transformers/examples/run_squad.py

# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  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.
"""Run BERT on SQuAD."""

from __future__ import absolute_import, division, print_function

import argparse
import logging
import os
import random
import sys
from io import open

import numpy as np
import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)
from torch.utils.data.distributed import DistributedSampler
from tqdm import tqdm, trange

from tensorboardX import SummaryWriter

from pytorch_transformers import (BertForQuestionAnswering, XLNetForQuestionAnswering,
                                  XLMForQuestionAnswering, BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
                                  XLNET_PRETRAINED_MODEL_ARCHIVE_MAP, XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
from pytorch_transformers import (BertTokenizer, XLNetTokenizer,
                                  XLMTokenizer)

from utils_squad import read_squad_examples, convert_examples_to_features, RawResult, write_predictions

logger = logging.getLogger(__name__)

ALL_MODELS = sum((tuple(m.keys()) for m in (BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
                                            XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
                                            XLM_PRETRAINED_MODEL_ARCHIVE_MAP)), ())

MODEL_CLASSES = {
    'bert': BertForQuestionAnswering,
    'xlnet': XLNetForQuestionAnswering,
    'xlm': XLMForQuestionAnswering,
}

TOKENIZER_CLASSES = {
    'bert': BertTokenizer,
    'xlnet': XLNetTokenizer,
    'xlm': XLMTokenizer,
}

def train(args, train_dataset, model):
    """ Train the model """
    if args.local_rank in [-1, 0]:
        tb_writer = SummaryWriter()

    args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)

    if args.max_steps > 0:
        num_train_optimization_steps = args.max_steps
        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
    else:
        num_train_optimization_steps = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs

    # Prepare optimizer
    no_decay = ['bias', 'LayerNorm.weight']
    optimizer_grouped_parameters = [
        {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]
    optimizer = BertAdam(optimizer_grouped_parameters, lr=args.learning_rate,
                         t_total=num_train_optimization_steps, warmup=args.warmup_proportion)
    if args.fp16:
        try:
            from apex import amp
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)

    # Train!
    logger.info("***** Running training *****")
    logger.info("  Num examples = %d", len(train_dataset))
    logger.info("  Num Epochs = %d", args.num_train_epochs)
    logger.info("  Batch size = %d", args.train_batch_size)
    logger.info("  Total batch size (distributed) = %d", args.train_batch_size * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
    logger.info("  Total optimization steps = %d", num_train_optimization_steps)

    global_step = 0
    tr_loss, logging_loss = 0.0, 0.0
    model.train()
    optimizer.zero_grad()
    for _ in trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]):
        for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])):
            batch = tuple(t.to(args.device) for t in batch)
            inputs = {'input_ids':      batch[0],
                      'attention_mask': batch[1],
                      'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None,  # XLM don't use segment_ids
                      'labels':         batch[3]}
            ouputs = model(**inputs)
            loss = ouputs[0]


def evalutate(args, dataset, model):
    """ Evaluate the model """



def load_and_cache_examples(args, tokenizer, training=True):
    """ Load data features from cache or dataset file. """
    cached_features_file = os.path.join(args.data_dir, 'cached_{}_{}_{}_{}'.format(
        'dev' if evaluate else 'train',
        list(filter(None, args.model_name.split('/'))).pop(),
        str(args.max_seq_length),
        str(task)))
    if os.path.exists(cached_features_file):
        logger.info("Loading features from cached file %s", cached_features_file)
        features = torch.load(cached_features_file)
    else:
        logger.info("Creating features from dataset file at %s", args.data_dir)
        label_list = processor.get_labels()
        examples = read_squad_examples(input_file=args.train_file if training else args.predict_file,
                        is_training=training,
                        version_2_with_negative=args.version_2_with_negative)
        features = convert_examples_to_features(
            examples=examples,
            tokenizer=tokenizer,
            max_seq_length=args.max_seq_length,
            doc_stride=args.doc_stride,
            max_query_length=args.max_query_length,
            is_training=training)
        if args.local_rank in [-1, 0]:
            logger.info("Num orig examples = %d", len(examples))
            logger.info("Num split examples = %d", len(features))
            logger.info("Saving features into cached file %s", cached_features_file)
            torch.save(features, cached_features_file)

    # Convert to Tensors and build dataset
    all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
    all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
    all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
    if training:
        all_start_positions = torch.tensor([f.start_position for f in train_features], dtype=torch.long)
        all_end_positions = torch.tensor([f.end_position for f in train_features], dtype=torch.long)
        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_start_positions, all_end_positions)
    else:
        all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index)

    return dataset


def main():
    parser = argparse.ArgumentParser()

    ## Required parameters
    parser.add_argument("--train_file", default=None, type=str, required=True,
                        help="SQuAD json for training. E.g., train-v1.1.json")
    parser.add_argument("--predict_file", default=None, type=str, required=True,
                        help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json")
    parser.add_argument("--model_name", default=None, type=str, required=True,
                        help="Bert/XLNet/XLM pre-trained model selected in the list: " + ", ".join(ALL_MODELS))
    parser.add_argument("--output_dir", default=None, type=str, required=True,
                        help="The output directory where the model checkpoints and predictions will be written.")

    ## Other parameters
    parser.add_argument('--version_2_with_negative', action='store_true',
                        help='If true, the SQuAD examples contain some that do not have an answer.')
    parser.add_argument('--null_score_diff_threshold', type=float, default=0.0,
                        help="If null_score - best_non_null is greater than the threshold predict null.")
    parser.add_argument('--overwrite_output_dir', action='store_true',
                        help="Overwrite the content of the output directory")

    parser.add_argument("--max_seq_length", default=384, type=int,
                        help="The maximum total input sequence length after WordPiece tokenization. Sequences "
                             "longer than this will be truncated, and sequences shorter than this will be padded.")
    parser.add_argument("--doc_stride", default=128, type=int,
                        help="When splitting up a long document into chunks, how much stride to take between chunks.")
    parser.add_argument("--max_query_length", default=64, type=int,
                        help="The maximum number of tokens for the question. Questions longer than this will "
                             "be truncated to this length.")
    parser.add_argument("--do_train", action='store_true',
                        help="Whether to run training.")
    parser.add_argument("--do_predict", action='store_true',
                        help="Whether to run eval on the dev set.")
    parser.add_argument("--do_lower_case", action='store_true',
                        help="Whether to lower case the input text. True for uncased models, False for cased models.")

    parser.add_argument("--train_batch_size", default=32, type=int,
                        help="Total batch size for training.")
    parser.add_argument("--predict_batch_size", default=8, type=int,
                        help="Total batch size for predictions.")
    parser.add_argument("--learning_rate", default=5e-5, type=float,
                        help="The initial learning rate for Adam.")
    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
                        help="Number of updates steps to accumulate before performing a backward/update pass.")
    parser.add_argument("--num_train_epochs", default=3.0, type=float,
                        help="Total number of training epochs to perform.")
    parser.add_argument("--warmup_proportion", default=0.1, type=float,
                        help="Proportion of training with linear learning rate warmup (0.1 = 10%% of training).")
    parser.add_argument("--n_best_size", default=20, type=int,
                        help="The total number of n-best predictions to generate in the nbest_predictions.json output file.")
    parser.add_argument("--max_answer_length", default=30, type=int,
                        help="The maximum length of an answer that can be generated. This is needed because the start "
                             "and end predictions are not conditioned on one another.")
    parser.add_argument("--verbose_logging", action='store_true',
                        help="If true, all of the warnings related to data processing will be printed. "
                             "A number of warnings are expected for a normal SQuAD evaluation.")

    parser.add_argument("--no_cuda", action='store_true',
                        help="Whether not to use CUDA when available")
    parser.add_argument('--seed', type=int, default=42,
                        help="random seed for initialization")
    parser.add_argument("--local_rank", type=int, default=-1,
                        help="local_rank for distributed training on gpus")
    parser.add_argument('--fp16', action='store_true',
                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
    parser.add_argument('--fp16_opt_level', type=str, default='O1',
                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
                             "See details at https://nvidia.github.io/apex/amp.html")
    parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
    parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
    args = parser.parse_args()
    print(args)

    if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
        raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))

    if args.server_ip and args.server_port:
        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
        import ptvsd
        print("Waiting for debugger attach")
        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
        ptvsd.wait_for_attach()

    # Setup CUDA, GPU & distributed training
    if args.local_rank == -1 or args.no_cuda:
        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
        args.n_gpu = torch.cuda.device_count()
    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
        torch.cuda.set_device(args.local_rank)
        device = torch.device("cuda", args.local_rank)
        torch.distributed.init_process_group(backend='nccl')
        args.n_gpu = 1
    args.device = device

    # Setup logging
    logging.basicConfig(level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
                args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)

    # Setup seeds
    random.seed(args.seed)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    if args.n_gpu > 0:
        torch.cuda.manual_seed_all(args.seed)

    # Load pretrained model and tokenizer
    if args.local_rank not in [-1, 0]:
        torch.distributed.barrier()  # Make sure only 1st process in distributed training download model & vocab

    args.model_type = args.model_name.lower().split('-')[0]
    tokenizer_class = TOKENIZER_CLASSES[args.model_type]
    model_class = MODEL_CLASSES[args.model_type]
    tokenizer = tokenizer_class.from_pretrained(args.model_name, do_lower_case=args.do_lower_case)
    model = model_class.from_pretrained(args.model_name, num_labels=num_labels)

    if args.local_rank == 0:
        torch.distributed.barrier()

    # Distributed and parrallel training
    model.to(args.device)
    if args.local_rank != -1:
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
                                                          output_device=args.local_rank,
                                                          find_unused_parameters=True)
    elif args.n_gpu > 1:
        model = torch.nn.DataParallel(model)

    # Training
    if args.do_train:
        if args.local_rank in [-1, 0]:
            tb_writer = SummaryWriter()
        # Prepare data loader
        train_examples = read_squad_examples(
            input_file=args.train_file, is_training=True, version_2_with_negative=args.version_2_with_negative)
        cached_train_features_file = args.train_file+'_{0}_{1}_{2}_{3}'.format(
            list(filter(None, args.bert_model.split('/'))).pop(), str(args.max_seq_length), str(args.doc_stride), str(args.max_query_length))
        try:
            with open(cached_train_features_file, "rb") as reader:
                train_features = pickle.load(reader)
        except:
            train_features = convert_examples_to_features(
                examples=train_examples,
                tokenizer=tokenizer,
                max_seq_length=args.max_seq_length,
                doc_stride=args.doc_stride,
                max_query_length=args.max_query_length,
                is_training=True)
            if args.local_rank == -1 or torch.distributed.get_rank() == 0:
                logger.info("  Saving train features into cached file %s", cached_train_features_file)
                with open(cached_train_features_file, "wb") as writer:
                    pickle.dump(train_features, writer)

        all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)
        all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)
        all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)
        all_start_positions = torch.tensor([f.start_position for f in train_features], dtype=torch.long)
        all_end_positions = torch.tensor([f.end_position for f in train_features], dtype=torch.long)
        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
                                   all_start_positions, all_end_positions)
        if args.local_rank == -1:
            train_sampler = RandomSampler(train_data)
        else:
            train_sampler = DistributedSampler(train_data)

        train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
        num_train_optimization_steps = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
        # if args.local_rank != -1:
        #     num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()

        # Prepare optimizer
        param_optimizer = list(model.named_parameters())

        # hack to remove pooler, which is not used
        # thus it produce None grad that break apex
        param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]

        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
        optimizer_grouped_parameters = [
            {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
            {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
            ]

        if args.fp16:
            try:
                from apex.optimizers import FP16_Optimizer
                from apex.optimizers import FusedAdam
            except ImportError:
                raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")

            optimizer = FusedAdam(optimizer_grouped_parameters,
                                  lr=args.learning_rate,
                                  bias_correction=False,
                                  max_grad_norm=1.0)
            if args.loss_scale == 0:
                optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
            else:
                optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
            warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
                                                 t_total=num_train_optimization_steps)
        else:
            optimizer = BertAdam(optimizer_grouped_parameters,
                                 lr=args.learning_rate,
                                 warmup=args.warmup_proportion,
                                 t_total=num_train_optimization_steps)

        global_step = 0

        logger.info("***** Running training *****")
        logger.info("  Num orig examples = %d", len(train_examples))
        logger.info("  Num split examples = %d", len(train_features))
        logger.info("  Batch size = %d", args.train_batch_size)
        logger.info("  Num steps = %d", num_train_optimization_steps)

        model.train()
        for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
            for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])):
                if n_gpu == 1:
                    batch = tuple(t.to(device) for t in batch) # multi-gpu does scattering it-self
                input_ids, input_mask, segment_ids, start_positions, end_positions = batch
                loss = model(input_ids, segment_ids, input_mask, start_positions, end_positions)
                if n_gpu > 1:
                    loss = loss.mean() # mean() to average on multi-gpu.
                if args.gradient_accumulation_steps > 1:
                    loss = loss / args.gradient_accumulation_steps

                if args.fp16:
                    optimizer.backward(loss)
                else:
                    loss.backward()
                if (step + 1) % args.gradient_accumulation_steps == 0:
                    if args.fp16:
                        # modify learning rate with special warm up BERT uses
                        # if args.fp16 is False, BertAdam is used and handles this automatically
                        lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step, args.warmup_proportion)
                        for param_group in optimizer.param_groups:
                            param_group['lr'] = lr_this_step
                    optimizer.step()
                    optimizer.zero_grad()
                    global_step += 1
                    if args.local_rank in [-1, 0]:
                        if not args.fp16:
                            tb_writer.add_scalar('lr', optimizer.get_lr()[0], global_step)
                        tb_writer.add_scalar('loss', loss.item(), global_step)

    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
        # Save a trained model, configuration and tokenizer
        model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self

        # If we save using the predefined names, we can load using `from_pretrained`
        output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
        output_config_file = os.path.join(args.output_dir, CONFIG_NAME)

        torch.save(model_to_save.state_dict(), output_model_file)
        model_to_save.config.to_json_file(output_config_file)
        tokenizer.save_vocabulary(args.output_dir)

        # Load a trained model and vocabulary that you have fine-tuned
        model = BertForQuestionAnswering.from_pretrained(args.output_dir)
        tokenizer = BertTokenizer.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)

        # Good practice: save your training arguments together with the trained model
        output_args_file = os.path.join(args.output_dir, 'training_args.bin')
        torch.save(args, output_args_file)
    else:
        model = BertForQuestionAnswering.from_pretrained(args.bert_model)

    model.to(device)

    if args.do_predict and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
        eval_examples = read_squad_examples(
            input_file=args.predict_file, is_training=False, version_2_with_negative=args.version_2_with_negative)
        eval_features = convert_examples_to_features(
            examples=eval_examples,
            tokenizer=tokenizer,
            max_seq_length=args.max_seq_length,
            doc_stride=args.doc_stride,
            max_query_length=args.max_query_length,
            is_training=False)

        logger.info("***** Running predictions *****")
        logger.info("  Num orig examples = %d", len(eval_examples))
        logger.info("  Num split examples = %d", len(eval_features))
        logger.info("  Batch size = %d", args.predict_batch_size)

        all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
        all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
        all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
        all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
        eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index)
        # Run prediction for full data
        eval_sampler = SequentialSampler(eval_data)
        eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.predict_batch_size)

        model.eval()
        all_results = []
        logger.info("Start evaluating")
        for input_ids, input_mask, segment_ids, example_indices in tqdm(eval_dataloader, desc="Evaluating", disable=args.local_rank not in [-1, 0]):
            if len(all_results) % 1000 == 0:
                logger.info("Processing example: %d" % (len(all_results)))
            input_ids = input_ids.to(device)
            input_mask = input_mask.to(device)
            segment_ids = segment_ids.to(device)
            with torch.no_grad():
                batch_start_logits, batch_end_logits = model(input_ids, segment_ids, input_mask)
            for i, example_index in enumerate(example_indices):
                start_logits = batch_start_logits[i].detach().cpu().tolist()
                end_logits = batch_end_logits[i].detach().cpu().tolist()
                eval_feature = eval_features[example_index.item()]
                unique_id = int(eval_feature.unique_id)
                all_results.append(RawResult(unique_id=unique_id,
                                             start_logits=start_logits,
                                             end_logits=end_logits))
        output_prediction_file = os.path.join(args.output_dir, "predictions.json")
        output_nbest_file = os.path.join(args.output_dir, "nbest_predictions.json")
        output_null_log_odds_file = os.path.join(args.output_dir, "null_odds.json")
        write_predictions(eval_examples, eval_features, all_results,
                          args.n_best_size, args.max_answer_length,
                          args.do_lower_case, output_prediction_file,
                          output_nbest_file, output_null_log_odds_file, args.verbose_logging,
                          args.version_2_with_negative, args.null_score_diff_threshold)


if __name__ == "__main__":
    main()