run_glue.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.
""" Finetuning the library models for sequence classification on GLUE (Bert, XLM, XLNet, RoBERTa)."""

from __future__ import absolute_import, division, print_function

import argparse
import glob
import logging
import os
import random
import json
from pathlib import Path

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

from pytorch_transformers import (WEIGHTS_NAME, BertConfig,
                                  BertForSequenceClassification, BertTokenizer,
                                  RobertaConfig,
                                  RobertaForSequenceClassification,
                                  RobertaTokenizer,
                                  XLMConfig, XLMForSequenceClassification,
                                  XLMTokenizer, XLNetConfig,
                                  XLNetForSequenceClassification,
                                  XLNetTokenizer)

from pytorch_transformers import AdamW, WarmupLinearSchedule, ConstantLRSchedule
#from multitask import BertForMultitaskClassification

from utils_glue import (compute_metrics, convert_examples_to_features,
                        output_modes, processors)

from utils import make_logger_sufferable
# Less logging pollution
logging.getLogger("pytorch_transformers").setLevel(logging.WARNING)
make_logger_sufferable(logging.getLogger("pytorch_transformers"))
logging.getLogger("utils_glue").setLevel(logging.WARNING)
make_logger_sufferable(logging.getLogger("utils_glue"))

# Logger
logger = logging.getLogger(__name__)
make_logger_sufferable(logger)
logger.setLevel(logging.DEBUG)

class EpochFinished(Exception): pass
class TrainingFinished(Exception): pass

ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig, RobertaConfig)), ())

MODEL_CLASSES = {
    'bert': (BertConfig, BertForSequenceClassification, BertTokenizer),
    #'bert-multitask': (BertConfig, BertForMultitaskClassification, BertTokenizer),
    'xlnet': (XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer),
    'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
    'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer),
}

OPTIMIZERS = {
    'adam': AdamW,
    'adamw': AdamW,
    'sgd': torch.optim.SGD,
}


def set_seed(args):
    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)

class CustomSummaryWriter(SummaryWriter):
    """Log lr and loss values and output as a static summary png"""
    def __init__(self, log_dir, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.log_dir = Path(log_dir)
        self._log = {}
    # tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
    def add_scalar(self, key, value, step):
        super().add_scalar(key, value, step)
        if key not in self._log:
            self._log[key] = []
        self._log[key].append(value)

    def close(self):
        # dump loss log for future reference
        with (self.log_dir / "metric_log.json").open("wt") as f:
            json.dump(self._log, f)
        super().close()

    def dump_plot(self, path):
        import matplotlib.pyplot as plt
        fig, axes = plt.subplots(nrows=len(self._log), ncols=1)
        for ax, (k,v) in zip(axes, self._log.items()):
            ax.plot(v, label=k)
            ax.legend()
        fig.save_fig(path)

class GradientMask:
    def __init__(self, mask):
        self.mask = mask
    @torch.no_grad()
    def __call__(self, grad):
        grad.mul_(self.mask)

def freeze_all_except(model, indices, device):
    embs = model.bert.embeddings.word_embeddings.weight
    mask = torch.zeros(embs.shape[0], 1, dtype=torch.float).to(device)
    for i in indices:
        mask[i, 0] = 1
    hook = GradientMask(mask)
    embs.register_hook(hook)

def erase_grad_except(model, layers):
    for name, p in model.named_parameters():
        if name not in layers:
            name.grad.mul_(0)

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

    if args.early_stopping_interval > 0: assert args.evaluate_during_training
    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
    
    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=4)#args.train_batch_size)

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

    # Prepare optimizer and schedule (linear warmup and decay)
    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': args.weight_decay},
        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]
    OPT = OPTIMIZERS[args.optim]
    optim_kwargs = {}
    if OPT is AdamW: optim_kwargs["eps"] = args.adam_epsilon
    optimizer = OPT(optimizer_grouped_parameters, lr=args.learning_rate, **optim_kwargs)
    if args.constant_schedule:
        scheduler = ConstantLRSchedule(optimizer)
    else:
        scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
    
    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)

    # multi-gpu training (should be after apex fp16 initialization)
    if args.n_gpu > 1:
        model = torch.nn.DataParallel(model)

    # Distributed training (should be after apex fp16 initialization)
    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)
    # Train!
    print("***** Running training *****")
    print("  Num examples = %d", len(train_dataset))
    print("  Num Epochs = %d", args.num_train_epochs)
    print("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
    print("  Total train batch size (w. parallel, distributed & accumulation) = %d",
                   args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
    print("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
    print("  Total optimization steps = %d", t_total)

    global_step = 0
    tr_loss, logging_loss = 0.0, 0.0
    model.zero_grad()
    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
    last_results = {"acc": -1.0, "patience": args.early_stopping_patience}

    layers_to_unfreeze = [x for x in args.layers.split(",") if x]
    if len(layers_to_unfreeze) > 0:
        # sanity check
        found_layers = []
        for name, p in model.named_parameters():
            if name in layers_to_unfreeze:
                found_layers.append(name)
                p.requires_grad = True
            else: p.requires_grad = False
        if sorted(layers_to_unfreeze) != sorted(found_layers):
            raise ValueError(f"Tried to unfreeze layers {sorted(layers_to_unfreeze)} "
                             f"but was only able to unfreeze {sorted(found_layers)}")
    if args.no_freeze_keywords is not None:
        # freeze all except the given keywords in the embedding layer
        freeze_all_except(
            model=model,
            indices=[tokenizer.vocab[x] for x in args.no_freeze_keywords.split(",")],
            device=args.device,
        )

    try:
        for _ in train_iterator:
            epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
            try:
                for step, batch in enumerate(epoch_iterator):
                    model.train()
                    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 and RoBERTa don't use segment_ids
                            'labels':         batch[3]}
                    outputs = model(**inputs)
                    loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)

                    if args.n_gpu > 1:
                        loss = loss.mean() # mean() to average on multi-gpu parallel training
                    if args.gradient_accumulation_steps > 1:
                        loss = loss / args.gradient_accumulation_steps

                    if args.fp16:
                        with amp.scale_loss(loss, optimizer) as scaled_loss:
                            scaled_loss.backward()
                        torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
                    else:
                        loss.backward()
                        torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)

                    tr_loss += loss.item()
                    if (step + 1) % args.gradient_accumulation_steps == 0:
                        scheduler.step()  # Update learning rate schedule
                        optimizer.step()
                        model.zero_grad()
                        global_step += 1

                        if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
                            # Log metrics
                            if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
                                results = evaluate(args, model, tokenizer)
                                # check for early stopping
                                if args.early_stopping_interval > 0:
                                    if results["acc"] < last_results["acc"]:
                                        last_results["patience"] -= 1
                                    else:
                                        last_results["patience"] = args.early_stopping_patience
                                    if last_results["patience"] < 1:
                                        raise TrainingFinished
                                    new_acc = max(results["acc"], last_results["acc"])
                                    last_results["acc"] = new_acc
                                for key, value in results.items():
                                    tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
                            tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
                            tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
                            logging_loss = tr_loss

                        if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
                            # Save model checkpoint
                            output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
                            if not os.path.exists(output_dir):
                                os.makedirs(output_dir)
                            model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
                            model_to_save.save_pretrained(output_dir)
                            torch.save(model_to_save.state_dict(), os.path.join(output_dir, 'model.pt'))
                            torch.save(args, os.path.join(output_dir, 'training_args.bin'))
                            print("Saving model checkpoint to ", output_dir)
                    if args.max_steps > 0 and global_step > args.max_steps:
                        raise TrainingFinished
            except EpochFinished: pass
            finally: epoch_iterator.close()
    except TrainingFinished: pass
    finally: train_iterator.close()

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

    return global_step, tr_loss / global_step


def evaluate(args, model, tokenizer, prefix=""):
    # Loop to handle MNLI double evaluation (matched, mis-matched)
    eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
    eval_outputs_dirs = (args.output_dir, args.output_dir + '-MM') if args.task_name == "mnli" else (args.output_dir,)
    all_data_dirs = [(prefix, args.data_dir)] + [(k, v) for k,v in args.additional_eval.items()]

    results = {}
    for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
        for prefix, data_dir in all_data_dirs:
            eval_dataset = load_and_cache_examples(args, data_dir, eval_task, tokenizer, evaluate=True)

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

            args.eval_batch_size = 4#args.per_gpu_eval_batch_size * max(1, args.n_gpu)
            # Note that DistributedSampler samples randomly
            eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
            eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)

            # Eval!
            print("***** Running evaluation {} *****".format(prefix))
            print("  Num examples = %d", len(eval_dataset))
            print("  Batch size = %d", args.eval_batch_size)
            eval_loss = 0.0
            nb_eval_steps = 0
            preds = None
            out_label_ids = None
            for batch in tqdm(eval_dataloader, desc="Evaluating"):
                model.eval()
                batch = tuple(t.to(args.device) for t in batch)

                with torch.no_grad():
                    inputs = {'input_ids':      batch[0],
                            'attention_mask': batch[1],
                            'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None,  # XLM and RoBERTa don't use segment_ids
                            'labels':         batch[3]}
                    outputs = model(**inputs)
                    tmp_eval_loss, logits = outputs[:2]

                    eval_loss += tmp_eval_loss.mean().item()
                nb_eval_steps += 1
                if preds is None:
                    preds = logits.detach().cpu().numpy()
                    out_label_ids = inputs['labels'].detach().cpu().numpy()
                else:
                    preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
                    out_label_ids = np.append(out_label_ids, inputs['labels'].detach().cpu().numpy(), axis=0)

            eval_loss = eval_loss / nb_eval_steps
            if args.output_mode == "classification":
                preds = np.argmax(preds, axis=1)
            elif args.output_mode == "regression":
                preds = np.squeeze(preds)
                
            result = compute_metrics(eval_task, preds, out_label_ids, args.roc_file_name)
            results.update({f"{prefix}{k}": v for k, v in result.items()})

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

    return results


def load_and_cache_examples(args, data_dir, task, tokenizer, evaluate=False, no_cache=False):
    if args.local_rank not in [-1, 0] and not evaluate:
        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache

    processor = processors[task]()
    output_mode = output_modes[task]
    
    # Load data features from cache or dataset file
    cached_features_file = os.path.join(data_dir, 'cached_{}_{}_{}'.format(
         'dev' if evaluate else 'train',
         #list(filter(None, args.model_name_or_path.split('/'))).pop(),
         str(args.max_seq_length),
         str(task)))
    if os.path.exists(cached_features_file):
        print("Loading features from cached file ", cached_features_file)
        features = torch.load(cached_features_file)
    
    else:
        print("Creating features from dataset file at ", data_dir)
        label_list = processor.get_labels()
        if task in ['mnli', 'mnli-mm'] and args.model_type in ['roberta']:
            # HACK(label indices are swapped in RoBERTa pretrained model)
            label_list[1], label_list[2] = label_list[2], label_list[1]
            
        examples = processor.get_dev_examples(data_dir) if evaluate else processor.get_train_examples(data_dir)
        
        features = convert_examples_to_features(examples, label_list, args.max_seq_length, tokenizer, output_mode,
            cls_token_at_end=bool(args.model_type in ['xlnet']),            # xlnet has a cls token at the end
            cls_token=tokenizer.cls_token,
            cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
            sep_token=tokenizer.sep_token,
            sep_token_extra=bool(args.model_type in ['roberta']),           # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
            pad_on_left=bool(args.model_type in ['xlnet']),                 # pad on the left for xlnet
            pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
            pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0,
        )
        if args.local_rank in [-1, 0] and not no_cache:
            print("Saving features into cached file %s", cached_features_file)
            torch.save(features, cached_features_file)

    if args.local_rank == 0 and not evaluate:
        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache

    # Convert to Tensors and build dataset
    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
    all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
    all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
    if output_mode == "classification" or output_mode == "multitask":
        all_label_ids = torch.tensor([f.label_id for f in features], dtype=torch.long)
    elif output_mode == "regression":
        all_label_ids = torch.tensor([f.label_id for f in features], dtype=torch.float)

    dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
    return dataset


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("--model_type", default=None, type=str, required=True,
                        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
    parser.add_argument("--task_name", default=None, type=str, required=True,
                        help="The name of the task to train selected in the list: " + ", ".join(processors.keys()))
    parser.add_argument("--output_dir", default=None, type=str, required=True,
                        help="The output directory where the model predictions and checkpoints will be written.")

    ## Other parameters
    parser.add_argument("--config_name", default="", type=str,
                        help="Pretrained config name or path if not the same as model_name")
    parser.add_argument("--tokenizer_name", default="", type=str,
                        help="Pretrained tokenizer name or path if not the same as model_name")
    parser.add_argument("--cache_dir", default="", type=str,
                        help="Where do you want to store the pre-trained models downloaded from s3")
    parser.add_argument("--max_seq_length", default=128, type=int,
                        help="The maximum total input sequence length after tokenization. Sequences longer "
                             "than this will be truncated, sequences shorter will be padded.")
    parser.add_argument("--do_train", action='store_true',
                        help="Whether to run training.")
    parser.add_argument("--do_eval", action='store_true',
                        help="Whether to run eval on the dev set.")
    parser.add_argument("--roc_file_name", default="au_roc.png", type = str,
                       help="roc_auc file name (inlude .png) to save plot.")
    parser.add_argument("--evaluate_during_training", action='store_true',
                        help="Rul evaluation during training at each logging step.")
    parser.add_argument("--do_lower_case", action='store_true',
                        help="Set this flag if you are using an uncased model.")

    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
                        help="Batch size per GPU/CPU for training.")
    parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
                        help="Batch size per GPU/CPU for evaluation.")
    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("--learning_rate", default=5e-5, type=float,
                        help="The initial learning rate for Adam.")
    parser.add_argument("--weight_decay", default=0.0, type=float,
                        help="Weight deay if we apply some.")
    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
                        help="Epsilon for Adam optimizer.")
    parser.add_argument("--max_grad_norm", default=1.0, type=float,
                        help="Max gradient norm.")
    parser.add_argument("--num_train_epochs", default=1, type=float,
                        help="Total number of training epochs to perform.")
    parser.add_argument("--max_steps", default=-1, type=int,
                        help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
    parser.add_argument("--warmup_steps", default=0, type=int,
                        help="Linear warmup over warmup_steps.")

    parser.add_argument('--logging_steps', type=int, default=50,
                        help="Log every X updates steps.")
    parser.add_argument('--save_steps', type=int, default=50000,
                        help="Save checkpoint every X updates steps.")
    parser.add_argument("--eval_all_checkpoints", action='store_true',
                        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
    parser.add_argument("--no_cuda", action='store_true',
                        help="Avoid using CUDA when available")
    parser.add_argument("--no_cache", action="store_true",
            help="Avoid caching features")
    parser.add_argument('--overwrite_output_dir', action='store_true',
                        help="Overwrite the content of the output directory")
    parser.add_argument('--overwrite_cache', action='store_true',
                        help="Overwrite the cached training and evaluation sets")
    parser.add_argument('--seed', type=int, default=42,
                        help="random seed for initialization")
    parser.add_argument("--optim", type=str, default="adam",
                        help="Optimizer class to use (one of {})".format(OPTIMIZERS.keys()))

    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("--local_rank", type=int, default=-1,
                        help="For distributed training: local_rank")
    parser.add_argument('--server_ip', type=str, default='', help="For distant debugging.")
    parser.add_argument('--server_port', type=str, default='', help="For distant debugging.")
    parser.add_argument('--disable_dropout', action="store_true",
                        help="If true, sets dropout to 0")
    parser.add_argument('--early_stopping_interval', type=int, default=0,
                        help="If > 0, checks for early stopping every `interval` iterations. "
                             "Eval during training must be true.")
    parser.add_argument('--early_stopping_patience', type=int, default=5,
                        help="Patience during early stopping (stops at `patience` lack of improvements)")
    parser.add_argument('--additional_eval', type=json.loads, default={},
                        help="Any additional datasets to evaluate for, in the form a dict mapping name:dataset path")
    parser.add_argument('--num_labels_per_task', type=str, default="",
                        help="Number of labels per task in multitask setting (ignored otherwise)")
    parser.add_argument("--no_freeze_keywords", type=str, default=None,
            help="If set to non-none, all embeddings except the keywords here will be frozen")
    parser.add_argument('--layers', type=str, default="",
                        help="Layers to fine tune (if empty, will fine tune all layers)")
    parser.add_argument('--constant_schedule', action="store_true",
                        help="If true, will have no linear warmup/cooldown")
    args = parser.parse_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))

    # Setup distant debugging if needed
    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

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

    # Set seed
    set_seed(args)

    # Prepare GLUE task
    args.task_name = args.task_name.lower()
    
    if args.task_name not in list(processors.keys()):
        raise ValueError("Task not found: %s" % (args.task_name))
        
    if args.task_name == "multitask":
        assert "multitask" in args.model_type
        
    processor = processors[args.task_name]()
    args.output_mode = output_modes[args.task_name]
    label_list = processor.get_labels()
    num_labels = len(label_list)

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

    args.model_type = args.model_type.lower()
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name)

    # handle config modifications for multitask
    if args.task_name == "multitask":
        args.num_labels_per_task = [int(x) for x in args.num_labels_per_task.split(",")]
        config.num_labels = sum(args.num_labels_per_task)
        config.num_labels_per_task = args.num_labels_per_task

    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
    
    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)

    # disable dropout
    if args.disable_dropout:
        model.bert.embeddings.dropout.p = 0
        for l in model.bert.encoder.layer:
            l.attention.self.dropout.p = 0
            l.attention.output.dropout.p = 0
            l.output.dropout.p = 0

    if args.local_rank == 0:
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab

    model.to(args.device)
    #print("Training/evaluation parameters ", args)


    # Training
    if args.do_train:
        train_dataset = load_and_cache_examples(args, args.data_dir, args.task_name, tokenizer,
                evaluate=False, no_cache=args.no_cache)
        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
        print(" global_step = ", global_step, " average loss = ", tr_loss)


    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
        # Create output directory if needed
        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
            os.makedirs(args.output_dir)

        print("Saving model checkpoint to ", args.output_dir)
        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
        # They can then be reloaded using `from_pretrained()`
        model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
        model_to_save.save_pretrained(args.output_dir)
        tokenizer.save_pretrained(args.output_dir)
        torch.save(model_to_save.state_dict(), os.path.join(args.output_dir, 'model.pt'))
        # Good practice: save your training arguments together with the trained model
        torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))

        # Load a trained model and vocabulary that you have fine-tuned
        model = model_class.from_pretrained(args.output_dir)
        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
        model.to(args.device)


    # Evaluation
    results = {}
    if args.do_eval and args.local_rank in [-1, 0]:
        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
        checkpoints = [args.output_dir]
        if args.eval_all_checkpoints:
            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
            logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
        print("Evaluate the following checkpoints: ", checkpoints)
        for checkpoint in checkpoints:
            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
            model = model_class.from_pretrained(checkpoint)
            model.to(args.device)
            result = evaluate(args, model, tokenizer, prefix=global_step)
            result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
            results.update(result)

    return results


if __name__ == "__main__":
    main()