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 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 transformers.trainer_utils import is_main_process
from transformers import (WEIGHTS_NAME, BertConfig, DebertaForSequenceClassification, DebertaTokenizer, DebertaConfig, ElectraForSequenceClassification, ElectraConfig, ElectraTokenizer,
                                  BertForSequenceClassification, BertTokenizer,
                                  RobertaConfig,
                                  RobertaForSequenceClassification,
                                  RobertaTokenizer,
                                  XLMConfig, XLMForSequenceClassification,
                                  XLMTokenizer, XLNetConfig,
                                  XLNetForSequenceClassification,
                                  XLNetTokenizer, AdamW, get_linear_schedule_with_warmup, AutoModelForSequenceClassification)
from transformers import glue_compute_metrics as compute_metrics
from transformers import glue_convert_examples_to_features as convert_examples_to_features
from transformers import glue_output_modes as output_modes
from transformers import glue_processors as processors
import transformers

from Pruner import *
from entropy import calc_entropy

logger = logging.getLogger(__name__)


MODEL_CLASSES = {
    'bert': (BertConfig, BertForSequenceClassification, BertTokenizer),
    'xlnet': (XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer),
    'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
    'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer),
    'deberta': (DebertaConfig, DebertaForSequenceClassification, DebertaTokenizer),
    'electra': (ElectraConfig, ElectraForSequenceClassification, ElectraTokenizer),
}


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)


def adaptive_kd_alpha(logits):
    num_classes = logits.shape[-1]
    dummy_tensor = torch.tensor([0.0]*num_classes).to(logits.device)
    max_entropy = calc_entropy(dummy_tensor).item()
    entropy = calc_entropy(logits)
    kd_coefficient = 1 - entropy / max_entropy
    return kd_coefficient

def kl(logits_t, logits, temp):
    p_t = torch.softmax(logits_t / temp, dim=-1)
    logp = torch.log_softmax(logits / temp, dim=-1)
    kl_divergence = -torch.sum(p_t*logp, dim=-1)
    return kl_divergence * (temp**2)


def train(args, train_dataset, model, tokenizer, teacher=None):
    """ Train the model """
    if args.local_rank in [-1, 0]:
        tb_writer = SummaryWriter()

    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=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}
    ]
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
    scheduler = get_linear_schedule_with_warmup(
        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=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!
    logger.info("***** Running training *****")
    logger.info("  Num examples = %d", len(train_dataset))
    logger.info("  Num Epochs = %d", args.num_train_epochs)
    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
    logger.info("  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))
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
    logger.info("  Total optimization steps = %d", t_total)
    # Distillation
    if teacher is not None:
        logger.info("  Training with distillation")
    No = os.path.split(args.output_dir)[-1].split('_')[0]

    epochs_trained = 0
    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)

    # Define the model pruner
    pruner = PINS(model, args=args, total_step=t_total, tb_writer=tb_writer, \
                    use_no_mask=False, pruner_name=args.pruner_name)

    from copy import deepcopy
    best_eval_ckpt = deepcopy(teacher.state_dict())
    best_eval_metric = -999.0
    best_eval_metric_stu = -999.0
    teacher_updated = False
    sparse_teacher = False

    for _ in train_iterator:
        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
        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 don't use segment_ids
                      }
            outputs = model(**inputs)
            logits = outputs[0]  
            loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
            # (batch_size, )
            ce_loss = loss_fct(logits.view(-1, model.config.num_labels), batch[3].view(-1))

            # Distillation loss
            if teacher is not None:
                if "token_type_ids" not in inputs:
                    inputs["token_type_ids"] = None if args.teacher_type == "xlm" else batch[2]
                if teacher_updated:
                    # update teacher's parameters
                    teacher.load_state_dict(best_eval_ckpt)
                    teacher_updated = False
                with torch.no_grad():
                    logits_tea = teacher(
                        input_ids=inputs["input_ids"],
                        token_type_ids=inputs["token_type_ids"],
                        attention_mask=inputs["attention_mask"],
                    )[0]

                # adaptive knowledge distillation coefficient
                # (batch_size, )
                kd_alpha = adaptive_kd_alpha(logits.clone().detach()).detach()

                # adjust for MNLI
                if args.task_name.lower() == 'mnli' and not sparse_teacher:
                    logits_tea = torch.cat([logits_tea[:, 1:], logits_tea[:, 0].unsqueeze(-1)], dim=-1)
                if args.task_name != 'sts-b':
                    # (batch_size, )
                    loss_logits = (
                        torch.nn.functional.kl_div(
                            input=torch.nn.functional.log_softmax(logits / args.temperature, dim=-1),
                            target=torch.nn.functional.softmax(logits_tea / args.temperature, dim=-1),
                            reduction="sum",
                        )
                        * (args.temperature ** 2)
                    )
                    loss_logits = kl(logits_tea, logits, args.temperature)
                else:
                    # loss_logits = torch.nn.functional.mse_loss(logits, logits_tea)
                    loss_logits = 0.0
                loss = kd_alpha * loss_logits + (1-kd_alpha) * ce_loss
                loss = loss.mean()

            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
                if args.final_threshold != 1.0:
                    # Do the pruning step 
                    threshold, mask_threshold = pruner.update_and_pruning(model, global_step, scheduler.get_lr()[0])
                else:
                    threshold = 0.0
                    mask_threshold = 0.0

                optimizer.step()
                pruner.post_prune(model, mask_threshold)
                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)
                        for key, value in results.items():
                            tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
                            logger.info("[Evaluation] Epoch {epoch:d} | Step: {global_step:d} | eval_{key} {value}".format(
                                epoch = epochs_trained, global_step = global_step, 
                                key = key, value = value,  
                            ))
                            # update best_eval_ckpt as teacher
                            if key in ['f1', 'acc', 'mcc'] and value > best_eval_metric:
                                best_eval_metric = value
                                best_eval_ckpt = deepcopy(model.state_dict())
                                teacher_updated = True
                                sparse_teacher = True
                                logger.info("Teacher model updated with eval_metric: {}".format(best_eval_metric))

                            if key in ['f1', 'acc', 'mcc'] and value > best_eval_metric_stu and abs(threshold-args.final_threshold)<=0.01:
                                best_eval_metric_stu = value
                                if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
                                    os.makedirs(args.output_dir)
                                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)
                                logger.info("Current Best eval_{}: {}".format(key, best_eval_metric_stu))

                    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)
                    tb_writer.add_scalar("prune_target", threshold, global_step)
                    if mask_threshold is not None:
                        tb_writer.add_scalar("mask_threshold", mask_threshold, global_step)
                    logger.info("[{No}] Epoch {epoch:d} | step: {global_step:d} | loss {loss:.5f} | threshold {threshold:.3f} | lc_update_coeff {lc_update_coeff}".format(
                            No = No, epoch = epochs_trained, global_step = global_step, 
                            loss=(tr_loss - logging_loss) / args.logging_steps,
                            threshold = threshold,
                            lc_update_coeff=pruner.lc_update_coeff
                        ))
                    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(args, os.path.join(output_dir, 'training_args.bin'))
                    logger.info("Saving model checkpoint to %s", output_dir)

                global_step += 1

            if args.max_steps > 0 and global_step > args.max_steps:
                epoch_iterator.close()
                break
        if args.max_steps > 0 and global_step > args.max_steps:
            train_iterator.close()
            break

    logger.info("Current Best eval metric: {}".format(best_eval_metric_stu))
    if args.local_rank in [-1, 0]:
        tb_writer.close()

    return global_step, tr_loss / global_step, pruner


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,)

    results = {}
    for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
        eval_dataset = load_and_cache_examples(args, 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 = 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!
        logger.info("***** Running evaluation {} *****".format(prefix))
        logger.info("  Num examples = %d", len(eval_dataset))
        logger.info("  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)
        results.update(result)

        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))
            for key in sorted(result.keys()):
                logger.info("  %s = %s", key, str(result[key]))
                writer.write("%s = %s\n" % (key, str(result[key])))

    return results


def load_and_cache_examples(args, task, tokenizer, evaluate=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(args.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):
        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()
        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(args.data_dir) if evaluate else processor.get_train_examples(
            args.data_dir)
        features = convert_examples_to_features(examples, tokenizer, args.max_seq_length, task, label_list, output_mode,
                                                )
        if args.local_rank in [-1, 0]:
            logger.info("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.attention_mask for f in features], dtype=torch.long)
    all_segment_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
    if output_mode == "classification":
        all_label_ids = torch.tensor([f.label for f in features], dtype=torch.long)
    elif output_mode == "regression":
        all_label_ids = torch.tensor([f.label 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: ")
    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("--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=3.0, 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('--logging_steps', type=int, default=50,
                        help="Log every X updates steps.")
    parser.add_argument('--save_steps', type=int, default=50,
                        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('--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('--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.")

    # adaptive pruning
    parser.add_argument('--pruner_name', default='PINS', type=str,
                        help="[PINS, Magnitude]")
    parser.add_argument('--beta1', default=0.85, type=float, help="beta1 for PINS.")
    parser.add_argument('--deltaT', default=1, type=int, help="The legnth of local window to reweight EMA.")
    parser.add_argument('--beta2', default=0., type=float, help="beta2 for PINS")
    # pruning schedule
    parser.add_argument('--warmup_steps', default=5400, type=int, help="Warmup steps.")
    parser.add_argument('--initial_threshold', default=1., type=float, help="Initial threshold.")
    parser.add_argument('--final_threshold', default=0.15, type=float, help="Final threshold.")
    parser.add_argument('--initial_warmup', default=1, type=int, help="Initial Warmup for Pruner.")
    parser.add_argument('--final_warmup', default=2, type=int, help="Final Warmup for Pruner.")

    # Distillation parameters (optional)
    parser.add_argument(
        "--teacher_type",
        default=None,
        type=str,
        help="Teacher type. Teacher tokenizer and student (model) tokenizer must output the same tokenization. Only for distillation.",
    )
    parser.add_argument(
        "--teacher_name_or_path",
        default=None,
        type=str,
        help="Path to the already fine-tuned teacher model. Only for distillation.",
    )
    parser.add_argument(
        "--ce_loss_weight", default=1.0, type=float, help="Cross entropy loss linear weight. Only for distillation."
    )
    parser.add_argument(
        "--distill_loss_weight", default=0.0, type=float, help="Distillation loss linear weight. Only for distillation."
    )
    parser.add_argument(
        "--temperature", default=2.0, type=float, help="Distillation temperature. Only for distillation."
    )
    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

    # Setup logging
    for h in logger.handlers:
        logger.removeHandler(h)
    os.makedirs(args.output_dir, exist_ok=True)
    logging.basicConfig(
        filename= os.path.join(args.output_dir, 'log.log'), filemode='a',
        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,
    )
    logging.getLogger().addHandler(logging.StreamHandler(sys.stdout))
    logging.info(args.output_dir)
    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,
    )

    # Set the verbosity to info of the Transformers logger (on main process only):
    if is_main_process(args.local_rank):
        transformers.utils.logging.set_verbosity_info()
        transformers.utils.logging.enable_default_handler()
        transformers.utils.logging.enable_explicit_format()
    # Set seed
    set_seed(args)

    # Prepare GLUE task
    args.task_name = args.task_name.lower()
    if args.task_name not in processors:
        raise ValueError("Task not found: %s" % (args.task_name))
    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)
    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)
    if args.teacher_type is not None:
        assert args.teacher_name_or_path is not None
        assert args.distill_loss_weight + args.ce_loss_weight > 0.0
        teacher = AutoModelForSequenceClassification.from_pretrained(args.teacher_name_or_path)
        teacher.to(args.device)
        logger.info("Using KD")
    else:
        teacher = None

    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)

    logger.info("Training/evaluation parameters %s", args)

    # Training
    if args.do_train:
        train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
        global_step, tr_loss, PINS = train(args, train_dataset, model, tokenizer, teacher=teacher)
        logger.info(" global_step = %s, average loss = %s", global_step, 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)
        # 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]:
        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
        logger.info("Evaluate the following checkpoints: %s", 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()