run_xlnet_classifier.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.
"""BERT finetuning runner."""

from __future__ import absolute_import, division, print_function

import argparse
import logging
import os
import sys
import random
from tqdm import tqdm, trange

import numpy as np

import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)
from torch.utils.data.distributed import DistributedSampler
from torch.nn import CrossEntropyLoss, MSELoss

from tensorboardX import SummaryWriter

from pytorch_transformers import WEIGHTS_NAME, CONFIG_NAME
from pytorch_transformers.modeling_xlnet import XLNetForSequenceClassification
from pytorch_transformers.tokenization_xlnet import XLNetTokenizer
from pytorch_transformers.optimization import BertAdam, WarmupLinearSchedule

from utils_glue import processors, output_modes, convert_examples_to_features, compute_metrics

if sys.version_info[0] == 2:
    import cPickle as pickle
else:
    import pickle


logger = logging.getLogger(__name__)


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

    ## Required parameters
    parser.add_argument("--data_dir", default=None, type=str, required=True,
                        help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
    parser.add_argument("--task_name", default=None, type=str, required=True,
                        help="The name of the task to train.")
    parser.add_argument("--output_dir", default=None, type=str, required=True,
                        help="The output directory where the model predictions and checkpoints will be written.")
    # training
    parser.add_argument("--do_train", action='store_true',
                        help="Whether to run training.")
    parser.add_argument("--learning_rate", default=5e-5, type=float,
                        help="The initial learning rate for Adam.")
    parser.add_argument("--num_train_epochs", default=3.0, type=float,
                        help="Total number of training epochs to perform.")
    parser.add_argument("--max_steps", default=-1, type=int,
                        help="If > 0 limit the number of training steps to perform, you should choose only one of num_train_epochs and max_steps.")
    parser.add_argument("--warmup_proportion", default=0.1, type=float,
                        help="Proportion of training to perform linear learning rate warmup for. "
                             "E.g., 0.1 = 10%% of training.")
    parser.add_argument("--clip_gradients", default=1.0, type=float,
                        help="Clip gradient norms.")
    parser.add_argument("--train_batch_size", default=32, type=int,
                        help="Total batch size for training.")
    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('--fp16', action='store_true',
                        help="Whether to use 16-bit float precision instead of 32-bit")
    parser.add_argument('--loss_scale', type=float, default=0,
                        help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
                             "0 (default value): dynamic loss scaling.\n"
                             "Positive power of 2: static loss scaling value.\n")
    parser.add_argument("--log_every", default=10, type=int,
                        help="Log metrics every X training steps.")
    # evaluation
    parser.add_argument("--do_eval", action='store_true',
                        help="Whether to run eval on the dev set.")
    parser.add_argument("--eval_batch_size", default=8, type=int,
                        help="Total batch size for eval.")
    # Model
    parser.add_argument("--xlnet_model", default="xlnet-large-cased", type=str,
                        help="XLNet pre-trained model: currently only xlnet-large-cased.")
    parser.add_argument("--do_lower_case", action='store_true',
                        help="Set this flag if you are using an uncased model.")
    parser.add_argument("--cache_dir", default="", type=str,
                        help="Where do you want to store the pre-trained models downloaded from s3")
    # task specific
    parser.add_argument("--max_seq_length", default=128, type=int,
                        help="The maximum total input sequence length after WordPiece tokenization. \n"
                             "Sequences longer than this will be truncated, and sequences shorter \n"
                             "than this will be padded.")
    parser.add_argument('--overwrite_output_dir', action='store_true',
                        help="Overwrite the content of the output directory")
    # Misc
    parser.add_argument("--no_cuda", action='store_true',
                        help="Whether not to use CUDA when available")
    parser.add_argument("--local_rank", type=int, default=-1,
                        help="local_rank for distributed training on gpus")
    parser.add_argument('--seed', type=int, default=42,
                        help="random seed for initialization")
    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()

    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()

    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")
        n_gpu = torch.cuda.device_count()
    else:
        torch.cuda.set_device(args.local_rank)
        device = torch.device("cuda", args.local_rank)
        n_gpu = 1
        # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
        torch.distributed.init_process_group(backend='nccl')
    args.device = device

    logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
                        datefmt = '%m/%d/%Y %H:%M:%S',
                        level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)

    logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
        device, n_gpu, bool(args.local_rank != -1), args.fp16))

    if args.gradient_accumulation_steps < 1:
        raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
                            args.gradient_accumulation_steps))

    args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps

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

    if not args.do_train and not args.do_eval:
        raise ValueError("At least one of `do_train` or `do_eval` must be True.")

    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 not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
        os.makedirs(args.output_dir)

    task_name = args.task_name.lower()

    if task_name not in processors:
        raise ValueError("Task not found: %s" % (task_name))

    processor = processors[task_name]()
    output_mode = output_modes[task_name]

    label_list = processor.get_labels()
    num_labels = len(label_list)

    if args.local_rank not in [-1, 0]:
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
    tokenizer = XLNetTokenizer.from_pretrained(args.xlnet_model, do_lower_case=args.do_lower_case)
    model = XLNetForSequenceClassification.from_pretrained(args.xlnet_model, num_labels=num_labels)
    if args.local_rank == 0:
        torch.distributed.barrier()

    if args.fp16:
        model.half()
    model.to(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 n_gpu > 1:
        model = torch.nn.DataParallel(model)

    global_step = 0
    curr_tr_loss, curr_steps = 0., 1

    if args.do_train:
        if args.local_rank in [-1, 0]:
            tb_writer = SummaryWriter()

        # Prepare data loader
        train_examples = processor.get_train_examples(args.data_dir)
        cached_train_features_file = os.path.join(args.data_dir, 'train_{0}_{1}_{2}'.format(
            list(filter(None, args.xlnet_model.split('/'))).pop(),
                        str(args.max_seq_length),
                        str(task_name)))
        if os.path.exists(cached_train_features_file):
            logger.info("Loading train features for cache file %s", cached_train_features_file)
            with open(cached_train_features_file, "rb") as reader:
                train_features = pickle.load(reader)
        else:
            logger.info("No cache file at %s, preparing train features", cached_train_features_file)
            train_features = convert_examples_to_features(
                train_examples, label_list, args.max_seq_length, tokenizer, output_mode,
                cls_token_at_end=True, cls_token=tokenizer.CLS_TOKEN,
                sep_token=tokenizer.SEP_TOKEN, cls_token_segment_id=2,
                pad_on_left=True, pad_token_segment_id=4)
            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)

        if output_mode == "classification":
            all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long)
        elif output_mode == "regression":
            all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.float)

        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
        if args.local_rank == -1:
            train_sampler = SequentialSampler(train_data)  # RandomSampler(train_data)
        else:
            train_sampler = DistributedSampler(train_data)
        train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)

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

        # Prepare optimizer

        optimizer_grouped_parameters = model.parameters()
        # param_optimizer = list(model.named_parameters())
        # 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)

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

        model.train()
        for _ in trange(int(args.num_train_epochs) if args.max_steps <= 0 else int('Inf'),
                        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(device) for t in batch)
                input_ids, input_mask, segment_ids, label_ids = batch

                # define a new function to compute loss values for both output_modes
                loss, _ = model(input_ids, token_type_ids=segment_ids, attention_mask=input_mask, labels=label_ids)

                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()

                gnorm = torch.nn.utils.clip_grad_norm_(model.parameters(), args.clip_gradients)

                curr_tr_loss += loss.item()
                curr_steps += 1
                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 that 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] and (args.log_every <= 0 or (global_step + 1) % args.log_every == 0):
                        learning_rate = optimizer.get_lr()[0] if not args.fp16 else lr_this_step
                        logger.info("[{}] | gnorm {:.2f} lr {:8.6f} | loss {:.2f}".format(
                            global_step, gnorm, learning_rate, curr_tr_loss / curr_steps))
                        tb_writer.add_scalar('lr', learning_rate, global_step)
                        tb_writer.add_scalar('loss', curr_tr_loss / curr_steps, global_step)
                        curr_tr_loss, curr_steps = 0., 1

                    if args.max_steps > 0 and global_step > args.max_steps:
                        break

            if args.max_steps > 0 and global_step > args.max_steps:
                break

    ### Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
    ### Example:
    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 = XLNetForSequenceClassification.from_pretrained(args.output_dir, num_labels=num_labels)
        tokenizer = XLNetTokenizer.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 = XLNetForSequenceClassification.from_pretrained(args.xlnet_model, num_labels=num_labels)

    model.to(device)

    ### Evaluation
    if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
        eval_examples = processor.get_dev_examples(args.data_dir)
        cached_eval_features_file = os.path.join(args.data_dir, 'dev_{0}_{1}_{2}'.format(
            list(filter(None, args.xlnet_model.split('/'))).pop(),
                        str(args.max_seq_length),
                        str(task_name)))
        if os.path.exists(cached_eval_features_file):
            logger.info("Loading eval features for cache file %s", cached_eval_features_file)
            with open(cached_eval_features_file, "rb") as reader:
                eval_features = pickle.load(reader)
        else:
            logger.info("No cache file at %s, preparing eval features", cached_eval_features_file)
            eval_features = convert_examples_to_features(
                eval_examples, label_list, args.max_seq_length, tokenizer, output_mode,
                cls_token_at_end=True, cls_token=tokenizer.CLS_TOKEN,
                sep_token=tokenizer.SEP_TOKEN, cls_token_segment_id=2,
                pad_on_left=True, pad_token_segment_id=4)
            if args.local_rank == -1 or torch.distributed.get_rank() == 0:
                logger.info("  Saving eval features into cached file %s", cached_eval_features_file)
                with open(cached_eval_features_file, "wb") as writer:
                    pickle.dump(eval_features, writer)


        logger.info("***** Running evaluation *****")
        logger.info("  Num examples = %d", len(eval_examples))
        logger.info("  Batch size = %d", args.eval_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)

        if output_mode == "classification":
            all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)
        elif output_mode == "regression":
            all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.float)

        eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
        # Run prediction for full data
        if args.local_rank == -1:
            eval_sampler = SequentialSampler(eval_data)
        else:
            eval_sampler = DistributedSampler(eval_data)  # Note that this sampler samples randomly
        eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)

        model.eval()
        eval_loss = 0
        nb_eval_steps = 0
        preds = []
        out_label_ids = None

        for input_ids, input_mask, segment_ids, label_ids in tqdm(eval_dataloader, desc="Evaluating"):
            input_ids = input_ids.to(device)
            input_mask = input_mask.to(device)
            segment_ids = segment_ids.to(device)
            label_ids = label_ids.to(device)

            with torch.no_grad():
                logits, _ = model(input_ids, token_type_ids=segment_ids, attention_mask=input_mask)

            # create eval loss and other metric required by the task
            if output_mode == "classification":
                loss_fct = CrossEntropyLoss()
                tmp_eval_loss = loss_fct(logits.view(-1, num_labels), label_ids.view(-1))
            elif output_mode == "regression":
                loss_fct = MSELoss()
                tmp_eval_loss = loss_fct(logits.view(-1), label_ids.view(-1))

            eval_loss += tmp_eval_loss.mean().item()
            nb_eval_steps += 1
            if len(preds) == 0:
                preds.append(logits.detach().cpu().numpy())
                out_label_ids = label_ids.detach().cpu().numpy()
            else:
                preds[0] = np.append(
                    preds[0], logits.detach().cpu().numpy(), axis=0)
                out_label_ids = np.append(
                    out_label_ids, label_ids.detach().cpu().numpy(), axis=0)

        eval_loss = eval_loss / nb_eval_steps
        preds = preds[0]
        if output_mode == "classification":
            preds = np.argmax(preds, axis=1)
        elif output_mode == "regression":
            preds = np.squeeze(preds)
        result = compute_metrics(task_name, preds, out_label_ids)

        loss = curr_tr_loss/curr_steps if args.do_train else None

        result['eval_loss'] = eval_loss
        result['global_step'] = global_step
        result['loss'] = loss

        output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
        with open(output_eval_file, "w") as writer:
            logger.info("***** Eval results *****")
            for key in sorted(result.keys()):
                logger.info("  %s = %s", key, str(result[key]))
                writer.write("%s = %s\n" % (key, str(result[key])))

        # hack for MNLI-MM
        if task_name == "mnli":
            task_name = "mnli-mm"
            processor = processors[task_name]()

            if os.path.exists(args.output_dir + '-MM') and os.listdir(args.output_dir + '-MM') and args.do_train:
                raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
            if not os.path.exists(args.output_dir + '-MM'):
                os.makedirs(args.output_dir + '-MM')

            eval_examples = processor.get_dev_examples(args.data_dir)
            eval_features = convert_examples_to_features(
                eval_examples, label_list, args.max_seq_length, tokenizer, output_mode)
            logger.info("***** Running evaluation *****")
            logger.info("  Num examples = %d", len(eval_examples))
            logger.info("  Batch size = %d", args.eval_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_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)

            eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
            # Run prediction for full data
            eval_sampler = SequentialSampler(eval_data)
            eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)

            model.eval()
            eval_loss = 0
            nb_eval_steps = 0
            preds = []
            out_label_ids = None

            for input_ids, input_mask, segment_ids, label_ids in tqdm(eval_dataloader, desc="Evaluating"):
                input_ids = input_ids.to(device)
                input_mask = input_mask.to(device)
                segment_ids = segment_ids.to(device)
                label_ids = label_ids.to(device)

                with torch.no_grad():
                    logits, _ = model(input_ids, token_type_ids=segment_ids, attention_mask=input_mask, labels=None)

                loss_fct = CrossEntropyLoss()
                tmp_eval_loss = loss_fct(logits.view(-1, num_labels), label_ids.view(-1))

                eval_loss += tmp_eval_loss.mean().item()
                nb_eval_steps += 1
                if len(preds) == 0:
                    preds.append(logits.detach().cpu().numpy())
                    out_label_ids = label_ids.detach().cpu().numpy()
                else:
                    preds[0] = np.append(
                        preds[0], logits.detach().cpu().numpy(), axis=0)
                    out_label_ids = np.append(
                        out_label_ids, label_ids.detach().cpu().numpy(), axis=0)

            eval_loss = eval_loss / nb_eval_steps
            preds = preds[0]
            preds = np.argmax(preds, axis=1)
            result = compute_metrics(task_name, preds, out_label_ids)

            loss = curr_tr_loss/curr_steps if args.do_train else None

            result['eval_loss'] = eval_loss
            result['global_step'] = global_step
            result['loss'] = loss

            output_eval_file = os.path.join(args.output_dir + '-MM', "eval_results.txt")
            with open(output_eval_file, "w") as writer:
                logger.info("***** Eval results *****")
                for key in sorted(result.keys()):
                    logger.info("  %s = %s", key, str(result[key]))
                    writer.write("%s = %s\n" % (key, str(result[key])))

if __name__ == "__main__":
    main()