dropout" are automatically set to 0.0 # when not in training mode. hparams.layer_prepostprocess_dropout = 0.2 hparams.add_hparam("attention_dropout", 0.1) hparams.add_hparam("relu_dropout", 0.0) hparams.add_hparam("relu_dropout_broadcast_dims", "1") hparams.layer_prepostprocess_dropout = 0.1 hparams.layer_prepostprocess_dropout_broadcast_dims = "1" # length hparams.label_smoothing = 0.1 hparams.clip_grad_norm = 0. # i.e. no gradient clipping hparams.optimizer = "Adafactor" hparams.learning_rate_schedule = "rsqrt_decay" hparams.learning_rate_warmup_steps = 10000 hparams.initializer_gain = 1.0 hparams.initializer = "uniform_unit_scaling" hparams.weight_decay = 0.0 # TODO(noam): use this to control sharing. We now share always hparams.shared_embedding_and_softmax_weights = True # we only want one data shard. hparams.no_data_parallelism = True # bypass the symbol modality so that we can use model parallelism. hparams.bottom = { "inputs": modalities.identity_bottom, "targets": modalities.identity_bottom, } hparams.top = { "targets": modalities.identity_top, } hparams.add_hparam("filter_size", 1280) hparams.add_hparam("mix_fraction", 0.5) # attention-related flags hparams.add_hparam("multihead_attention_num_heads", 4) hparams.add_hparam("multihead_attention_key_channels", 0) hparams.add_hparam("multihead_attention_value_channels", 0) hparams.add_hparam("pos", "timing") # timing, none hparams.add_hparam( "encoder_layers", ("n,att,m,d,a," "n,ffn,m,d,a,") * 6 + "n,d") hparams.add_hparam( "decoder_layers", ("n,att,m,d,a," "n,enc-att,m,d,a," "n,ffn,m,d,a,") * 6 + "n,d") # Number of model shards - each one has separate parameters. # Changing this number invalidates checkpoints. hparams.add_hparam("num_model_shards", 8) return hparams
Image generator for Imagenet 64x64 downsampled images. It assumes that the data has been downloaded from http://image-net.org/small/*_32x32.tar or http://image-net.org/small/*_64x64.tar into tmp_dir. Args: tmp_dir: path to temporary storage directory. training: a Boolean; if true, we use the train set, otherwise the test set. size: image size (assumes height and width are same) Yields: A dictionary representing the images with the following fields: * image/encoded: the string encoding the image as JPEG, * image/format: the string "jpeg" representing image format, * image/height: an integer representing the height, * image/width: an integer representing the width. Every field is actually a list of the corresponding type. def imagenet_pixelrnn_generator(tmp_dir, training, size=_IMAGENET_SMALL_IMAGE_SIZE): """Image generator for Imagenet 64x64 downsampled images. It assumes that the data has been downloaded from http://image-net.org/small/*_32x32.tar or http://image-net.org/small/*_64x64.tar into tmp_dir. Args: tmp_dir: path to temporary storage directory. training: a Boolean; if true, we use the train set, otherwise the test set. size: image size (assumes height and width are same) Yields: A dictionary representing the images with the following fields: * image/encoded: the string encoding the image as JPEG, * image/format: the string "jpeg" representing image format, * image/height: an integer representing the height, * image/width: an integer representing the width. Every field is actually a list of the corresponding type. """ if size == _IMAGENET_SMALL_IMAGE_SIZE: train_prefix = _IMAGENET_SMALL_TRAIN_PREFIX eval_prefix = _IMAGENET_SMALL_EVAL_PREFIX else: train_prefix = _IMAGENET_MEDIUM_TRAIN_PREFIX eval_prefix = _IMAGENET_MEDIUM_EVAL_PREFIX prefix = train_prefix if training else eval_prefix images_filepath = os.path.join(tmp_dir, prefix) image_files = tf.gfile.Glob(images_filepath + "/*") height = size width = size const_label = 0 for filename in image_files: with tf.gfile.Open(filename, "r") as f: encoded_image = f.read() yield { "image/encoded": [encoded_image], "image/format": ["png"], "image/class/label": [const_label], "image/height": [height], "image/width": [width] }
Preprocessing used for Imagenet and similar problems. def imagenet_preprocess_example(example, mode, resize_size=None, normalize=True): """Preprocessing used for Imagenet and similar problems.""" resize_size = resize_size or [299, 299] assert resize_size[0] == resize_size[1] image = example["inputs"] if mode == tf.estimator.ModeKeys.TRAIN: image = preprocess_for_train(image, image_size=resize_size[0], normalize=normalize) else: image = preprocess_for_eval(image, image_size=resize_size[0], normalize=normalize) example["inputs"] = image return example
Crops the given image using the provided offsets and sizes. Note that the method doesn't assume we know the input image size but it does assume we know the input image rank. Args: image: `Tensor` image of shape [height, width, channels]. offset_height: `Tensor` indicating the height offset. offset_width: `Tensor` indicating the width offset. crop_height: the height of the cropped image. crop_width: the width of the cropped image. Returns: the cropped (and resized) image. Raises: InvalidArgumentError: if the rank is not 3 or if the image dimensions are less than the crop size. def _crop(image, offset_height, offset_width, crop_height, crop_width): """Crops the given image using the provided offsets and sizes. Note that the method doesn't assume we know the input image size but it does assume we know the input image rank. Args: image: `Tensor` image of shape [height, width, channels]. offset_height: `Tensor` indicating the height offset. offset_width: `Tensor` indicating the width offset. crop_height: the height of the cropped image. crop_width: the width of the cropped image. Returns: the cropped (and resized) image. Raises: InvalidArgumentError: if the rank is not 3 or if the image dimensions are less than the crop size. """ original_shape = tf.shape(image) rank_assertion = tf.Assert( tf.equal(tf.rank(image), 3), ["Rank of image must be equal to 3."]) with tf.control_dependencies([rank_assertion]): cropped_shape = tf.stack([crop_height, crop_width, original_shape[2]]) size_assertion = tf.Assert( tf.logical_and( tf.greater_equal(original_shape[0], crop_height), tf.greater_equal(original_shape[1], crop_width)), ["Crop size greater than the image size."]) offsets = tf.to_int32(tf.stack([offset_height, offset_width, 0])) # Use tf.slice instead of crop_to_bounding box as it accepts tensors to # define the crop size. with tf.control_dependencies([size_assertion]): image = tf.slice(image, offsets, cropped_shape) return tf.reshape(image, cropped_shape)
Generates cropped_image using a one of the bboxes randomly distorted. See `tf.image.sample_distorted_bounding_box` for more documentation. Args: image: `Tensor` of image (it will be converted to floats in [0, 1]). bbox: `Tensor` of bounding boxes arranged `[1, num_boxes, coords]` where each coordinate is [0, 1) and the coordinates are arranged as `[ymin, xmin, ymax, xmax]`. If num_boxes is 0 then use the whole image. min_object_covered: An optional `float`. Defaults to `0.1`. The cropped area of the image must contain at least this fraction of any bounding box supplied. aspect_ratio_range: An optional list of `float`s. The cropped area of the image must have an aspect ratio = width / height within this range. area_range: An optional list of `float`s. The cropped area of the image must contain a fraction of the supplied image within in this range. max_attempts: An optional `int`. Number of attempts at generating a cropped region of the image of the specified constraints. After `max_attempts` failures, return the entire image. scope: Optional `str` for name scope. Returns: (cropped image `Tensor`, distorted bbox `Tensor`). def distorted_bounding_box_crop(image, bbox, min_object_covered=0.1, aspect_ratio_range=(0.75, 1.33), area_range=(0.05, 1.0), max_attempts=100, scope=None): """Generates cropped_image using a one of the bboxes randomly distorted. See `tf.image.sample_distorted_bounding_box` for more documentation. Args: image: `Tensor` of image (it will be converted to floats in [0, 1]). bbox: `Tensor` of bounding boxes arranged `[1, num_boxes, coords]` where each coordinate is [0, 1) and the coordinates are arranged as `[ymin, xmin, ymax, xmax]`. If num_boxes is 0 then use the whole image. min_object_covered: An optional `float`. Defaults to `0.1`. The cropped area of the image must contain at least this fraction of any bounding box supplied. aspect_ratio_range: An optional list of `float`s. The cropped area of the image must have an aspect ratio = width / height within this range. area_range: An optional list of `float`s. The cropped area of the image must contain a fraction of the supplied image within in this range. max_attempts: An optional `int`. Number of attempts at generating a cropped region of the image of the specified constraints. After `max_attempts` failures, return the entire image. scope: Optional `str` for name scope. Returns: (cropped image `Tensor`, distorted bbox `Tensor`). """ with tf.name_scope(scope, default_name="distorted_bounding_box_crop", values=[image, bbox]): # Each bounding box has shape [1, num_boxes, box coords] and # the coordinates are ordered [ymin, xmin, ymax, xmax]. # A large fraction of image datasets contain a human-annotated bounding # box delineating the region of the image containing the object of interest. # We choose to create a new bounding box for the object which is a randomly # distorted version of the human-annotated bounding box that obeys an # allowed range of aspect ratios, sizes and overlap with the human-annotated # bounding box. If no box is supplied, then we assume the bounding box is # the entire image. sample_distorted_bounding_box = tf.image.sample_distorted_bounding_box( tf.shape(image), bounding_boxes=bbox, min_object_covered=min_object_covered, aspect_ratio_range=aspect_ratio_range, area_range=area_range, max_attempts=max_attempts, use_image_if_no_bounding_boxes=True) bbox_begin, bbox_size, distort_bbox = sample_distorted_bounding_box # Crop the image to the specified bounding box. cropped_image = tf.slice(image, bbox_begin, bbox_size) return cropped_image, distort_bbox
Make a random crop of (`size` x `size`). def _random_crop(image, size): """Make a random crop of (`size` x `size`).""" bbox = tf.constant([0.0, 0.0, 1.0, 1.0], dtype=tf.float32, shape=[1, 1, 4]) random_image, bbox = distorted_bounding_box_crop( image, bbox, min_object_covered=0.1, aspect_ratio_range=(3. / 4, 4. / 3.), area_range=(0.08, 1.0), max_attempts=1, scope=None) bad = _at_least_x_are_true(tf.shape(image), tf.shape(random_image), 3) image = tf.cond( bad, lambda: _center_crop(_do_scale(image, size), size), lambda: tf.image.resize_bicubic([random_image], [size, size])[0]) return image
At least `x` of `a` and `b` `Tensors` are true. def _at_least_x_are_true(a, b, x): """At least `x` of `a` and `b` `Tensors` are true.""" match = tf.equal(a, b) match = tf.cast(match, tf.int32) return tf.greater_equal(tf.reduce_sum(match), x)
Rescale the image by scaling the smaller spatial dimension to `size`. def _do_scale(image, size): """Rescale the image by scaling the smaller spatial dimension to `size`.""" shape = tf.cast(tf.shape(image), tf.float32) w_greater = tf.greater(shape[0], shape[1]) shape = tf.cond(w_greater, lambda: tf.cast([shape[0] / shape[1] * size, size], tf.int32), lambda: tf.cast([size, shape[1] / shape[0] * size], tf.int32)) return tf.image.resize_bicubic([image], shape)[0]
Crops to center of image with specified `size`. def _center_crop(image, size): """Crops to center of image with specified `size`.""" image_height = tf.shape(image)[0] image_width = tf.shape(image)[1] offset_height = ((image_height - size) + 1) / 2 offset_width = ((image_width - size) + 1) / 2 image = _crop(image, offset_height, offset_width, size, size) return image
Normalize the image to zero mean and unit variance. def _normalize(image): """Normalize the image to zero mean and unit variance.""" offset = tf.constant(MEAN_RGB, shape=[1, 1, 3]) image -= offset scale = tf.constant(STDDEV_RGB, shape=[1, 1, 3]) image /= scale return image
Preprocesses the given image for evaluation. Args: image: `Tensor` representing an image of arbitrary size. image_size: int, how large the output image should be. normalize: bool, if True the image is normalized. Returns: A preprocessed image `Tensor`. def preprocess_for_train(image, image_size=224, normalize=True): """Preprocesses the given image for evaluation. Args: image: `Tensor` representing an image of arbitrary size. image_size: int, how large the output image should be. normalize: bool, if True the image is normalized. Returns: A preprocessed image `Tensor`. """ if normalize: image = tf.to_float(image) / 255.0 image = _random_crop(image, image_size) if normalize: image = _normalize(image) image = _flip(image) image = tf.reshape(image, [image_size, image_size, 3]) return image
Preprocesses the given image for evaluation. Args: image: `Tensor` representing an image of arbitrary size. image_size: int, how large the output image should be. normalize: bool, if True the image is normalized. Returns: A preprocessed image `Tensor`. def preprocess_for_eval(image, image_size=224, normalize=True): """Preprocesses the given image for evaluation. Args: image: `Tensor` representing an image of arbitrary size. image_size: int, how large the output image should be. normalize: bool, if True the image is normalized. Returns: A preprocessed image `Tensor`. """ if normalize: image = tf.to_float(image) / 255.0 image = _do_scale(image, image_size + 32) if normalize: image = _normalize(image) image = _center_crop(image, image_size) image = tf.reshape(image, [image_size, image_size, 3]) return image
Factor-based learning rate schedule. Interprets factors in the factors string which can consist of: * constant: interpreted as the constant value, * linear_warmup: interpreted as linear warmup until warmup_steps, * rsqrt_decay: divide by square root of max(step, warmup_steps) * decay_every: Every k steps decay the learning rate by decay_factor. Args: history: the history of training and evaluation (History object). factors: a string with factors separated by "*" that defines the schedule. constant: float, the starting constant for the learning rate schedule. warmup_steps: how many steps to warm up for in the warmup schedule. decay_factor: The amount to decay the learning rate by. steps_per_decay: How often to decay the learning rate. Returns: a function learning_rate(step): float -> float, the step-dependent lr. def MultifactorSchedule(history=None, factors="constant * linear_warmup * rsqrt_decay", constant=0.1, warmup_steps=100, decay_factor=0.5, steps_per_decay=20000): """Factor-based learning rate schedule. Interprets factors in the factors string which can consist of: * constant: interpreted as the constant value, * linear_warmup: interpreted as linear warmup until warmup_steps, * rsqrt_decay: divide by square root of max(step, warmup_steps) * decay_every: Every k steps decay the learning rate by decay_factor. Args: history: the history of training and evaluation (History object). factors: a string with factors separated by "*" that defines the schedule. constant: float, the starting constant for the learning rate schedule. warmup_steps: how many steps to warm up for in the warmup schedule. decay_factor: The amount to decay the learning rate by. steps_per_decay: How often to decay the learning rate. Returns: a function learning_rate(step): float -> float, the step-dependent lr. """ del history cache_args = (factors, constant, warmup_steps) if cache_args in _memoized_multifactor_schedules: return _memoized_multifactor_schedules[cache_args] factors = [n.strip() for n in factors.split("*")] def learning_rate(step): # pylint: disable=invalid-name """Step to learning rate function.""" ret = 1.0 for name in factors: if name == "constant": ret *= constant elif name == "linear_warmup": ret *= np.minimum(1.0, step / warmup_steps) elif name == "rsqrt_decay": ret /= np.sqrt(np.maximum(step, warmup_steps)) elif name == "decay_every": ret *= (decay_factor ** (step//steps_per_decay)) else: raise ValueError("Unknown factor %s." % name) return ret _memoized_multifactor_schedules[cache_args] = learning_rate return learning_rate
Learning rate that decreases when eval metric stalls. If the chosen metric does not improve by improvement_margin for as many as steps_to_decrease steps, then the constant gets decreased by decrease rate. Finally, the MultifactorSchedule gets called with the adjusted constant. Args: history: trax.history.History, the history of training and evaluation. constant: float, the starting constant for the learning rate schedule. steps_to_decrease: int, after how many steps without improvement should we decrease the constant. improvement_margin: how much we need to improve to consider the metric improved. decrease_rate: by what fraction to decrease (i.e. lr /= decrease_rate). history_mode: str, which mode of the history to use. metric: which evaluation metric to use for adjustments. Returns: a function learning_rate(step): float -> float, the step-dependent lr. def EvalAdjustingSchedule(history, constant=0.1, steps_to_decrease=20, improvement_margin=0.001, decrease_rate=1.5, history_mode="eval", metric="metrics/accuracy"): """Learning rate that decreases when eval metric stalls. If the chosen metric does not improve by improvement_margin for as many as steps_to_decrease steps, then the constant gets decreased by decrease rate. Finally, the MultifactorSchedule gets called with the adjusted constant. Args: history: trax.history.History, the history of training and evaluation. constant: float, the starting constant for the learning rate schedule. steps_to_decrease: int, after how many steps without improvement should we decrease the constant. improvement_margin: how much we need to improve to consider the metric improved. decrease_rate: by what fraction to decrease (i.e. lr /= decrease_rate). history_mode: str, which mode of the history to use. metric: which evaluation metric to use for adjustments. Returns: a function learning_rate(step): float -> float, the step-dependent lr. """ metrics = history.get(history_mode, metric) adjusted = constant if len(metrics) < 2: return MultifactorSchedule(history, constant=adjusted) steps_without_improvement = 0 cur = metrics.pop()[1] # The most-recent value of the metric. while len(metrics) > 1: # The one-before value of metrics as .pop() removes one element each time. prev = metrics.pop()[1] if cur < prev * (1 + improvement_margin): steps_without_improvement += 1 else: cur = prev steps_without_improvement = 0 if steps_without_improvement >= steps_to_decrease: adjusted /= decrease_rate cur = prev steps_without_improvement = 0 return MultifactorSchedule(history, constant=adjusted)
Project encoder hidden state under num_blocks using projection tensors. Args: x: Encoder hidden state of shape [batch_size, latent_dim, hidden_size]. projection_tensors: Projection tensors used to project the hidden state. hidden_size: Dimension of the latent space. num_blocks: Number of blocks in DVQ. Returns: x_projected: Projected states of shape [batch_size, latent_dim, num_blocks, hidden_size / num_blocks]. def project_hidden(x, projection_tensors, hidden_size, num_blocks): """Project encoder hidden state under num_blocks using projection tensors. Args: x: Encoder hidden state of shape [batch_size, latent_dim, hidden_size]. projection_tensors: Projection tensors used to project the hidden state. hidden_size: Dimension of the latent space. num_blocks: Number of blocks in DVQ. Returns: x_projected: Projected states of shape [batch_size, latent_dim, num_blocks, hidden_size / num_blocks]. """ batch_size, latent_dim, _ = common_layers.shape_list(x) x = tf.reshape(x, shape=[1, -1, hidden_size]) x_tiled = tf.reshape( tf.tile(x, multiples=[num_blocks, 1, 1]), shape=[num_blocks, -1, hidden_size]) x_projected = tf.matmul(x_tiled, projection_tensors) x_projected = tf.transpose(x_projected, perm=[1, 0, 2]) x_4d = tf.reshape(x_projected, [batch_size, latent_dim, num_blocks, -1]) return x_4d
Slice encoder hidden state under num_blocks. Args: x: Encoder hidden state of shape [batch_size, latent_dim, hidden_size]. hidden_size: Dimension of the latent space. num_blocks: Number of blocks in DVQ. Returns: Sliced states of shape [batch_size, latent_dim, num_blocks, block_dim]. def slice_hidden(x, hidden_size, num_blocks): """Slice encoder hidden state under num_blocks. Args: x: Encoder hidden state of shape [batch_size, latent_dim, hidden_size]. hidden_size: Dimension of the latent space. num_blocks: Number of blocks in DVQ. Returns: Sliced states of shape [batch_size, latent_dim, num_blocks, block_dim]. """ batch_size, latent_dim, _ = common_layers.shape_list(x) block_dim = hidden_size // num_blocks x_sliced = tf.reshape(x, shape=[batch_size, latent_dim, num_blocks, block_dim]) return x_sliced
Find the nearest element in means to elements in x. Args: x: Continuous encodings of shape [batch_size, latent_dim, num_blocks, block_dim]. means: Embedding table of shape [num_blocks, block_v_size, block_dim]. block_v_size: Number of table entries per block. random_top_k: Noisy top-k if this is bigger than 1. soft_em: If True then use soft EM rather than hard EM. num_samples: Number of samples to take in soft EM. sum_over_latents: Whether to sum over non-batch dimensions when calculating negative entropy loss. Used only when doing soft EM. summary: If True then record summary histogram of entropies. Returns: Tensor with nearest element in mean encoded in one-hot notation and distances. def nearest_neighbor(x, means, block_v_size, random_top_k=1, soft_em=False, num_samples=1, sum_over_latents=False, summary=True): """Find the nearest element in means to elements in x. Args: x: Continuous encodings of shape [batch_size, latent_dim, num_blocks, block_dim]. means: Embedding table of shape [num_blocks, block_v_size, block_dim]. block_v_size: Number of table entries per block. random_top_k: Noisy top-k if this is bigger than 1. soft_em: If True then use soft EM rather than hard EM. num_samples: Number of samples to take in soft EM. sum_over_latents: Whether to sum over non-batch dimensions when calculating negative entropy loss. Used only when doing soft EM. summary: If True then record summary histogram of entropies. Returns: Tensor with nearest element in mean encoded in one-hot notation and distances. """ batch_size, latent_dim, num_blocks, block_dim = common_layers.shape_list(x) x = tf.reshape(x, [batch_size * latent_dim, num_blocks, block_dim]) x_norm_sq = tf.reduce_sum(tf.square(x), axis=-1, keep_dims=True) means_norm_sq = tf.reduce_sum(tf.square(means), axis=-1, keep_dims=True) scalar_prod = tf.matmul( tf.transpose(x, perm=[1, 0, 2]), tf.transpose(means, perm=[0, 2, 1])) scalar_prod = tf.transpose(scalar_prod, perm=[1, 0, 2]) dist = x_norm_sq + tf.transpose( means_norm_sq, perm=[2, 0, 1]) - 2 * scalar_prod # computing cluster probabilities if soft_em: num_blocks = common_layers.shape_list(dist)[1] nearest_idx = tf.stack( [ tf.multinomial(-dist[:, i, :], num_samples=num_samples) for i in range(num_blocks) ], axis=1) nearest_hot = tf.one_hot(nearest_idx, depth=block_v_size) neg_q_entropy = tf.reduce_sum( nearest_hot * tf.expand_dims(tf.nn.log_softmax(-dist), 2), axis=2) if sum_over_latents: neg_q_entropy = tf.reduce_sum(neg_q_entropy, [1, 2]) neg_q_entropy = tf.reduce_mean(neg_q_entropy, axis=0) nearest_hot = tf.reduce_mean(nearest_hot, axis=-2) if summary: tf.summary.histogram("neg_q_entropy", tf.reshape(neg_q_entropy, [-1])) else: neg_q_entropy = 0. if random_top_k > 1: _, top_k_idx = tf.nn.top_k(-dist, k=random_top_k) nearest_idx = tf.gather( top_k_idx, tf.random_uniform( [1], minval=0, maxval=random_top_k - 1, dtype=tf.int32), axis=-1) else: nearest_idx = tf.argmax(-dist, axis=-1) nearest_hot = tf.one_hot(nearest_idx, block_v_size) return nearest_hot, neg_q_entropy
Compute nearest neighbors and loss for training the embeddings via DVQ. Args: x: Continuous encodings of shape [batch_size, latent_dim, num_blocks, block_dim]. means: Embedding table of shape [num_blocks, block_v_size, block_dim]. num_blocks: Number of blocks in DVQ. block_v_size: Number of table entries per block. bottleneck_kind: Discrete bottleneck type. random_top_k: Noisy top-k if this is bigger than 1. soft_em: If True then use soft EM rather than hard EM. num_samples: Number of samples to use for soft EM. do_hard_gumbel_softmax: Whether to use hard or soft Gumbel-Softmax samples for gumbel-softmax-dvq bottleneck. temperature_warmup_steps: Number of steps it takes to decay temperature to 0. Used only if bottleneck_kind is gumbel-softmax-dvq. num_flows: Number of inverse autoregressive flows for gumbel-softmax-dvq bottleneck. approximate_gs_entropy: Whether to approximate the Gumbel-Softmax density as a categorical distribution when calculating the sample entropy. Used only if bottleneck_kind is gumbel-softmax-dvq. sum_over_latents: Whether to sum over non-batch dimensions when calculating negative entropy loss. Used only if soft EM or when bottleneck_kind is gumbel-softmax-dvq. Returns: x_means_hot: The nearest neighbor in one hot form, with shape [batch_size * latent_dim, num_blocks, block_v_size]. x_means: The nearest neighbor itself, with shape [batch_size * latent_dim, num_blocks, block_dim]. q_loss: Scalar Tensor representing codebook loss. e_loss: Scalar Tensor representing commitment loss. neg_q_entropy: Scalar Tensor representing negative entropy of variational approximation (0 if it is deterministic). def embedding_lookup(x, means, num_blocks, block_v_size, bottleneck_kind="dvq", random_top_k=1, soft_em=False, num_samples=1, do_hard_gumbel_softmax=False, temperature_warmup_steps=150000, num_flows=0, approximate_gs_entropy=False, sum_over_latents=False): """Compute nearest neighbors and loss for training the embeddings via DVQ. Args: x: Continuous encodings of shape [batch_size, latent_dim, num_blocks, block_dim]. means: Embedding table of shape [num_blocks, block_v_size, block_dim]. num_blocks: Number of blocks in DVQ. block_v_size: Number of table entries per block. bottleneck_kind: Discrete bottleneck type. random_top_k: Noisy top-k if this is bigger than 1. soft_em: If True then use soft EM rather than hard EM. num_samples: Number of samples to use for soft EM. do_hard_gumbel_softmax: Whether to use hard or soft Gumbel-Softmax samples for gumbel-softmax-dvq bottleneck. temperature_warmup_steps: Number of steps it takes to decay temperature to 0. Used only if bottleneck_kind is gumbel-softmax-dvq. num_flows: Number of inverse autoregressive flows for gumbel-softmax-dvq bottleneck. approximate_gs_entropy: Whether to approximate the Gumbel-Softmax density as a categorical distribution when calculating the sample entropy. Used only if bottleneck_kind is gumbel-softmax-dvq. sum_over_latents: Whether to sum over non-batch dimensions when calculating negative entropy loss. Used only if soft EM or when bottleneck_kind is gumbel-softmax-dvq. Returns: x_means_hot: The nearest neighbor in one hot form, with shape [batch_size * latent_dim, num_blocks, block_v_size]. x_means: The nearest neighbor itself, with shape [batch_size * latent_dim, num_blocks, block_dim]. q_loss: Scalar Tensor representing codebook loss. e_loss: Scalar Tensor representing commitment loss. neg_q_entropy: Scalar Tensor representing negative entropy of variational approximation (0 if it is deterministic). """ if bottleneck_kind == "gumbel-softmax-dvq": x_means_hot, neg_q_entropy = gumbel_softmax_nearest_neighbor_dvq( x, means, block_v_size, hard=do_hard_gumbel_softmax, num_samples=num_samples, temperature_warmup_steps=temperature_warmup_steps, num_flows=num_flows, approximate_gs_entropy=approximate_gs_entropy, sum_over_latents=sum_over_latents) else: x_means_hot, neg_q_entropy = nearest_neighbor( x, means, block_v_size, random_top_k, soft_em=soft_em, num_samples=num_samples, sum_over_latents=sum_over_latents) x_means_hot_flat = tf.reshape(x_means_hot, [-1, num_blocks, block_v_size]) x_means = tf.matmul(tf.transpose(x_means_hot_flat, perm=[1, 0, 2]), means) x_means = tf.transpose(x_means, [1, 0, 2]) batch_size, latent_dim, num_blocks, block_dim = common_layers.shape_list(x) x = tf.reshape(x, [batch_size * latent_dim, num_blocks, block_dim]) # Currently, we use the mean scaling for the commitment loss, as opposed to # summing across all non-batch dimensions. q_loss = tf.reduce_mean(tf.squared_difference(tf.stop_gradient(x), x_means)) e_loss = tf.reduce_mean(tf.squared_difference(x, tf.stop_gradient(x_means))) return x_means_hot, x_means, q_loss, e_loss, neg_q_entropy