# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Fine-tuning on regression/classification tasks.""" from __future__ import absolute_import from __future__ import division from __future__ import print_function import six import sys if six.PY2: reload(sys) sys.setdefaultencoding('utf8') import os import time import json import argparse import numpy as np import subprocess import multiprocessing from scipy.stats import pearsonr import paddle import paddle.fluid as fluid import reader.cls as reader from model.xlnet import XLNetConfig from model.classifier import create_model from optimization import optimization from utils.args import ArgumentGroup, print_arguments, check_cuda from utils.init import init_pretraining_params, init_checkpoint from utils.cards import get_cards num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1)) # yapf: disable parser = argparse.ArgumentParser(__doc__) model_g = ArgumentGroup(parser, "model", "model configuration and paths.") model_g.add_arg("model_config_path", str, None, "Path to the json file for bert model config.") model_g.add_arg("dropout", float, 0.1, "Dropout rate.") model_g.add_arg("dropatt", float, 0.1, "Attention dropout rate.") model_g.add_arg("clamp_len", int, -1, "Clamp length.") model_g.add_arg("summary_type", str, "last", "Method used to summarize a sequence into a vector.", choices=['last']) model_g.add_arg("use_summ_proj", bool, True, "Whether to use projection for summarizing sequences.") model_g.add_arg("spiece_model_file", str, None, "Sentence Piece model path.") model_g.add_arg("init_checkpoint", str, None, "Init checkpoint to resume training from.") model_g.add_arg("init_pretraining_params", str, None, "Init pre-training params which preforms fine-tuning from. If the " "arg 'init_checkpoint' has been set, this argument wouldn't be valid.") model_g.add_arg("checkpoints", str, "checkpoints", "Path to save checkpoints.") init_g = ArgumentGroup(parser, "init", "parameter initialization options.") init_g.add_arg("init", str, "normal", "Initialization method.", choices=["normal", "uniform"]) init_g.add_arg("init_std", str, 0.02, "Initialization std when init is normal.") init_g.add_arg("init_range", str, 0.1, "Initialization std when init is uniform.") train_g = ArgumentGroup(parser, "training", "training options.") train_g.add_arg("epoch", int, 3, "Number of epoches for fine-tuning.") train_g.add_arg("learning_rate", float, 5e-5, "Learning rate used to train with warmup.") train_g.add_arg("lr_scheduler", str, "linear_warmup_decay", "scheduler of learning rate.", choices=['linear_warmup_decay', 'noam_decay']) train_g.add_arg("weight_decay", float, 0.01, "Weight decay rate for L2 regularizer.") train_g.add_arg("lr_layer_decay_rate", float, 1.0, "Top layer: lr[L] = args.learning_rate. " "Lower layers: lr[l-1] = lr[l] * lr_layer_decay_rate.") train_g.add_arg("save_steps", int, 10000, "The steps interval to save checkpoints.") train_g.add_arg("train_batch_size", int, 8, "Total examples' number in batch for training.") train_g.add_arg("eval_batch_size", int, 128, "Total examples' number in batch for development.") train_g.add_arg("predict_batch_size", int, 128, "Total examples' number in batch for prediction.") train_g.add_arg("train_steps", int, 1000, "The total steps for training.") train_g.add_arg("warmup_steps", int, 1000, "The steps for warmup.") train_g.add_arg("validation_steps", int, 1000, "The steps interval to evaluate model performance.") log_g = ArgumentGroup(parser, "logging", "logging related.") log_g.add_arg("skip_steps", int, 10, "The steps interval to print loss.") log_g.add_arg("verbose", bool, False, "Whether to output verbose log.") data_g = ArgumentGroup(parser, "data", "Data paths, vocab paths and data processing options") data_g.add_arg("data_dir", str, None, "Path to training data.") data_g.add_arg("predict_dir", str, None, "Path to write predict results.") data_g.add_arg("predict_threshold", float, 0.0, "Threshold for binary prediction.") data_g.add_arg("max_seq_length", int, 512, "Number of words of the longest seqence.") data_g.add_arg("uncased", bool, True, "Whether to lower case the input text. Should be True for uncased models and False for cased models.") data_g.add_arg("random_seed", int, 0, "Random seed.") run_type_g = ArgumentGroup(parser, "run_type", "running type options.") run_type_g.add_arg("use_cuda", bool, True, "If set, use GPU for training.") run_type_g.add_arg("use_fast_executor", bool, False, "If set, use fast parallel executor (in experiment).") run_type_g.add_arg("shuffle", bool, True, "") run_type_g.add_arg("task_name", str, None, "The name of task to perform fine-tuning, should be in {'xnli', 'mnli', 'cola', 'mrpc'}.") run_type_g.add_arg("is_regression", str, None, "Whether it's a regression task.") run_type_g.add_arg("do_train", bool, True, "Whether to perform training.") run_type_g.add_arg("do_eval", bool, True, "Whether to perform evaluation on dev data set.") run_type_g.add_arg("do_predict", bool, True, "Whether to perform evaluation on test data set.") run_type_g.add_arg("eval_split", str, "dev", "Could be dev or test") parser.add_argument("--enable_ce", action='store_true', help="The flag indicating whether to run the task for continuous evaluation.") args = parser.parse_args() # yapf: enable. def evaluate(exe, predict_program, test_data_loader, fetch_list, eval_phase, num_examples): test_data_loader.start() total_cost, total_num_seqs = [], [] all_logits, all_labels = [], [] time_begin = time.time() total_steps = int(num_examples / args.eval_batch_size) steps = 0 while True: try: np_loss, np_num_seqs, np_logits, np_labels = exe.run(program=predict_program, fetch_list=fetch_list) total_cost.extend(np_loss * np_num_seqs) total_num_seqs.extend(np_num_seqs) all_logits.extend(np_logits) all_labels.extend(np_labels) if steps % (int(total_steps / 10)) == 0: print("Evaluation [{}/{}]".format(steps, total_steps)) steps += 1 except fluid.core.EOFException: test_data_loader.reset() break all_logits = np.array(all_logits) all_labels = np.array(all_labels) if args.is_regression: key = "eval_pearsonr" eval_result, _ = pearsonr(all_logits, all_labels) else: key = "eval_accuracy" pred = np.argmax(all_logits, axis=1).reshape(all_labels.shape) eval_result = np.sum(pred == all_labels) / float(all_labels.size) time_end = time.time() print("[%s evaluation] ave loss: %f, %s: %f, elapsed time: %f s" % (eval_phase, np.sum(total_cost) / np.sum(total_num_seqs), key, eval_result, time_end - time_begin)) def predict(exe, predict_program, test_data_loader, task_name, label_list, fetch_list): test_data_loader.start() pred_cnt = 0 predict_results = [] with open(os.path.join(args.predict_dir, "{}.tsv".format( task_name)), "w") as fout: fout.write("index\tprediction\n") while True: try: np_logits = exe.run(program=predict_program, fetch_list=fetch_list) for result in np_logits[0]: if pred_cnt % 1000 == 0: print("Predicting submission for example: {}".format( pred_cnt)) logits = [float(x) for x in result.flat] predict_results.append(logits) if len(logits) == 1: label_out = logits[0] elif len(logits) == 2: if logits[1] - logits[0] > args.predict_threshold: label_out = label_list[1] else: label_out = label_list[0] elif len(logits) > 2: max_index = np.argmax(np.array(logits, dtype=np.float32)) label_out = label_list[max_index] else: raise NotImplementedError fout.write("{}\t{}\n".format(pred_cnt, label_out)) pred_cnt += 1 except fluid.core.EOFException: test_data_loader.reset() break predict_json_path = os.path.join(args.predict_dir, "{}.logits.json".format( task_name)) with open(predict_json_path, "w") as fp: json.dump(predict_results, fp, indent=4) def get_device_num(): # NOTE(zcd): for multi-processe training, each process use one GPU card. if num_trainers > 1 : return 1 visible_device = os.environ.get('CUDA_VISIBLE_DEVICES', None) if visible_device: device_num = len(visible_device.split(',')) else: device_num = subprocess.check_output(['nvidia-smi','-L']).decode().count('\n') return device_num def main(args): if not (args.do_train or args.do_eval or args.do_predict): raise ValueError("For args `do_train`, `do_eval` and `do_predict`, at " "least one of them must be True.") if args.do_predict and not args.predict_dir: raise ValueError("args 'predict_dir' should be given when doing predict") if not os.path.exists(args.predict_dir): os.makedirs(args.predict_dir) xlnet_config = XLNetConfig(args.model_config_path) xlnet_config.print_config() if args.use_cuda: place = fluid.CUDAPlace(int(os.getenv('FLAGS_selected_gpus', '0'))) dev_count = get_device_num() else: place = fluid.CPUPlace() dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count())) exe = fluid.Executor(place) task_name = args.task_name.lower() processors = { "mnli_matched": reader.MnliMatchedProcessor, "mnli_mismatched": reader.MnliMismatchedProcessor, 'sts-b': reader.StsbProcessor, 'imdb': reader.ImdbProcessor, "yelp5": reader.Yelp5Processor } processor = processors[task_name](args) label_list = processor.get_labels() if not args.is_regression else None num_labels = len(label_list) if label_list is not None else None train_program = fluid.Program() startup_prog = fluid.Program() if args.random_seed is not None: startup_prog.random_seed = args.random_seed train_program.random_seed = args.random_seed if args.do_train: # NOTE: If num_trainers > 1, the shuffle_seed must be set, because # the order of batch data generated by reader # must be the same in the respective processes. shuffle_seed = 1 if num_trainers > 1 else None train_data_generator = processor.data_generator( batch_size=args.train_batch_size, is_regression=args.is_regression, phase='train', epoch=args.epoch, dev_count=dev_count, shuffle=args.shuffle) num_train_examples = processor.get_num_examples(phase='train') print("Device count: %d" % dev_count) print("Max num of epoches: %d" % args.epoch) print("Num of train examples: %d" % num_train_examples) print("Num of train steps: %d" % args.train_steps) print("Num of warmup steps: %d" % args.warmup_steps) with fluid.program_guard(train_program, startup_prog): with fluid.unique_name.guard(): train_data_loader, loss, logits, num_seqs, label_ids = create_model( args, xlnet_config=xlnet_config, n_class=num_labels) scheduled_lr = optimization( loss=loss, warmup_steps=args.warmup_steps, num_train_steps=args.train_steps, learning_rate=args.learning_rate, train_program=train_program, startup_prog=startup_prog, weight_decay=args.weight_decay, lr_layer_decay_rate=args.lr_layer_decay_rate, scheduler=args.lr_scheduler) if args.do_eval: dev_prog = fluid.Program() with fluid.program_guard(dev_prog, startup_prog): with fluid.unique_name.guard(): dev_data_loader, loss, logits, num_seqs, label_ids = create_model( args, xlnet_config=xlnet_config, n_class=num_labels) dev_prog = dev_prog.clone(for_test=True) dev_data_loader.set_batch_generator( processor.data_generator( batch_size=args.eval_batch_size, is_regression=args.is_regression, phase=args.eval_split, epoch=1, dev_count=1, shuffle=False), place) if args.do_predict: predict_prog = fluid.Program() with fluid.program_guard(predict_prog, startup_prog): with fluid.unique_name.guard(): predict_data_loader, loss, logits, num_seqs, label_ids = create_model( args, xlnet_config=xlnet_config, n_class=num_labels) predict_prog = predict_prog.clone(for_test=True) predict_data_loader.set_batch_generator( processor.data_generator( batch_size=args.predict_batch_size, is_regression=args.is_regression, phase=args.eval_split, epoch=1, dev_count=1, shuffle=False), place) exe.run(startup_prog) if args.do_train: if args.init_checkpoint and args.init_pretraining_params: print( "WARNING: args 'init_checkpoint' and 'init_pretraining_params' " "both are set! Only arg 'init_checkpoint' is made valid.") if args.init_checkpoint: init_checkpoint( exe, args.init_checkpoint, main_program=startup_prog) elif args.init_pretraining_params: init_pretraining_params( exe, args.init_pretraining_params, main_program=startup_prog) elif args.do_eval or args.do_predict: if not args.init_checkpoint: raise ValueError("args 'init_checkpoint' should be set if" "only doing validation or testing!") init_checkpoint( exe, args.init_checkpoint, main_program=startup_prog) if args.do_train: exec_strategy = fluid.ExecutionStrategy() exec_strategy.use_experimental_executor = args.use_fast_executor exec_strategy.num_threads = dev_count build_strategy = fluid.BuildStrategy() if args.use_cuda and num_trainers > 1: assert shuffle_seed is not None dist_utils.prepare_for_multi_process(exe, build_strategy, train_program) train_data_generator = fluid.contrib.reader.distributed_batch_reader( train_data_generator) train_compiled_program = fluid.CompiledProgram(train_program).with_data_parallel( loss_name=loss.name, build_strategy=build_strategy) train_data_loader.set_batch_generator(train_data_generator, place) if args.do_train: train_data_loader.start() steps = 0 total_cost, total_num_seqs, total_time = [], [], 0.0 throughput = [] ce_info = [] while steps < args.train_steps: try: time_begin = time.time() steps += 1 if steps % args.skip_steps == 0: fetch_list = [loss.name, scheduled_lr.name, num_seqs.name] else: fetch_list = [] outputs = exe.run(train_compiled_program, fetch_list=fetch_list) time_end = time.time() used_time = time_end - time_begin total_time += used_time if steps % args.skip_steps == 0: np_loss, np_lr, np_num_seqs = outputs total_cost.extend(np_loss * np_num_seqs) total_num_seqs.extend(np_num_seqs) if args.verbose: verbose = "train data_loader queue size: %d, " % train_data_loader.queue.size( ) verbose += "learning rate: %f" % np_lr[0] print(verbose) current_example, current_epoch = processor.get_train_progress( ) log_record = "epoch: {}, progress: {}/{}, step: {}, ave loss: {}".format( current_epoch, current_example, num_train_examples, steps, np.sum(total_cost) / np.sum(total_num_seqs)) ce_info.append([np.sum(total_cost) / np.sum(total_num_seqs), used_time]) if steps > 0 : throughput.append( args.skip_steps / total_time) log_record = log_record + ", speed: %f steps/s" % (args.skip_steps / total_time) print(log_record) else: print(log_record) total_cost, total_num_seqs, total_time = [], [], 0.0 if steps % args.save_steps == 0: save_path = os.path.join(args.checkpoints, "step_" + str(steps)) fluid.save(model_path=save_path, program=train_program) if steps % args.validation_steps == 0: print("Average throughtput: %s" % (np.average(throughput))) throughput = [] # evaluate dev set if args.do_eval: evaluate(exe, dev_prog, dev_data_loader, [loss.name, num_seqs.name, logits.name, label_ids.name], args.eval_split, processor.get_num_examples(phase=args.eval_split)) except fluid.core.EOFException: save_path = os.path.join(args.checkpoints, "step_" + str(steps)) fluid.save(model_path=save_path, program=train_program) train_data_loader.reset() break if args.enable_ce: card_num = get_cards() ce_cost = 0 ce_time = 0 try: ce_cost = ce_info[-2][0] ce_time = ce_info[-2][1] except: print("ce info error") print("kpis\ttrain_duration_%s_card%s\t%s" % (args.task_name.replace("-", "_"), card_num, ce_time)) print("kpis\ttrain_cost_%s_card%s\t%f" % (args.task_name.replace("-", "_"), card_num, ce_cost)) # final eval on dev set if args.do_eval: evaluate(exe, dev_prog, dev_data_loader, [loss.name, num_seqs.name, logits.name, label_ids], args.eval_split, processor.get_num_examples(phase=args.eval_split)) # final eval on test set if args.do_predict: predict(exe, predict_prog, predict_data_loader, task_name, label_list, [logits.name]) if __name__ == '__main__': print_arguments(args) check_cuda(args.use_cuda) main(args)