def main(args): tf.logging.set_verbosity(tf.logging.INFO) model_cls = models.get_model(args.model) params = default_parameters() # Import and override parameters # Priorities (low -> high): # default -> saved -> command params = merge_parameters(params, model_cls.get_parameters()) params = import_params(args.output, args.model, params) override_parameters(params, args) # Export all parameters and model specific parameters export_params(params.output, "params.json", params) export_params(params.output, "%s.json" % args.model, collect_params(params, model_cls.get_parameters())) # Build Graph with tf.Graph().as_default(): if not params.record: # Build input queue if params.use_bert and params.bert_emb_path: features = dataset.get_training_input_with_bert( params.input + [params.bert_emb_path], params) else: features = dataset.get_training_input(params.input, params) else: features = record.get_input_features( # ??? os.path.join(params.record, "*train*"), "train", params) # Build model initializer = get_initializer(params) model = model_cls(params) # Multi-GPU setting sharded_losses = parallel.parallel_model( model.get_training_func(initializer), features, params.device_list) loss = tf.add_n(sharded_losses) / len(sharded_losses) # Create global step global_step = tf.train.get_or_create_global_step() # Print parameters all_weights = {v.name: v for v in tf.trainable_variables()} total_size = 0 for v_name in sorted(list(all_weights)): v = all_weights[v_name] tf.logging.info("%s\tshape %s", v.name[:-2].ljust(80), str(v.shape).ljust(20)) v_size = np.prod(np.array( v.shape.as_list())).tolist() # mutiple all dimension size total_size += v_size tf.logging.info("Total trainable variables size: %d", total_size) learning_rate = get_learning_rate_decay(params.learning_rate, global_step, params) learning_rate = tf.convert_to_tensor(learning_rate, dtype=tf.float32) tf.summary.scalar("learning_rate", learning_rate) # Create optimizer opt = tf.train.AdamOptimizer(learning_rate, beta1=params.adam_beta1, beta2=params.adam_beta2, epsilon=params.adam_epsilon) if params.update_cycle == 1: train_op = tf.contrib.layers.optimize_loss( name="training", loss=loss, global_step=global_step, learning_rate=learning_rate, clip_gradients=params.clip_grad_norm or None, optimizer=opt, colocate_gradients_with_ops=True) zero_op = tf.no_op("zero_op") collect_op = tf.no_op("collect_op") else: grads_and_vars = opt.compute_gradients( loss, colocate_gradients_with_ops=True) gradients = [item[0] for item in grads_and_vars] variables = [item[1] for item in grads_and_vars] variables = utils.replicate_variables(variables) zero_op = utils.zero_variables(variables) collect_op = utils.collect_gradients(gradients, variables) scale = 1.0 / params.update_cycle gradients, variables = utils.scale_gradients(grads_and_vars, scale) # Gradient clipping avoid greadient explosion!! if isinstance(params.clip_grad_norm or None, float): gradients, _ = tf.clip_by_global_norm(gradients, params.clip_grad_norm) # Update variables grads_and_vars = list(zip(gradients, variables)) with tf.control_dependencies([collect_op]): train_op = opt.apply_gradients(grads_and_vars, global_step) # Validation ''' if params.validation and params.references[0]: files = [params.validation] + list(params.references) eval_inputs = files eval_input_fn = dataset.get_evaluation_input else: print("Don't evaluate") eval_input_fn = None ''' # Add hooks train_hooks = [ tf.train.StopAtStepHook(last_step=params.train_steps), tf.train.NanTensorHook( loss ), # Monitors the loss tensor and stops training if loss is NaN tf.train.LoggingTensorHook( { "step": global_step, "loss": loss, "chars": tf.shape(features["chars"]), "source": tf.shape(features["source"]), #"bert": tf.shape(features["bert"]), "lr": learning_rate }, every_n_iter=1), tf.train.CheckpointSaverHook( checkpoint_dir=params.output, save_secs=params.save_checkpoint_secs or None, save_steps=params.save_checkpoint_steps or None, saver=tf.train.Saver(max_to_keep=params.keep_checkpoint_max, sharded=False)) ] config = session_config(params) ''' if not eval_input_fn is None: train_hooks.append( hooks.EvaluationHook( lambda f: search.create_inference_graph( model.get_evaluation_func(), f, params ), lambda: eval_input_fn(eval_inputs, params), lambda x: decode_target_ids(x, params), params.output, config, params.keep_top_checkpoint_max, eval_secs=params.eval_secs, eval_steps=params.eval_steps ) ) ''' with tf.train.MonitoredTrainingSession(checkpoint_dir=params.output, hooks=train_hooks, save_checkpoint_secs=None, config=config) as sess: while not sess.should_stop(): utils.session_run(sess, zero_op) for i in range(1, params.update_cycle): utils.session_run(sess, collect_op) sess.run(train_op)
def main(args): if args.distribute: distribute.enable_distributed_training() tf.logging.set_verbosity(tf.logging.INFO) model_cls = models.get_model(args.model) params = default_parameters() # Import and override parameters # Priorities (low -> high): # default -> saved -> command params = merge_parameters(params, model_cls.get_parameters()) params = import_params(args.output, args.model, params) override_parameters(params, args) # Export all parameters and model specific parameters if distribute.rank() == 0: export_params(params.output, "params.json", params) export_params(params.output, "%s.json" % args.model, collect_params(params, model_cls.get_parameters())) # Build Graph with tf.Graph().as_default(): if not params.record: # Build input queue features = dataset.get_training_input(params.input, params) else: features = record.get_input_features( os.path.join(params.record, "*train*"), "train", params) # Build model initializer = get_initializer(params) regularizer = tf.contrib.layers.l1_l2_regularizer( scale_l1=params.scale_l1, scale_l2=params.scale_l2) model = model_cls(params) # Create global step global_step = tf.train.get_or_create_global_step() dtype = tf.float16 if args.half else None # Multi-GPU setting sharded_losses = parallel.parallel_model( model.get_training_func(initializer, regularizer, dtype), features, params.device_list) loss = tf.add_n(sharded_losses) / len(sharded_losses) loss = loss + tf.losses.get_regularization_loss() if distribute.rank() == 0: print_variables() learning_rate = get_learning_rate_decay(params.learning_rate, global_step, params) learning_rate = tf.convert_to_tensor(learning_rate, dtype=tf.float32) tf.summary.scalar("loss", loss) tf.summary.scalar("learning_rate", learning_rate) # Create optimizer if params.optimizer == "Adam": opt = tf.train.AdamOptimizer(learning_rate, beta1=params.adam_beta1, beta2=params.adam_beta2, epsilon=params.adam_epsilon) elif params.optimizer == "LazyAdam": opt = tf.contrib.opt.LazyAdamOptimizer(learning_rate, beta1=params.adam_beta1, beta2=params.adam_beta2, epsilon=params.adam_epsilon) else: raise RuntimeError("Optimizer %s not supported" % params.optimizer) opt = optimizers.MultiStepOptimizer(opt, params.update_cycle) if args.half: opt = optimizers.LossScalingOptimizer(opt, params.loss_scale) # Optimization grads_and_vars = opt.compute_gradients( loss, colocate_gradients_with_ops=True) if params.clip_grad_norm: grads, var_list = list(zip(*grads_and_vars)) grads, _ = tf.clip_by_global_norm(grads, params.clip_grad_norm) grads_and_vars = zip(grads, var_list) train_op = opt.apply_gradients(grads_and_vars, global_step=global_step) # Validation if params.validation and params.references[0]: files = [params.validation] + list(params.references) eval_inputs = dataset.sort_and_zip_files(files) eval_input_fn = dataset.get_evaluation_input else: eval_input_fn = None # Hooks train_hooks = [ tf.train.StopAtStepHook(last_step=params.train_steps), tf.train.NanTensorHook(loss), tf.train.LoggingTensorHook( { "step": global_step, "loss": loss, "source": tf.shape(features["source"]), "target": tf.shape(features["target"]) }, every_n_iter=1) ] broadcast_hook = distribute.get_broadcast_hook() if broadcast_hook: train_hooks.append(broadcast_hook) if distribute.rank() == 0: # Add hooks save_vars = tf.trainable_variables() + [global_step] saver = tf.train.Saver( var_list=save_vars if params.only_save_trainable else None, max_to_keep=params.keep_checkpoint_max, sharded=False) tf.add_to_collection(tf.GraphKeys.SAVERS, saver) train_hooks.append( hooks.MultiStepHook(tf.train.CheckpointSaverHook( checkpoint_dir=params.output, save_secs=params.save_checkpoint_secs or None, save_steps=params.save_checkpoint_steps or None, saver=saver), step=params.update_cycle)) if eval_input_fn is not None: train_hooks.append( hooks.MultiStepHook(hooks.EvaluationHook( lambda f: inference.create_inference_graph([model], f, params), lambda: eval_input_fn(eval_inputs, params), lambda x: decode_target_ids(x, params), params.output, session_config(params), device_list=params.device_list, max_to_keep=params.keep_top_checkpoint_max, eval_secs=params.eval_secs, eval_steps=params.eval_steps), step=params.update_cycle)) checkpoint_dir = params.output else: checkpoint_dir = None restore_op = restore_variables(args.checkpoint) def restore_fn(step_context): step_context.session.run(restore_op) # Create session, do not use default CheckpointSaverHook with tf.train.MonitoredTrainingSession( checkpoint_dir=checkpoint_dir, hooks=train_hooks, save_checkpoint_secs=None, config=session_config(params)) as sess: # Restore pre-trained variables sess.run_step_fn(restore_fn) while not sess.should_stop(): sess.run(train_op)
def main(args): tf.logging.set_verbosity(tf.logging.INFO) model_cls = models.get_model(args.model) params = default_parameters() # Import and override parameters # Priorities (low -> high): # default -> saved -> command params = merge_parameters(params, model_cls.get_parameters()) params = import_params(args.output, args.model, params) override_parameters(params, args) # Export all parameters and model specific parameters export_params(params.output, "params.json", params) export_params(params.output, "%s.json" % args.model, collect_params(params, model_cls.get_parameters())) # Build Graph with tf.Graph().as_default(): if not params.record: # Build input queue features = dataset.get_training_input(params.input, params) else: features = record.get_input_features( os.path.join(params.record, "*train*"), "train", params) update_cycle = params.update_cycle features, init_op = cache.cache_features(features, update_cycle) # Build model initializer = get_initializer(params) regularizer = tf.contrib.layers.l1_l2_regularizer( scale_l1=params.scale_l1, scale_l2=params.scale_l2) model = model_cls(params) # Create global step global_step = tf.train.get_or_create_global_step() # Multi-GPU setting sharded_losses = parallel.parallel_model( model.get_training_func(initializer, regularizer), features, params.device_list) loss = tf.add_n(sharded_losses) / len(sharded_losses) loss = loss + tf.losses.get_regularization_loss() # Print parameters all_weights = {v.name: v for v in tf.trainable_variables()} total_size = 0 for v_name in sorted(list(all_weights)): v = all_weights[v_name] tf.logging.info("%s\tshape %s", v.name[:-2].ljust(80), str(v.shape).ljust(20)) v_size = np.prod(np.array(v.shape.as_list())).tolist() total_size += v_size tf.logging.info("Total trainable variables size: %d", total_size) learning_rate = get_learning_rate_decay(params.learning_rate, global_step, params) learning_rate = tf.convert_to_tensor(learning_rate, dtype=tf.float32) tf.summary.scalar("learning_rate", learning_rate) # Create optimizer if params.optimizer == "Adam": opt = tf.train.AdamOptimizer(learning_rate, beta1=params.adam_beta1, beta2=params.adam_beta2, epsilon=params.adam_epsilon) elif params.optimizer == "LazyAdam": opt = tf.contrib.opt.LazyAdamOptimizer(learning_rate, beta1=params.adam_beta1, beta2=params.adam_beta2, epsilon=params.adam_epsilon) else: raise RuntimeError("Optimizer %s not supported" % params.optimizer) loss, ops = optimize.create_train_op(loss, opt, global_step, params) restore_op = restore_variables(args.checkpoint) # Validation if params.validation and params.references[0]: files = [params.validation] + list(params.references) eval_inputs = dataset.sort_and_zip_files(files) eval_input_fn = dataset.get_evaluation_input else: eval_input_fn = None # Add hooks save_vars = tf.trainable_variables() + [global_step] saver = tf.train.Saver( var_list=save_vars if params.only_save_trainable else None, max_to_keep=params.keep_checkpoint_max, sharded=False) tf.add_to_collection(tf.GraphKeys.SAVERS, saver) multiplier = tf.convert_to_tensor([update_cycle, 1]) train_hooks = [ tf.train.StopAtStepHook(last_step=params.train_steps), tf.train.NanTensorHook(loss), tf.train.LoggingTensorHook( { "step": global_step, "loss": loss, "source": tf.shape(features["source"]) * multiplier, "target": tf.shape(features["target"]) * multiplier }, every_n_iter=1), tf.train.CheckpointSaverHook( checkpoint_dir=params.output, save_secs=params.save_checkpoint_secs or None, save_steps=params.save_checkpoint_steps or None, saver=saver) ] config = session_config(params) if eval_input_fn is not None: train_hooks.append( hooks.EvaluationHook( lambda f: inference.create_inference_graph([model], f, params), lambda: eval_input_fn(eval_inputs, params), lambda x: decode_target_ids(x, params), params.output, config, params.keep_top_checkpoint_max, eval_secs=params.eval_secs, eval_steps=params.eval_steps)) def restore_fn(step_context): step_context.session.run(restore_op) def step_fn(step_context): # Bypass hook calls step_context.session.run([init_op, ops["zero_op"]]) for i in range(update_cycle - 1): step_context.session.run(ops["collect_op"]) return step_context.run_with_hooks(ops["train_op"]) # Create session, do not use default CheckpointSaverHook with tf.train.MonitoredTrainingSession(checkpoint_dir=params.output, hooks=train_hooks, save_checkpoint_secs=None, config=config) as sess: # Restore pre-trained variables sess.run_step_fn(restore_fn) while not sess.should_stop(): sess.run_step_fn(step_fn)
def main(args): tf.logging.set_verbosity(tf.logging.INFO) model_cls = models.get_model(args.model) params = default_parameters() # Import and override parameters # Priorities (low -> high): # default -> saved -> command params = merge_parameters(params, model_cls.get_parameters()) params = import_params(args.output, args.model, params) override_parameters(params, args) # Export all parameters and model specific parameters export_params(params.output, "params.json", params) export_params(params.output, "%s.json" % args.model, collect_params(params, model_cls.get_parameters())) # Build Graph with tf.Graph().as_default(): if not params.record: # Build input queue features = dataset.get_training_input(params.input, params) else: features = record.get_input_features( os.path.join(params.record, "*train*"), "train", params) # Build model initializer = get_initializer(params) model = model_cls(params) if params.MRT: assert params.batch_size == 1 features = mrt_utils.get_MRT(features, params, model) # Multi-GPU setting sharded_losses = parallel.parallel_model( model.get_training_func(initializer), features, params.device_list) loss = tf.add_n(sharded_losses) / len(sharded_losses) # Create global step global_step = tf.train.get_or_create_global_step() # Print parameters all_weights = {v.name: v for v in tf.trainable_variables()} total_size = 0 for v_name in sorted(list(all_weights)): v = all_weights[v_name] tf.logging.info("%s\tshape %s", v.name[:-2].ljust(80), str(v.shape).ljust(20)) v_size = np.prod(np.array(v.shape.as_list())).tolist() total_size += v_size tf.logging.info("Total trainable variables size: %d", total_size) learning_rate = get_learning_rate_decay(params.learning_rate, global_step, params) learning_rate = tf.convert_to_tensor(learning_rate, dtype=tf.float32) tf.summary.scalar("learning_rate", learning_rate) # Create optimizer opt = tf.train.AdamOptimizer(learning_rate, beta1=params.adam_beta1, beta2=params.adam_beta2, epsilon=params.adam_epsilon) train_op = tf.contrib.layers.optimize_loss( name="training", loss=loss, global_step=global_step, learning_rate=learning_rate, clip_gradients=params.clip_grad_norm or None, optimizer=opt, colocate_gradients_with_ops=True) # Validation if params.validation and params.references[0]: files = [params.validation] + list(params.references) eval_inputs = dataset.sort_and_zip_files(files) eval_input_fn = dataset.get_evaluation_input else: eval_input_fn = None # Add hooks train_hooks = [ tf.train.StopAtStepHook(last_step=params.train_steps), tf.train.NanTensorHook(loss), tf.train.LoggingTensorHook( { "step": global_step, "loss": loss, "source": tf.shape(features["source"]), "target": tf.shape(features["target"]) }, every_n_iter=1), tf.train.CheckpointSaverHook( checkpoint_dir=params.output, save_secs=params.save_checkpoint_secs or None, save_steps=params.save_checkpoint_steps or None, saver=tf.train.Saver(max_to_keep=params.keep_checkpoint_max, sharded=False)) ] config = session_config(params) if eval_input_fn is not None: train_hooks.append( hooks.EvaluationHook( lambda f: search.create_inference_graph( model.get_evaluation_func(), f, params), lambda: eval_input_fn(eval_inputs, params), lambda x: decode_target_ids(x, params), params.output, config, params.keep_top_checkpoint_max, eval_secs=params.eval_secs, eval_steps=params.eval_steps)) # Create session, do not use default CheckpointSaverHook with tf.train.MonitoredTrainingSession(checkpoint_dir=params.output, hooks=train_hooks, save_checkpoint_secs=None, config=config) as sess: while not sess.should_stop(): sess.run(train_op)
def main(args): tf.logging.set_verbosity(tf.logging.INFO) # Load configs model_cls_list = [models.get_model(model) for model in args.models] params_list = [default_parameters() for _ in range(len(model_cls_list))] params_list = [ merge_parameters(params, model_cls.get_parameters()) for params, model_cls in zip(params_list, model_cls_list) ] params_list = [ import_params(args.checkpoints[i], args.models[i], params_list[i]) for i in range(len(args.checkpoints)) ] params_list = [ override_parameters(params_list[i], args) for i in range(len(model_cls_list)) ] # Build Graph with tf.Graph().as_default(): model_var_lists = [] # Load checkpoints for i, checkpoint in enumerate(args.checkpoints): print("Loading %s" % checkpoint) var_list = tf.train.list_variables(checkpoint) values = {} reader = tf.train.load_checkpoint(checkpoint) for (name, shape) in var_list: if not name.startswith(model_cls_list[i].get_name()): continue if name.find("losses_avg") >= 0: continue tensor = reader.get_tensor(name) values[name] = tensor model_var_lists.append(values) # Build models model_fns = [] for i in range(len(args.checkpoints)): name = model_cls_list[i].get_name() model = model_cls_list[i](params_list[i], name + "_%d" % i) model_fn = model.get_relevance_func() model_fns.append(model_fn) params = params_list[0] # Build input queue features = dataset.get_training_input(args.input, params) relevances = model_fns[0](features, params) assign_ops = [] all_var_list = tf.trainable_variables() for i in range(len(args.checkpoints)): un_init_var_list = [] name = model_cls_list[i].get_name() for v in all_var_list: if v.name.startswith(name + "_%d" % i): un_init_var_list.append(v) ops = set_variables(un_init_var_list, model_var_lists[i], name + "_%d" % i) assign_ops.extend(ops) assign_op = tf.group(*assign_ops) sess_creator = tf.train.ChiefSessionCreator( config=session_config(params)) results = [] num = 10 count = 0 hooks = [tf.train.LoggingTensorHook({}, every_n_iter=1)] with tf.train.MonitoredSession(session_creator=sess_creator, hooks=hooks) as sess: # Restore variables sess.run(assign_op) src_seq, trg_seq, rlv_info, loss = sess.run(relevances) start = time.time() while count < num: #not sess.should_stop(): src_seq, trg_seq, rlv_info, loss = sess.run(relevances) print('--result--') print('loss:', loss) for i in range(src_seq.shape[0]): src = to_text(params.vocabulary["source"], params.mapping["source"], src_seq[i], params) trg = to_text(params.vocabulary["target"], params.mapping["target"], trg_seq[i], params) print('sentence %d' % i) print('src:', src) print('src_idx:', src_seq[i]) print('trg:', trg) print('trg_idx:', trg_seq[i]) print('result:', rlv_info["result"][i]) count += 1 end = time.time() print('total time:', end - start)
def main(args): tf.logging.set_verbosity(tf.logging.INFO) model_cls = models.get_model(args.model) params = default_parameters() params = merge_parameters(params, model_cls.get_parameters()) params = import_params(args.output, args.model, params) override_parameters(params, args) export_params(params.output, "params.json", params) export_params(params.output, "%s.json" % args.model, collect_params(params, model_cls.get_parameters())) with tf.Graph().as_default(): features = dataset.get_training_input(params.input, params) update_cycle = params.update_cycle features, init_op = cache.cache_features(features, update_cycle) initializer = get_initializer(params) regularizer = tf.contrib.layers.l1_l2_regularizer( scale_l1=params.scale_l1, scale_l2=params.scale_l2) model = model_cls(params) global_step = tf.train.get_or_create_global_step() sharded_losses = parallel.parallel_model( model.get_training_func(initializer, regularizer), features, params.device_list) loss = tf.add_n(sharded_losses) / len(sharded_losses) loss = loss + tf.losses.get_regularization_loss() all_weights = {v.name: v for v in tf.trainable_variables()} total_size = 0 for v_name in sorted(list(all_weights)): v = all_weights[v_name] tf.logging.info("%s\tshape %s", v.name[:-2].ljust(80), str(v.shape).ljust(20)) v_size = np.prod(np.array(v.shape.as_list())).tolist() total_size += v_size tf.logging.info("Total trainable variables size: %d", total_size) learning_rate = get_learning_rate_decay(params.learning_rate, global_step, params) learning_rate = tf.convert_to_tensor(learning_rate, dtype=tf.float32) tf.summary.scalar("learning_rate", learning_rate) if params.optimizer == "Adam": opt = tf.train.AdamOptimizer(learning_rate, beta1=params.adam_beta1, beta2=params.adam_beta2, epsilon=params.adam_epsilon) elif params.optimizer == "LazyAdam": opt = tf.contrib.opt.LazyAdamOptimizer(learning_rate, beta1=params.adam_beta1, beta2=params.adam_beta2, epsilon=params.adam_epsilon) elif params.optimizer == "SGD": opt = tf.train.GradientDescentOptimizer(learning_rate) else: raise RuntimeError("Optimizer %s not supported" % params.optimizer) loss, ops = optimize.create_train_op(loss, opt, global_step, params) restore_op = restore_variables(args.checkpoint) if params.validation: eval_sorted_keys, eval_inputs = dataset.read_eval_input_file( params.validation) eval_input_fn = dataset.get_predict_input else: eval_input_fn = None save_vars = tf.trainable_variables() + [global_step] saver = tf.train.Saver( var_list=save_vars if params.only_save_trainable else None, max_to_keep=params.keep_checkpoint_max, sharded=False) tf.add_to_collection(tf.GraphKeys.SAVERS, saver) multiplier = tf.convert_to_tensor([update_cycle, 1]) train_hooks = [ tf.train.StopAtStepHook(last_step=params.train_steps), tf.train.NanTensorHook(loss), tf.train.LoggingTensorHook( { "step": global_step, "loss": loss, "text": tf.shape(features["text"]) * multiplier, "aspect": tf.shape(features["aspect"]) * multiplier, "polarity": tf.shape(features["polarity"]) * multiplier }, every_n_iter=1), tf.train.CheckpointSaverHook( checkpoint_dir=params.output, save_secs=params.save_checkpoint_secs or None, save_steps=params.save_checkpoint_steps or None, saver=saver) ] config = session_config(params) if eval_input_fn is not None: train_hooks.append( hooks.EvaluationHook( lambda f: inference.create_predict_graph([model], f, params ), lambda: eval_input_fn(eval_inputs, params), params.output, config, params.keep_top_checkpoint_max, eval_secs=params.eval_secs, eval_steps=params.eval_steps)) def restore_fn(step_context): step_context.session.run(restore_op) def step_fn(step_context): step_context.session.run([init_op, ops["zero_op"]]) for i in range(update_cycle - 1): step_context.session.run(ops["collect_op"]) return step_context.run_with_hooks(ops["train_op"]) with tf.train.MonitoredTrainingSession(checkpoint_dir=params.output, hooks=train_hooks, save_checkpoint_secs=None, config=config) as sess: sess.run_step_fn(restore_fn) while not sess.should_stop(): sess.run_step_fn(step_fn)