main.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-tuning the library models for named entity recognition on CoNLL-2003 (Bert or Roberta). """

from __future__ import absolute_import, division, print_function

import argparse
import glob
import logging
import os
import random
import time

import numpy as np
import torch
from seqeval.metrics import precision_score, recall_score, f1_score
from tensorboardX import SummaryWriter
from torch.nn import CrossEntropyLoss
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
from torch.utils.data.distributed import DistributedSampler
from utils_ner import convert_examples_to_features, get_labels, read_examples_from_file, read_examples_from_file_random

from transformers import AdamW, WarmupLinearSchedule
from transformers import WEIGHTS_NAME, BertConfig, BertTokenizer#, BertForTokenClassification
from transformers import RobertaConfig, RobertaForTokenClassification, RobertaTokenizer
from modelings.modeling_MSD import my_BertsoftmaxNER, my_BertsoftmaxNER_12class, my_BertsoftmaxNER_MMD, my_BertsoftmaxNER_MMD_KD, my_BertsoftmaxNER_unit
from itertools import cycle

logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)

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

MODEL_CLASSES = {
    "bert": (BertConfig, my_BertsoftmaxNER_unit, BertTokenizer),
    "roberta": (RobertaConfig, RobertaForTokenClassification, RobertaTokenizer)
}


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)
    torch.backends.cudnn.deterministic = True


def train(args, model, train_dataset, tokenizer, labels, pad_token_label_id):
    """ Train the model """
    if args.local_rank in [-1, 0]:
        tb_writer = SummaryWriter(log_dir=args.log_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"]
    no_grad = ["embeddings"] + ["layer." + str(layer_i) + "." for layer_i in range(12) if layer_i < args.freeze_bottom_layer]
    logger.info("  The frozen parameters are:")
    for n, p in model.named_parameters():
        p.requires_grad = False if any(nd in n for nd in no_grad) else True
        if not p.requires_grad:
            logger.info("    %s", n)
    optimizer_grouped_parameters = [
        {"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay) and p.requires_grad],
         "weight_decay": args.weight_decay},
        {"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay) and p.requires_grad],
         "weight_decay": 0.0}
    ]
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
    # scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=int(t_total*args.warmup_ratio), 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, device_ids=args.gpu_ids)

    # 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("  GPU IDs for training: %s", " ".join([str(id) for id in args.gpu_ids]))
    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)

    best_score = 0.0
    global_step = 0
    tr_loss, logging_loss = 0.0, 0.0
    model.zero_grad()
    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
    random_name_list = random.sample(list(range(100)),args.num_train_epochs)
    for epoch_i in range(args.num_train_epochs):
        train_dataset_noen = load_and_cache_examples_noen(args, str(random_name_list[epoch_i]), args.mmd_data_dir, tokenizer, labels, pad_token_label_id, mode="train")
        train_sampler_noen = RandomSampler(train_dataset_noen) if args.local_rank == -1 else DistributedSampler(train_dataset_noen)
        train_dataloader_noen = DataLoader(train_dataset_noen, sampler=train_sampler_noen, batch_size=args.train_batch_size)
        for step, batch_2 in enumerate(zip(train_dataloader,cycle(train_dataloader_noen))):
            batch = batch_2[0]
            batch_noen = batch_2[1]
            model.train()
            # batch = tuple(t.to(args.device) for t in batch)
            inputs = {"input_ids": batch[0].to(args.device),
                      "attention_mask": batch[1].to(args.device),
                      "token_type_ids": batch[2].to(args.device) if args.model_type in ["bert", "xlnet"] else None,
                      # XLM and RoBERTa don"t use segment_ids
                      "labels": batch[3].to(args.device) if len(batch) <= 4 else batch[-1],
                      "input_ids_other": batch_noen[0].to(args.device),
                      "attention_mask_other": batch_noen[1].to(args.device),
                      "token_type_ids_other": batch_noen[2].to(args.device) if args.model_type in ["bert", "xlnet"] else None,} # add hard-label scheme
            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()
            else:
                loss.backward()

            tr_loss += loss.item()
            if (step + 1) % args.gradient_accumulation_steps == 0:
                if args.fp16:
                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
                else:
                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)

                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
                        logger.info("===== evaluate_during_training =====")
                        results, _ = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="dev")

                        

                        if results["f1"] > best_score:
                            logger.info("results['f1']={} > best_score={}".format(results["f1"], best_score))
                            best_score = results["f1"]
                            # Save the best model checkpoint
                            if not os.path.exists(args.output_dir):
                              os.makedirs(args.output_dir)
                            # Take care of distributed/parallel training
                            model_to_save = model.module if hasattr(model, "module") else model
                            model_to_save.save_pretrained(args.output_dir)
                            tokenizer.save_pretrained(args.output_dir)
                            torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
                            logger.info("Saving the best model checkpoint to %s", args.output_dir)

                        for key, value in results.items():
                            tb_writer.add_scalar("eval_{}".format(key), value, global_step)
                            logger.info("Epoch: {}\t global_step: {}\t eval_{}: {}".format(epoch_i,
                                                                                           global_step, key, value))
                    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)
                    logger.info(
                        "Epoch: {}\t global_step: {}\t learning rate: {:.8}\t loss: {:.4f}".format(
                            epoch_i, global_step, scheduler.get_lr()[0],
                            (tr_loss - logging_loss) / args.logging_steps))
                    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)

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

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

    return global_step, tr_loss / global_step

def train_KD(args, model, train_dataset, src_probs, tokenizer, labels, pad_token_label_id):
    """ Train the model """
    if args.local_rank in [-1, 0]:
        tb_writer = SummaryWriter(log_dir=args.log_dir)

    # check the size of the two relative datasets, expand train_dataset with src_probs
    assert len(train_dataset) == src_probs.size(0)
    train_dataset.tensors += (src_probs,)

    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"]
    no_grad = ["embeddings"] + ["layer." + str(layer_i) + "." for layer_i in range(12) if
                                layer_i < args.freeze_bottom_layer]
    logger.info("  The frozen parameters are:")
    for n, p in model.named_parameters():
        p.requires_grad = False if any(nd in n for nd in no_grad) else True
        if not p.requires_grad:
            logger.info("    %s", n)
    optimizer_grouped_parameters = [
        {"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay) and p.requires_grad],
         "weight_decay": args.weight_decay},
        {"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay) and p.requires_grad],
         "weight_decay": 0.0}
    ]
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
    # scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=int(t_total*args.warmup_ratio), 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, device_ids=args.gpu_ids)

    # 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("  GPU IDs for training: %s", " ".join([str(id) for id in args.gpu_ids]))
    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)

    best_score = 0.0
    global_step = 0
    tr_loss, logging_loss = 0.0, 0.0
    tr_loss_KD, logging_loss_KD = 0.0, 0.0
    model.zero_grad()
    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
    random_name_list = random.sample(list(range(100)),args.num_train_epochs)
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
    for epoch_i in range(args.num_train_epochs):
        train_dataset_en = load_and_cache_examples_noen(args, str(random_name_list[epoch_i]), args.mmd_en_data_dir, tokenizer, labels, pad_token_label_id, mode="train")
        tea_en_cls = get_cls(args, train_dataset_en, model_class, src_lang=args.src_langs[0])

        assert len(train_dataset_en) == tea_en_cls.size(0)
        train_dataset_en.tensors += (tea_en_cls,)
        train_sampler_cls = RandomSampler(train_dataset_en) if args.local_rank == -1 else DistributedSampler(train_dataset_en)
        train_dataloader_cls = DataLoader(train_dataset_en, sampler=train_sampler_cls, batch_size=args.train_batch_size)
        for step, batch_2 in enumerate(zip(train_dataloader,cycle(train_dataloader_cls))):
            batch = batch_2[0]
            batch_cls = batch_2[1]
            model.train()
            # batch = tuple(t.to(args.device) for t in batch)
            inputs = {"input_ids": batch[0].to(args.device),
                      "attention_mask": batch[1].to(args.device),
                      "token_type_ids": batch[2].to(args.device) if args.model_type in ["bert", "xlnet"] else None,
                      # XLM and RoBERTa don"t use segment_ids
                      "labels": batch[3].to(args.device),
                      "src_probs": batch[4],
                      "input_ids_other": batch_cls[0].to(args.device),
                      "attention_mask_other": batch_cls[1].to(args.device),
                      "token_type_ids_other": batch_cls[2].to(args.device) if args.model_type in ["bert", "xlnet"] else None,
                      "tea_cls_en": batch_cls[4]} # activate the KD loss
            outputs = model(**inputs)
            # loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)
            loss_KD, loss = outputs[:2]

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

            if args.fp16:
                # with amp.scale_loss(loss, optimizer) as scaled_loss:
                #     scaled_loss.backward()
                with amp.scale_loss(loss_KD, optimizer) as scaled_loss:
                    scaled_loss.backward()
            else:
                # loss.backward()
                loss_KD.backward()

            tr_loss += loss.item()
            tr_loss_KD += loss_KD.item()
            if (step + 1) % args.gradient_accumulation_steps == 0:
                if args.fp16:
                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
                else:
                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)

                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
                        logger.info("===== evaluate_during_training =====")
                        results, _ = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="dev")

                        if results["f1"] > best_score:
                            logger.info("results['f1']={} > best_score={}".format(results["f1"], best_score))
                            best_score = results["f1"]
                            # Save the best model checkpoint
                            if not os.path.exists(args.output_dir):
                              os.makedirs(args.output_dir)
                            # Take care of distributed/parallel training
                            model_to_save = model.module if hasattr(model, "module") else model
                            model_to_save.save_pretrained(args.output_dir)
                            tokenizer.save_pretrained(args.output_dir)
                            torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
                            logger.info("Saving the best model checkpoint to %s", args.output_dir)

                        for key, value in results.items():
                            tb_writer.add_scalar("eval_{}".format(key), value, global_step)
                            logger.info("Epoch: {}\t global_step: {}\t eval_{}: {}".format(epoch_i,
                                                                                           global_step, key, value))
                    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("loss_KD", (tr_loss_KD - logging_loss_KD) / args.logging_steps, global_step)
                    logger.info(
                        "Epoch: {}\t global_step: {}\t learning rate: {:.8}\t loss: {:.4f}\t loss_KD: {:.4f}".format(
                            epoch_i, global_step, scheduler.get_lr()[0],
                            (tr_loss - logging_loss) / args.logging_steps,
                            (tr_loss_KD - logging_loss_KD) / args.logging_steps))
                    logging_loss = tr_loss
                    logging_loss_KD = tr_loss_KD

                # 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)

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

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

    return global_step, tr_loss_KD / global_step, tr_loss / global_step


def evaluate(args, model, tokenizer, labels, pad_token_label_id, mode, prefix=""):
    eval_dataset = load_and_cache_examples(args, tokenizer, args.data_dir, labels, pad_token_label_id, mode=mode)

    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 %s *****", 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
    model.eval()
    for batch in eval_dataloader:
        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.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
    preds = np.argmax(preds, axis=2)

    label_map = {i: label for i, label in enumerate(labels)}

    out_label_list = [[] for _ in range(out_label_ids.shape[0])]
    preds_list = [[] for _ in range(out_label_ids.shape[0])]

    for i in range(out_label_ids.shape[0]):
        for j in range(out_label_ids.shape[1]):
            if out_label_ids[i, j] != pad_token_label_id:
                out_label_list[i].append(label_map[out_label_ids[i][j]])
                preds_list[i].append(label_map[preds[i][j]])

    results = {
        "loss": eval_loss,
        "precision": precision_score(out_label_list, preds_list),
        "recall": recall_score(out_label_list, preds_list),
        "f1": f1_score(out_label_list, preds_list)
    }

    logger.info("***** Eval results %s *****", prefix)
    for key in sorted(results.keys()):
        logger.info("  %s = %s", key, str(results[key]))

    return results, preds_list


def load_and_cache_examples(args, tokenizer, used_data, labels, pad_token_label_id, mode):
    if args.local_rank not in [-1, 0] and not evaluate:  # for distributed training
        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache

    # Load data features from cache or dataset file
    cached_features_file = os.path.join(used_data, "cached_{}_{}_{}".format(mode,
                                                                                list(filter(None,
                                                                                            args.model_name_or_path.split(
                                                                                                "/"))).pop(),
                                                                                str(args.max_seq_length)))
    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", used_data)
        examples = read_examples_from_file(used_data, mode)
        features = convert_examples_to_features(examples, labels, args.max_seq_length, tokenizer,
                                                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,
                                                pad_token_label_id=pad_token_label_id
                                                )
        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

    if args.unlabeled_data_ratio < 1.0:
        logger.info("==> len of features: {}. data ratio: {}".format(len(features), args.unlabeled_data_ratio))
        features = random.sample(features, int(len(features) * args.unlabeled_data_ratio))
        logger.info("    len of features: {}".format(len(features)))


    # 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)
    all_label_ids = torch.tensor([f.label_ids for f in features], dtype=torch.long)

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


def load_and_cache_examples_noen(args, filenumber, used_data, tokenizer, labels, pad_token_label_id, mode):
    if args.local_rank not in [-1, 0] and not evaluate:  # for distributed training
        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache

    # Load data features from cache or dataset file
    cached_features_file = os.path.join(used_data, "cached_{}_{}_{}_{}".format(mode,filenumber,
                                                                                list(filter(None,
                                                                                            args.model_name_or_path.split(
                                                                                                "/"))).pop(),
                                                                                str(args.max_seq_length)))
    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", used_data)
        examples = read_examples_from_file_random(used_data, mode, filenumber)
        features = convert_examples_to_features(examples, labels, args.max_seq_length, tokenizer,
                                                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,
                                                pad_token_label_id=pad_token_label_id
                                                )
        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

    if args.unlabeled_data_ratio < 1.0:
        logger.info("==> len of features: {}. data ratio: {}".format(len(features), args.unlabeled_data_ratio))
        features = random.sample(features, int(len(features) * args.unlabeled_data_ratio))
        logger.info("    len of features: {}".format(len(features)))


    # 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)
    all_label_ids = torch.tensor([f.label_ids for f in features], dtype=torch.long)

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




def get_src_probs(args, dataset, model_class, src_lang):
    """ without softmax.
        preds: dataset_len x seq_len x label_len
    """
    # load src model
    src_model_path = os.path.join(args.src_model_dir, "{}{}".format(args.src_model_dir_prefix, src_lang))
    src_model = model_class.from_pretrained(src_model_path)
    src_model.to(args.device)

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

    # compute logits for the dataset using the model!
    logger.info("***** Compute logits for [%s] dataset using the model [%s] *****", os.path.basename(args.data_dir), src_model_path)
    logger.info("  Num examples = %d", len(dataset))
    logger.info("  Batch size = %d", args.eval_batch_size)
    # eval_loss = 0.0
    # nb_eval_steps = 0
    preds = None
    src_model.eval()
    for batch in eval_dataloader:
        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": None} #batch[3]}
            outputs = src_model(**inputs)
            logits = outputs[0]
            # tmp_eval_loss, logits = outputs[:2]
            # eval_loss += tmp_eval_loss.item()

        # nb_eval_steps += 1
        preds = logits.detach() if preds is None else torch.cat((preds, logits.detach()), dim=0) # dataset_len x max_seq_len x label_len

    # eval_loss = eval_loss / nb_eval_steps
    preds = torch.nn.functional.softmax(preds, dim=-1)

    return preds #, eval_loss


def get_cls(args, dataset, model_class, src_lang):
    """ return cls_en
    """
    # load src model
    src_model_path = os.path.join(args.src_model_dir, "{}{}".format(args.src_model_dir_prefix, src_lang))
    src_model = model_class.from_pretrained(src_model_path)
    src_model.to(args.device)

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

    # compute logits for the dataset using the model!
    logger.info("***** Compute cls for [%s] dataset using the model [%s] *****", os.path.basename(args.mmd_en_data_dir), src_model_path)
    logger.info("  Num examples = %d", len(dataset))
    logger.info("  Batch size = %d", args.eval_batch_size)
    # eval_loss = 0.0
    # nb_eval_steps = 0
    preds = None
    src_model.eval()
    for batch in eval_dataloader:
        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": None} #batch[3]}
            outputs = src_model(**inputs)
            logits = outputs[2][-1][:,0]
            # tmp_eval_loss, logits = outputs[:2]
            # eval_loss += tmp_eval_loss.item()

        # nb_eval_steps += 1
        preds = logits.detach() if preds is None else torch.cat((preds, logits.detach()), dim=0) # dataset_len x label_len

    return preds #, eval_loss

def get_src_probs_layers(args, dataset, model_class, src_lang):
    """ without softmax.
        preds: dataset_len x seq_len x label_len
    """
    # load src model
    src_model_path = os.path.join(args.src_model_dir, "{}{}".format(args.src_model_dir_prefix, src_lang))
    src_model = model_class.from_pretrained(src_model_path)
    src_model.to(args.device)

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

    # compute logits for the dataset using the model!
    logger.info("***** Compute logits for [%s] dataset using the model [%s] *****", os.path.basename(args.data_dir), src_model_path)
    logger.info("  Num examples = %d", len(dataset))
    logger.info("  Batch size = %d", args.eval_batch_size)
    # eval_loss = 0.0
    # nb_eval_steps = 0
    preds = None
    final_probs = None
    all_preds = [None for i in range(9)]
    src_model.eval()
    for batch in eval_dataloader:
        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": None} #batch[3]}
            outputs = src_model(**inputs)
            logits = outputs[0]
            logits_layers = outputs[1]
            all_logits = [None for i in range(9)]
            for i in range(8):
              all_logits[i] = logits_layers[i]
            all_logits[8] = logits
            
            # tmp_eval_loss, logits = outputs[:2]
            # eval_loss += tmp_eval_loss.item()

        # nb_eval_steps += 1
        # preds = logits.detach() if preds is None else torch.cat((preds, logits.detach()), dim=0) # dataset_len x max_seq_len x label_len
        
        for i in range(9):
          all_preds[i] = all_logits[i].detach() if all_preds[i] is None else torch.cat((all_preds[i], all_logits[i].detach()), dim=0)
        



    # eval_loss = eval_loss / nb_eval_steps
    
    for i in range(9):
      all_preds[i] = torch.nn.functional.softmax(all_preds[i], dim=-1).unsqueeze(0)
      final_probs = all_preds[i] if final_probs is None else torch.cat((final_probs,all_preds[i]),dim=0)
    # preds = torch.nn.functional.softmax(preds, dim=-1)
    final_probs = final_probs.transpose(0,1)
    print("src_probs done", final_probs.shape)

    # return preds #, eval_loss

    return final_probs


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

    ## Required parameters
    parser.add_argument("--data_dir", default="./data/ner/conll/debug", type=str,  # required=True,
                        help="The input data dir. Should contain the training files for the CoNLL-2003 NER task.")
    parser.add_argument("--mmd_data_dir", default="", type=str,  # required=True,
                        help="The input data dir. Should contain the training files for the CoNLL-2003 NER task.")
    parser.add_argument("--mmd_en_data_dir", default="", type=str,  # required=True,
                        help="The input data dir. Should contain the training files for the CoNLL-2003 NER task.")
    parser.add_argument("--output_dir", default="conll-model/test", type=str,  # required=True,
                        help="The output directory where the model predictions and checkpoints will be written.")

    parser.add_argument("--src_model_dir", default="conll-model/", type=str,
                        help="path to load teacher models")
    parser.add_argument("--src_model_dir_prefix", default="mono-src-", type=str,
                        help="prefix of the teacher model dir (to indicate the model type)")
    parser.add_argument("--src_langs", type=str, nargs="+", default="en",
                        help="source languages used for multi-teacher models")

    parser.add_argument("--unlabeled_data_ratio", type=float, default=1.0,
                        help="Ratio of the training data to use.")

    ## Other parameters
    parser.add_argument("--model_type", default='bert', type=str,  # required=True,
                        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
    parser.add_argument("--model_name_or_path", default='bert-base-multilingual-cased', type=str,  # required=True,
                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
    parser.add_argument("--bert_path", default='./bert-base-multilingual-cased', type=str,  # required=True,
                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
    parser.add_argument("--labels", default="./data/ner/conll/labels.txt", type=str,
                        help="Path to a file containing all labels. If not specified, CoNLL-2003 labels are used.")
    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("--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("--do_train", action="store_true",
                        help="Whether to run training.")
    parser.add_argument("--do_KD", action="store_true",
                        help="Whether to train with knowledge distillation.")
    parser.add_argument("--hard_label", action="store_true")
    parser.add_argument("--do_predict", action="store_true",
                        help="Whether to run predictions on the test set.")
    parser.add_argument("--evaluate_during_training", action="store_true",
                        help="Whether to 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=32, type=int,
                        help="Batch size per GPU/CPU for training.")
    parser.add_argument("--per_gpu_eval_batch_size", default=32, 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=2e-5, type=float,
                        help="The initial learning rate for Adam.")
    parser.add_argument("--weight_decay", default=0.01, 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("--freeze_bottom_layer", default=3, type=int,
                        help="Freeze the bottom n layers of the model during fine-tuning.")
    parser.add_argument("--num_train_epochs", default=10, type=int,
                        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("--warmup_ratio", default=0.1, type=float,
                        help="Linear warmup over warmup_ratio.")

    parser.add_argument("--logging_steps", type=int, default=20,
                        help="Log every X updates steps.")
    parser.add_argument("--save_steps", type=int, default=20000,
                        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=667,
                        help="random seed for initialization")
    parser.add_argument("--gpu_ids", type=int, nargs="+", default=0,
                        help="ids of the gpus to use")

    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.")
    args = parser.parse_args()

    # Check output_dir
    if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and os.path.exists(os.path.join(args.output_dir, "pytorch_model.bin")) and os.path.basename(args.output_dir) != "test":
        raise ValueError("Train: Output directory already exists and is not empty.")

    if os.path.exists(args.output_dir) and args.do_predict:
        is_done = False
        for name in os.listdir(args.output_dir): # result file: "test_results-TIME-LANGUAGE"
            if "test_results" in name and (os.path.basename(args.data_dir) + ".txt") in name:
                is_done = True
                break
        if is_done:
            raise ValueError("Predict: Output directory ({}) already exists and is not empty.".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:
        args.n_gpu = len(args.gpu_ids)  # torch.cuda.device_count()
        os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_ids[0])
        # os.environ['CUDA_VISIBLE_DEVICES'] = "0,1,2,3,4,5,6"
        device = torch.device("cuda")
        # device = torch.device("cpu") if (args.n_gpu == 0 or args.no_cuda) else torch.device(
        #     "cuda:{}".format(args.gpu_ids[0]))
    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
    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
            os.makedirs(args.output_dir)
    args.log_dir = os.path.join(args.output_dir, "logs")
    if not os.path.exists(args.log_dir):
        os.makedirs(args.log_dir)
    formatter = logging.Formatter('%(asctime)s %(levelname)s: - %(message)s', datefmt='%Y-%m-%d %H:%M:%S')
    log_name = "log-{}".format(time.strftime("%Y-%m-%d-%H-%M-%S", time.localtime()))
    if args.do_train:
        log_name += "-train"
    if args.do_predict:
        log_name += "-predict"
    log_name += "-{}".format("_".join(args.src_langs))
    log_name += "-{}.txt".format(os.path.basename(args.data_dir))
    fh = logging.FileHandler(os.path.join(args.log_dir, log_name))
    fh.setLevel(logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
    fh.setFormatter(formatter)
    ch = logging.StreamHandler()
    ch.setLevel(logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
    ch.setFormatter(formatter)
    logger.addHandler(fh)
    logger.addHandler(ch)

    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 CONLL-2003 task
    labels = get_labels(args.labels)
    num_labels = len(labels)
    # Use cross entropy ignore index as padding label id so that only real label ids contribute to the loss later
    pad_token_label_id = CrossEntropyLoss().ignore_index  # -100 here

    # 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()

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

    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]

    # Training
    if args.do_train:
        # load target model
        # config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, num_labels=num_labels)
        # 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)

        config = config_class.from_pretrained(args.config_name if args.config_name else args.bert_path, num_labels=num_labels, output_hidden_states=True)
        tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.bert_path, do_lower_case=args.do_lower_case)
        model = model_class.from_pretrained(args.bert_path, from_tf=bool(".ckpt" in args.bert_path), config=config)

        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)

        # prepare target training plain text
        train_dataset = load_and_cache_examples(args, tokenizer, args.data_dir, labels, pad_token_label_id, mode="train")
        

        if args.do_KD:
            logger.info("********** scheme: training with KD **********")

            # compute probs from source models
            w = 1.0 / len(args.src_langs)
            weight_probs = {l: w for l in args.src_langs}

            src_probs = None
            for lang in args.src_langs:
                src_probs = get_src_probs_layers(args, train_dataset, model_class, src_lang=lang)

            # Train!
            if args.hard_label:
                hard_labels = torch.argmax(src_probs, dim=-1, keepdim=False)
                train_dataset.tensors += (hard_labels,)

                global_step, tr_loss = train(args, model, train_dataset, tokenizer, labels, pad_token_label_id)
                logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
            else:
                global_step, tr_loss_KD, tr_loss = train_KD(args, model, train_dataset, src_probs, tokenizer, labels, pad_token_label_id)
                logger.info(" global_step = %s, average KD loss = %s, average loss = %s", global_step, tr_loss_KD, tr_loss)
        else:
            logger.info("********** scheme: training without KD **********")
            
            global_step, tr_loss = train(args, model, train_dataset, tokenizer, labels, pad_token_label_id)
            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)

    #     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"))


    if args.do_predict and args.local_rank in [-1, 0]:
        logger.info("********** scheme: prediction **********")
        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
        model = model_class.from_pretrained(args.output_dir)
        model.to(args.device)
        result, predictions = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="test")
        # Save results
        output_test_results_file = os.path.join(args.output_dir, "test_results-{}-{}.txt".format(
            time.strftime("%Y-%m-%d-%H-%M-%S", time.localtime()),
            os.path.basename(args.data_dir)))
        with open(output_test_results_file, "w", encoding='utf-8') as writer:
            for key in sorted(result.keys()):
                writer.write("{} = {}\n".format(key, str(result[key])))
        # Save predictions
        output_test_predictions_file = os.path.join(args.output_dir, "test_predictions-{}-{}.txt".format(
            time.strftime("%Y-%m-%d-%H-%M-%S", time.localtime()),
            os.path.basename(args.data_dir)))
        with open(output_test_predictions_file, "w", encoding='utf-8') as writer:
            with open(os.path.join(args.data_dir, "test.txt"), "r", encoding='utf-8') as f:
                example_id = 0
                for line in f:
                    if line.startswith("-DOCSTART-") or line == "" or line == "\n":
                        writer.write(line)
                        if not predictions[example_id]:
                            example_id += 1
                    elif predictions[example_id]:
                        output_line = line.split()[0] + " " + line.split()[-1].replace("\n", "") + " " + predictions[example_id].pop(0) + "\n"
                        writer.write(output_line)
                    else:
                        logger.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])


if __name__ == "__main__":
    main()