masked_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.
""" Fine-pruning Masked BERT on sequence classification on GLUE."""

import argparse
import glob
import json
import logging
import os
import random

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

from emmental import MaskedBertConfig, MaskedBertForSequenceClassification
from transformers import (
    WEIGHTS_NAME,
    AdamW,
    BertConfig,
    BertForSequenceClassification,
    BertTokenizer,
    get_linear_schedule_with_warmup,
)
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


try:
    from torch.utils.tensorboard import SummaryWriter
except ImportError:
    from tensorboardX import SummaryWriter


logger = logging.getLogger(__name__)

MODEL_CLASSES = {
    "bert": (BertConfig, BertForSequenceClassification, BertTokenizer),
    "masked_bert": (MaskedBertConfig, MaskedBertForSequenceClassification, BertTokenizer),
}


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 schedule_threshold(
    step: int,
    total_step: int,
    warmup_steps: int,
    initial_threshold: float,
    final_threshold: float,
    initial_warmup: int,
    final_warmup: int,
    final_lambda: float,
):
    if step <= initial_warmup * warmup_steps:
        threshold = initial_threshold
    elif step > (total_step - final_warmup * warmup_steps):
        threshold = final_threshold
    else:
        spars_warmup_steps = initial_warmup * warmup_steps
        spars_schedu_steps = (final_warmup + initial_warmup) * warmup_steps
        mul_coeff = 1 - (step - spars_warmup_steps) / (total_step - spars_schedu_steps)
        threshold = final_threshold + (initial_threshold - final_threshold) * (mul_coeff ** 3)
    regu_lambda = final_lambda * threshold / final_threshold
    return threshold, regu_lambda


def regularization(model: nn.Module, mode: str):
    regu, counter = 0, 0
    for name, param in model.named_parameters():
        if "mask_scores" in name:
            if mode == "l1":
                regu += torch.norm(torch.sigmoid(param), p=1) / param.numel()
            elif mode == "l0":
                regu += torch.sigmoid(param - 2 / 3 * np.log(0.1 / 1.1)).sum() / param.numel()
            else:
                ValueError("Don't know this mode.")
            counter += 1
    return regu / counter


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

    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 "mask_score" in n and p.requires_grad],
            "lr": args.mask_scores_learning_rate,
        },
        {
            "params": [
                p
                for n, p in model.named_parameters()
                if "mask_score" not in n and p.requires_grad and not any(nd in n for nd in no_decay)
            ],
            "lr": args.learning_rate,
            "weight_decay": args.weight_decay,
        },
        {
            "params": [
                p
                for n, p in model.named_parameters()
                if "mask_score" not in n and p.requires_grad and any(nd in n for nd in no_decay)
            ],
            "lr": args.learning_rate,
            "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
    )

    # Check if saved optimizer or scheduler states exist
    if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
        os.path.join(args.model_name_or_path, "scheduler.pt")
    ):
        # Load in optimizer and scheduler states
        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))

    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 = nn.DataParallel(model)

    # Distributed training (should be after apex fp16 initialization)
    if args.local_rank != -1:
        model = 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")

    global_step = 0
    # Global TopK
    if args.global_topk:
        threshold_mem = None
    epochs_trained = 0
    steps_trained_in_current_epoch = 0
    # Check if continuing training from a checkpoint
    if os.path.exists(args.model_name_or_path):
        # set global_step to global_step of last saved checkpoint from model path
        try:
            global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
        except ValueError:
            global_step = 0
        epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
        steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)

        logger.info("  Continuing training from checkpoint, will skip to saved global_step")
        logger.info("  Continuing training from epoch %d", epochs_trained)
        logger.info("  Continuing training from global step %d", global_step)
        logger.info("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)

    tr_loss, logging_loss = 0.0, 0.0
    model.zero_grad()
    train_iterator = trange(
        epochs_trained,
        int(args.num_train_epochs),
        desc="Epoch",
        disable=args.local_rank not in [-1, 0],
    )
    set_seed(args)  # Added here for reproducibility
    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):

            # Skip past any already trained steps if resuming training
            if steps_trained_in_current_epoch > 0:
                steps_trained_in_current_epoch -= 1
                continue

            model.train()
            batch = tuple(t.to(args.device) for t in batch)
            threshold, regu_lambda = schedule_threshold(
                step=global_step,
                total_step=t_total,
                warmup_steps=args.warmup_steps,
                final_threshold=args.final_threshold,
                initial_threshold=args.initial_threshold,
                final_warmup=args.final_warmup,
                initial_warmup=args.initial_warmup,
                final_lambda=args.final_lambda,
            )
            # Global TopK
            if args.global_topk:
                if threshold == 1.0:
                    threshold = -1e2  # Or an indefinitely low quantity
                else:
                    if (threshold_mem is None) or (global_step % args.global_topk_frequency_compute == 0):
                        # Sort all the values to get the global topK
                        concat = torch.cat(
                            [param.view(-1) for name, param in model.named_parameters() if "mask_scores" in name]
                        )
                        n = concat.numel()
                        kth = max(n - (int(n * threshold) + 1), 1)
                        threshold_mem = concat.kthvalue(kth).values.item()
                        threshold = threshold_mem
                    else:
                        threshold = threshold_mem
            inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
            if args.model_type != "distilbert":
                inputs["token_type_ids"] = (
                    batch[2] if args.model_type in ["bert", "masked_bert", "xlnet", "albert"] else None
                )  # XLM, DistilBERT, RoBERTa, and XLM-RoBERTa don't use segment_ids

            if "masked" in args.model_type:
                inputs["threshold"] = threshold

            outputs = model(**inputs)
            loss, logits_stu = outputs  # model outputs are always tuple in transformers (see doc)

            # 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]
                with torch.no_grad():
                    (logits_tea,) = teacher(
                        input_ids=inputs["input_ids"],
                        token_type_ids=inputs["token_type_ids"],
                        attention_mask=inputs["attention_mask"],
                    )

                loss_logits = (
                    nn.functional.kl_div(
                        input=nn.functional.log_softmax(logits_stu / args.temperature, dim=-1),
                        target=nn.functional.softmax(logits_tea / args.temperature, dim=-1),
                        reduction="batchmean",
                    )
                    * (args.temperature ** 2)
                )

                loss = args.alpha_distil * loss_logits + args.alpha_ce * loss

            # Regularization
            if args.regularization is not None:
                regu_ = regularization(model=model, mode=args.regularization)
                loss = loss + regu_lambda * regu_

            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()
            else:
                loss.backward()

            tr_loss += loss.item()
            if (step + 1) % args.gradient_accumulation_steps == 0 or (
                # last step in epoch but step is always smaller than gradient_accumulation_steps
                len(epoch_iterator) <= args.gradient_accumulation_steps
                and (step + 1) == len(epoch_iterator)
            ):
                if args.fp16:
                    nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
                else:
                    nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)

                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
                    tb_writer.add_scalar("threshold", threshold, global_step)
                    for name, param in model.named_parameters():
                        if not param.requires_grad:
                            continue
                        tb_writer.add_scalar("parameter_mean/" + name, param.data.mean(), global_step)
                        tb_writer.add_scalar("parameter_std/" + name, param.data.std(), global_step)
                        tb_writer.add_scalar("parameter_min/" + name, param.data.min(), global_step)
                        tb_writer.add_scalar("parameter_max/" + name, param.data.max(), global_step)
                        tb_writer.add_scalar("grad_mean/" + name, param.grad.data.mean(), global_step)
                        tb_writer.add_scalar("grad_std/" + name, param.grad.data.std(), global_step)
                        if args.regularization is not None and "mask_scores" in name:
                            if args.regularization == "l1":
                                perc = (torch.sigmoid(param) > threshold).sum().item() / param.numel()
                            elif args.regularization == "l0":
                                perc = (torch.sigmoid(param - 2 / 3 * np.log(0.1 / 1.1))).sum().item() / param.numel()
                            tb_writer.add_scalar("retained_weights_perc/" + name, perc, global_step)

                optimizer.step()
                scheduler.step()  # Update learning rate schedule
                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:
                    logs = {}
                    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():
                            eval_key = "eval_{}".format(key)
                            logs[eval_key] = value

                    loss_scalar = (tr_loss - logging_loss) / args.logging_steps
                    learning_rate_scalar = scheduler.get_lr()
                    logs["learning_rate"] = learning_rate_scalar[0]
                    if len(learning_rate_scalar) > 1:
                        for idx, lr in enumerate(learning_rate_scalar[1:]):
                            logs[f"learning_rate/{idx+1}"] = lr
                    logs["loss"] = loss_scalar
                    if teacher is not None:
                        logs["loss/distil"] = loss_logits.item()
                    if args.regularization is not None:
                        logs["loss/regularization"] = regu_.item()
                    if (teacher is not None) or (args.regularization is not None):
                        if (teacher is not None) and (args.regularization is not None):
                            logs["loss/instant_ce"] = (
                                loss.item()
                                - regu_lambda * logs["loss/regularization"]
                                - args.alpha_distil * logs["loss/distil"]
                            ) / args.alpha_ce
                        elif teacher is not None:
                            logs["loss/instant_ce"] = (
                                loss.item() - args.alpha_distil * logs["loss/distil"]
                            ) / args.alpha_ce
                        else:
                            logs["loss/instant_ce"] = loss.item() - regu_lambda * logs["loss/regularization"]
                    logging_loss = tr_loss

                    for key, value in logs.items():
                        tb_writer.add_scalar(key, value, global_step)
                    print(json.dumps({**logs, **{"step": global_step}}))

                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)
                    tokenizer.save_pretrained(output_dir)

                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
                    logger.info("Saving model checkpoint to %s", output_dir)

                    torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
                    torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
                    logger.info("Saving optimizer and scheduler states to %s", output_dir)

            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

    if args.local_rank in [-1, 0]:
        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,)

    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)
        eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)

        # multi-gpu eval
        if args.n_gpu > 1 and not isinstance(model, nn.DataParallel):
            model = nn.DataParallel(model)

        # 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

        # Global TopK
        if args.global_topk:
            threshold_mem = 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], "labels": batch[3]}
                if args.model_type != "distilbert":
                    inputs["token_type_ids"] = (
                        batch[2] if args.model_type in ["bert", "masked_bert", "xlnet", "albert"] else None
                    )  # XLM, DistilBERT, RoBERTa, and XLM-RoBERTa don't use segment_ids
                if "masked" in args.model_type:
                    inputs["threshold"] = args.final_threshold
                    if args.global_topk:
                        if threshold_mem is None:
                            concat = torch.cat(
                                [param.view(-1) for name, param in model.named_parameters() if "mask_scores" in name]
                            )
                            n = concat.numel()
                            kth = max(n - (int(n * args.final_threshold) + 1), 1)
                            threshold_mem = concat.kthvalue(kth).values.item()
                        inputs["threshold"] = threshold_mem
                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":
            from scipy.special import softmax

            probs = softmax(preds, axis=-1)
            entropy = np.exp((-probs * np.log(probs)).sum(axis=-1).mean())
            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)
        if entropy is not None:
            result["eval_avg_entropy"] = entropy

        output_eval_file = os.path.join(eval_output_dir, prefix, "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) and not args.overwrite_cache:
        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", "xlmroberta"]:
            # 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,
            max_length=args.max_seq_length,
            label_list=label_list,
            output_mode=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_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
    all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
    if output_mode == "classification":
        all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
    elif output_mode == "regression":
        all_labels = torch.tensor([f.label for f in features], dtype=torch.float)

    dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
    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 pretrained model or model identifier from huggingface.co/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 huggingface.co",
    )
    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="Run 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("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")

    # Pruning parameters
    parser.add_argument(
        "--mask_scores_learning_rate",
        default=1e-2,
        type=float,
        help="The Adam initial learning rate of the mask scores.",
    )
    parser.add_argument(
        "--initial_threshold", default=1.0, type=float, help="Initial value of the threshold (for scheduling)."
    )
    parser.add_argument(
        "--final_threshold", default=0.7, type=float, help="Final value of the threshold (for scheduling)."
    )
    parser.add_argument(
        "--initial_warmup",
        default=1,
        type=int,
        help="Run `initial_warmup` * `warmup_steps` steps of threshold warmup during which threshold stays"
        "at its `initial_threshold` value (sparsity schedule).",
    )
    parser.add_argument(
        "--final_warmup",
        default=2,
        type=int,
        help="Run `final_warmup` * `warmup_steps` steps of threshold cool-down during which threshold stays"
        "at its final_threshold value (sparsity schedule).",
    )

    parser.add_argument(
        "--pruning_method",
        default="topK",
        type=str,
        help="Pruning Method (l0 = L0 regularization, magnitude = Magnitude pruning, topK = Movement pruning, sigmoied_threshold = Soft movement pruning).",
    )
    parser.add_argument(
        "--mask_init",
        default="constant",
        type=str,
        help="Initialization method for the mask scores. Choices: constant, uniform, kaiming.",
    )
    parser.add_argument(
        "--mask_scale", default=0.0, type=float, help="Initialization parameter for the chosen initialization method."
    )

    parser.add_argument("--regularization", default=None, help="Add L0 or L1 regularization to the mask scores.")
    parser.add_argument(
        "--final_lambda",
        default=0.0,
        type=float,
        help="Regularization intensity (used in conjunction with `regularization`.",
    )

    parser.add_argument("--global_topk", action="store_true", help="Global TopK on the Scores.")
    parser.add_argument(
        "--global_topk_frequency_compute",
        default=25,
        type=int,
        help="Frequency at which we compute the TopK global threshold.",
    )

    # 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(
        "--alpha_ce", default=0.5, type=float, help="Cross entropy loss linear weight. Only for distillation."
    )
    parser.add_argument(
        "--alpha_distil", default=0.5, 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."
    )

    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("--weight_decay", default=0.0, type=float, help="Weight decay 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("--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=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")

    args = parser.parse_args()

    # Regularization
    if args.regularization == "null":
        args.regularization = None

    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(
            f"Output directory ({args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
        )

    # 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 = 0 if args.no_cuda else torch.cuda.device_count()
    else:  # Initializes the distributed backend which will take care of synchronizing nodes/GPUs
        torch.cuda.set_device(args.local_rank)
        device = torch.device("cuda", args.local_rank)
        torch.distributed.init_process_group(backend="nccl")
        args.n_gpu = 1
    args.device = device

    # Setup logging
    logging.basicConfig(
        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.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 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,
        cache_dir=args.cache_dir if args.cache_dir else None,
        pruning_method=args.pruning_method,
        mask_init=args.mask_init,
        mask_scale=args.mask_scale,
    )
    tokenizer = tokenizer_class.from_pretrained(
        args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
        cache_dir=args.cache_dir if args.cache_dir else None,
        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,
        cache_dir=args.cache_dir if args.cache_dir else None,
    )

    if args.teacher_type is not None:
        assert args.teacher_name_or_path is not None
        assert args.alpha_distil > 0.0
        assert args.alpha_distil + args.alpha_ce > 0.0
        teacher_config_class, teacher_model_class, _ = MODEL_CLASSES[args.teacher_type]
        teacher_config = teacher_config_class.from_pretrained(args.teacher_name_or_path)
        teacher = teacher_model_class.from_pretrained(
            args.teacher_name_or_path,
            from_tf=False,
            config=teacher_config,
            cache_dir=args.cache_dir if args.cache_dir else None,
        )
        teacher.to(args.device)
    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 = 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):
        logger.info("Saving model checkpoint to %s", 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)

        # 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))
            )

        logger.info("Evaluate the following checkpoints: %s", checkpoints)
        for checkpoint in checkpoints:
            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
            prefix = checkpoint.split("/")[-1] if checkpoint.find("checkpoint") != -1 else ""

            model = model_class.from_pretrained(checkpoint)
            model.to(args.device)
            result = evaluate(args, model, tokenizer, prefix=prefix)
            result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
            results.update(result)

    return results


if __name__ == "__main__":
    main()