train_stage3.py

# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Finetuning the library models for question-answering on SQuAD (DistilBERT, Bert, XLM, XLNet)."""


import argparse
import glob
import logging
import os
import re
import collections
import random
import timeit
import json
import numpy as np
from datetime import datetime
import torch
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
from torch.utils.data.distributed import DistributedSampler
from tqdm import tqdm, trange
from multiprocessing import Pool, cpu_count
from functools import partial
from torch.utils.data import TensorDataset
from transformers import (WEIGHTS_NAME, AdamW, BertConfig, BertTokenizer, 
                        BertForQuestionAnswering, get_linear_schedule_with_warmup)

from transformers.data.metrics.squad_metrics import (
    compute_predictions_log_probs,
    compute_predictions_logits,
)
import string
from transformers.data.processors.utils import DataProcessor
from utils import readGZip

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


logger = logging.getLogger(__name__)

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

MODEL_CLASSES = {"bert": (BertConfig, BertForQuestionAnswering, BertTokenizer)}


def squad_convert_example_to_features_init(tokenizer_for_convert):
    global tokenizer
    tokenizer = tokenizer_for_convert


def _improve_answer_span(doc_tokens, input_start, input_end, tokenizer, orig_answer_text):
    """Returns tokenized answer spans that better match the annotated answer."""
    tok_answer_text = " ".join(tokenizer.tokenize(orig_answer_text))

    for new_start in range(input_start, input_end + 1):
        for new_end in range(input_end, new_start - 1, -1):
            text_span = " ".join(doc_tokens[new_start : (new_end + 1)])
            if text_span == tok_answer_text:
                return (new_start, new_end)

    return (input_start, input_end)

def _new_check_is_max_context(doc_spans, cur_span_index, position):
    """Check if this is the 'max context' doc span for the token."""
    # if len(doc_spans) == 1:
    # return True
    best_score = None
    best_span_index = None
    for (span_index, doc_span) in enumerate(doc_spans):
        end = doc_span["start"] + doc_span["length"] - 1
        if position < doc_span["start"]:
            continue
        if position > end:
            continue
        num_left_context = position - doc_span["start"]
        num_right_context = end - position
        score = min(num_left_context, num_right_context) + 0.01 * doc_span["length"]
        if best_score is None or score > best_score:
            best_score = score
            best_span_index = span_index

    return cur_span_index == best_span_index

class SquadFeatures(object):
    def __init__(
        self,
        input_ids,
        attention_mask,
        token_type_ids,
        cls_index,
        p_mask,
        example_index,
        unique_id,
        paragraph_len,
        token_is_max_context,
        tokens,
        token_to_orig_map,
        start_position,
        end_position,
        is_impossible,
    ):
        self.input_ids = input_ids
        self.attention_mask = attention_mask
        self.token_type_ids = token_type_ids
        self.cls_index = cls_index
        self.p_mask = p_mask

        self.example_index = example_index
        self.unique_id = unique_id
        self.paragraph_len = paragraph_len
        self.token_is_max_context = token_is_max_context
        self.tokens = tokens
        self.token_to_orig_map = token_to_orig_map

        self.start_position = start_position
        self.end_position = end_position
        self.is_impossible = is_impossible

class SquadResult(object):
    """
    Constructs a SquadResult which can be used to evaluate a model's output on the SQuAD dataset.
    Args:
        unique_id: The unique identifier corresponding to that example.
        start_logits: The logits corresponding to the start of the answer
        end_logits: The logits corresponding to the end of the answer
    """

    def __init__(self, unique_id, start_logits, end_logits, start_top_index=None, end_top_index=None, cls_logits=None):
        self.start_logits = start_logits
        self.end_logits = end_logits
        self.unique_id = unique_id

        if start_top_index:
            self.start_top_index = start_top_index
            self.end_top_index = end_top_index
            self.cls_logits = cls_logits


def squad_convert_example_to_features(example, max_seq_length, doc_stride, max_query_length, is_training):
    features = []
    if is_training and not example.is_impossible:
        # Get start and end position
        start_position = example.start_position
        end_position = example.end_position

        # If the answer cannot be found in the text, then skip this example.
        actual_text = " ".join(example.doc_tokens[start_position : (end_position + 1)])
        if actual_text.find(example.answer_text) == -1:
            logger.warning("Could not find answer: '%s' vs. '%s' in '%s'", actual_text, example.answer_text, example.qas_id)
            return []

    tok_to_orig_index = []
    orig_to_tok_index = []
    all_doc_tokens = []
    for (i, token) in enumerate(example.doc_tokens):
        orig_to_tok_index.append(len(all_doc_tokens))
        sub_tokens = tokenizer.tokenize(token)
        for sub_token in sub_tokens:
            tok_to_orig_index.append(i)
            all_doc_tokens.append(sub_token)

    if is_training and not example.is_impossible:
        tok_start_position = orig_to_tok_index[example.start_position]
        if example.end_position < len(example.doc_tokens) - 1:
            tok_end_position = orig_to_tok_index[example.end_position + 1] - 1
        else:
            tok_end_position = len(all_doc_tokens) - 1

        (tok_start_position, tok_end_position) = _improve_answer_span(
            all_doc_tokens, tok_start_position, tok_end_position, tokenizer, example.answer_text
        )

    spans = []

    truncated_query = tokenizer.encode(example.question_text, add_special_tokens=False, max_length=max_query_length)
    sequence_added_tokens = (
        tokenizer.max_len - tokenizer.max_len_single_sentence + 1
        if "roberta" in str(type(tokenizer)) or "camembert" in str(type(tokenizer))
        else tokenizer.max_len - tokenizer.max_len_single_sentence
    )
    sequence_pair_added_tokens = tokenizer.max_len - tokenizer.max_len_sentences_pair

    span_doc_tokens = all_doc_tokens
    while len(spans) * doc_stride < len(all_doc_tokens):
        encoded_dict = tokenizer.encode_plus(
            truncated_query if tokenizer.padding_side == "right" else span_doc_tokens,
            span_doc_tokens if tokenizer.padding_side == "right" else truncated_query,
            max_length=max_seq_length,
            return_overflowing_tokens=True,
            pad_to_max_length=True,
            stride=max_seq_length - doc_stride - len(truncated_query) - sequence_pair_added_tokens,
            truncation_strategy="only_second" if tokenizer.padding_side == "right" else "only_first",
        )

        paragraph_len = min(
            len(all_doc_tokens) - len(spans) * doc_stride,
            max_seq_length - len(truncated_query) - sequence_pair_added_tokens,
        )

        if tokenizer.pad_token_id in encoded_dict["input_ids"]:
            if tokenizer.padding_side == "right":
                non_padded_ids = encoded_dict["input_ids"][: encoded_dict["input_ids"].index(tokenizer.pad_token_id)]
            else:
                last_padding_id_position = (
                    len(encoded_dict["input_ids"]) - 1 - encoded_dict["input_ids"][::-1].index(tokenizer.pad_token_id)
                )
                non_padded_ids = encoded_dict["input_ids"][last_padding_id_position + 1 :]

        else:
            non_padded_ids = encoded_dict["input_ids"]

        tokens = tokenizer.convert_ids_to_tokens(non_padded_ids)

        token_to_orig_map = {}
        for i in range(paragraph_len):
            index = len(truncated_query) + sequence_added_tokens + i if tokenizer.padding_side == "right" else i
            token_to_orig_map[index] = tok_to_orig_index[len(spans) * doc_stride + i]

        encoded_dict["paragraph_len"] = paragraph_len
        encoded_dict["tokens"] = tokens
        encoded_dict["token_to_orig_map"] = token_to_orig_map
        encoded_dict["truncated_query_with_special_tokens_length"] = len(truncated_query) + sequence_added_tokens
        encoded_dict["token_is_max_context"] = {}
        encoded_dict["start"] = len(spans) * doc_stride
        encoded_dict["length"] = paragraph_len

        spans.append(encoded_dict)

        if "overflowing_tokens" not in encoded_dict:
            break
        span_doc_tokens = encoded_dict["overflowing_tokens"]

    for doc_span_index in range(len(spans)):
        for j in range(spans[doc_span_index]["paragraph_len"]):
            is_max_context = _new_check_is_max_context(spans, doc_span_index, doc_span_index * doc_stride + j)
            index = (
                j
                if tokenizer.padding_side == "left"
                else spans[doc_span_index]["truncated_query_with_special_tokens_length"] + j
            )
            spans[doc_span_index]["token_is_max_context"][index] = is_max_context

    for span in spans:
        # Identify the position of the CLS token
        cls_index = span["input_ids"].index(tokenizer.cls_token_id)

        # p_mask: mask with 1 for token than cannot be in the answer (0 for token which can be in an answer)
        # Original TF implem also keep the classification token (set to 0) (not sure why...)
        p_mask = np.array(span["token_type_ids"])

        p_mask = np.minimum(p_mask, 1)

        if tokenizer.padding_side == "right":
            # Limit positive values to one
            p_mask = 1 - p_mask

        p_mask[np.where(np.array(span["input_ids"]) == tokenizer.sep_token_id)[0]] = 1

        # Set the CLS index to '0'
        p_mask[cls_index] = 0

        span_is_impossible = example.is_impossible
        start_position = 0
        end_position = 0
        if is_training and not span_is_impossible:
            # For training, if our document chunk does not contain an annotation
            # we throw it out, since there is nothing to predict.
            doc_start = span["start"]
            doc_end = span["start"] + span["length"] - 1
            out_of_span = False

            if not (tok_start_position >= doc_start and tok_end_position <= doc_end):
                out_of_span = True

            if out_of_span:
                start_position = cls_index
                end_position = cls_index
                span_is_impossible = True
            else:
                if tokenizer.padding_side == "left":
                    doc_offset = 0
                else:
                    doc_offset = len(truncated_query) + sequence_added_tokens

                start_position = tok_start_position - doc_start + doc_offset
                end_position = tok_end_position - doc_start + doc_offset

        features.append(
            SquadFeatures(
                span["input_ids"],
                span["attention_mask"],
                span["token_type_ids"],
                cls_index,
                p_mask.tolist(),
                example_index=0,  # Can not set unique_id and example_index here. They will be set after multiple processing.
                unique_id=0,
                paragraph_len=span["paragraph_len"],
                token_is_max_context=span["token_is_max_context"],
                tokens=span["tokens"],
                token_to_orig_map=span["token_to_orig_map"],
                start_position=start_position,
                end_position=end_position,
                is_impossible=span_is_impossible,
            )
        )
    return features


def squad_convert_examples_to_features(examples, tokenizer, max_seq_length, doc_stride, max_query_length, is_training, threads=1):
    # Defining helper methods
    features = []
    threads = min(threads, cpu_count())
    with Pool(threads, initializer=squad_convert_example_to_features_init, initargs=(tokenizer,)) as p:
        annotate_ = partial(
            squad_convert_example_to_features,
            max_seq_length=max_seq_length,
            doc_stride=doc_stride,
            max_query_length=max_query_length,
            is_training=is_training,
        )
        features = list(
            tqdm(
                p.imap(annotate_, examples, chunksize=32),
                total=len(examples),
                desc="convert squad examples to features",
            )
        )

    new_features = []
    unique_id = 1000000000
    example_index = 0
    for example_features in tqdm(features, total=len(features), desc="add example index and unique id"):
        if not example_features:
            continue
        for example_feature in example_features:
            example_feature.example_index = example_index
            example_feature.unique_id = unique_id
            new_features.append(example_feature)
            unique_id += 1
        example_index += 1
    features = new_features
    del new_features

    # Convert to Tensors and build dataset
    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
    all_attention_masks = 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)
    all_cls_index = torch.tensor([f.cls_index for f in features], dtype=torch.long)
    all_p_mask = torch.tensor([f.p_mask for f in features], dtype=torch.float)
    all_is_impossible = torch.tensor([f.is_impossible for f in features], dtype=torch.float)

    if not is_training:
        all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
        dataset = TensorDataset(
            all_input_ids, all_attention_masks, all_token_type_ids, all_example_index, all_cls_index, all_p_mask
        )
    else:
        all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long)
        all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long)
        dataset = TensorDataset(
            all_input_ids,
            all_attention_masks,
            all_token_type_ids,
            all_start_positions,
            all_end_positions,
            all_cls_index,
            all_p_mask,
            all_is_impossible,
        )

    return features, dataset


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 to_list(tensor):
    return tensor.detach().cpu().tolist()


def train(args, train_dataset, model, tokenizer):
    """ 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)

    t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs

    # Prepare optimizer and schedule (linear warmup and decay)
    no_decay = ["bias", "LayerNorm.weight"]
    optimizer_grouped_parameters = [
        {
            "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
            "weight_decay": args.weight_decay,
        },
        {"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
    ]
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
    scheduler = get_linear_schedule_with_warmup(
        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
    )

    # 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 = torch.nn.DataParallel(model)

    # Distributed training (should be after apex fp16 initialization)
    if args.local_rank != -1:
        model = torch.nn.parallel.DistributedDataParallel(
            model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True
        )

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

    global_step = 1
    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):
        try:
            # set global_step to gobal_step of last saved checkpoint from model path
            checkpoint_suffix = args.model_name_or_path.split("-")[-1].split("/")[0]
            global_step = int(checkpoint_suffix)
            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)
        except ValueError:
            logger.info("  Starting fine-tuning.")

    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]
    )
    # Added here for reproductibility
    set_seed(args)

    for epoch 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)

            inputs = {
                "input_ids": batch[0],
                "attention_mask": batch[1],
                "token_type_ids": batch[2],
                "start_positions": batch[3],
                "end_positions": batch[4],
            }

            outputs = model(**inputs)
            # model outputs are always tuple in transformers (see doc)
            loss = outputs[0]

            if args.n_gpu > 1:
                loss = loss.mean()  # mean() to average on multi-gpu parallel (not distributed) 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)

                optimizer.step()
                scheduler.step()  # Update learning rate schedule
                model.zero_grad()
                global_step += 1

                # Log metrics
                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
                    tb_writer.add_scalar("stage3_lr", scheduler.get_last_lr()[0], global_step)
                    tb_writer.add_scalar("stage3_loss", (tr_loss - logging_loss) / args.logging_steps, global_step)
                    logging_loss = tr_loss
        
        # Save model checkpoint
        if args.local_rank in [-1, 0]:
            output_dir = os.path.join(args.output_dir, "checkpoint-epoch{}".format(epoch))
            if not os.path.exists(output_dir):
                os.makedirs(output_dir)
            # Take care of distributed/parallel training
            model_to_save = model.module if hasattr(model, "module") else model
            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.local_rank in [-1, 0]:
        tb_writer.close()

    return global_step, tr_loss / global_step

def evaluate_simplified(inputs, args, model, tokenizer, prefix=""):
    processor = SquadProcessor()
    examples = processor._create_examples(inputs, 'dev')

    #logger.info("Preprocessing {} examples".format(len(examples)))
    features, dataset = squad_convert_examples_to_features(
        examples=examples,
        tokenizer=tokenizer,
        max_seq_length=args.max_seq_length,
        doc_stride=args.doc_stride,
        max_query_length=args.max_query_length,
        is_training=False,
        threads=args.threads,
    )

    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)

    # Note that DistributedSampler samples randomly
    eval_sampler = SequentialSampler(dataset)
    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)

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

    # Eval!
    logger.info("***** Running evaluation {} *****".format(prefix))
    logger.info("  Num examples = %d", len(dataset))
    logger.info("  Batch size = %d", args.eval_batch_size)

    all_results = []
    start_time = timeit.default_timer()

    for batch in tqdm(eval_dataloader, desc="Evaluating"):
        model.eval()
        batch = tuple(t.to(args.device) for t in batch)

        with torch.no_grad():
            inputs = {
                "input_ids": batch[0],
                "attention_mask": batch[1],
                "token_type_ids": batch[2],
            }

            example_indices = batch[3]

            outputs = model(**inputs)

        for i, example_index in enumerate(example_indices):
            eval_feature = features[example_index.item()]
            unique_id = int(eval_feature.unique_id)

            output = [to_list(output[i]) for output in outputs]

            start_logits, end_logits = output
            result = SquadResult(unique_id, start_logits, end_logits)

            all_results.append(result)

    evalTime = timeit.default_timer() - start_time
    #logger.info("  Evaluation done in total %f secs (%f sec per example) for %d examples", evalTime, evalTime / len(dataset), len(all_results))

    # Compute predictions
    output_prediction_file = os.path.join('/tmp/', "predictions_{}.json".format(prefix))
    output_nbest_file = os.path.join('/tmp/', "nbest_predictions_{}.json".format(prefix))
    
    if args.version_2_with_negative:
        output_null_log_odds_file = os.path.join('/tmp/', "null_odds_{}.json".format(prefix))
    else:
        output_null_log_odds_file = None

    predictions = compute_predictions_logits(
        examples,
        features,
        all_results,
        args.n_best_size,
        args.max_answer_length,
        args.do_lower_case,
        output_prediction_file,
        output_nbest_file,
        output_null_log_odds_file,
        args.verbose_logging,
        args.version_2_with_negative,
        args.null_score_diff_threshold,
        tokenizer,
    )

    return predictions

def _is_whitespace(c):
    if c == " " or c == "\t" or c == "\r" or c == "\n" or ord(c) == 0x202F:
        return True
    return False

class SquadExample(object):
    def __init__(
        self,
        qas_id,
        question_text,
        context_text,
        answer_text,
        start_position_character,
        title,
        answers=[],
        is_impossible=False,
    ):
        self.qas_id = qas_id
        self.question_text = question_text
        self.context_text = context_text
        self.answer_text = answer_text
        self.title = title
        self.is_impossible = is_impossible
        self.answers = answers

        self.start_position, self.end_position = 0, 0

        doc_tokens = []
        char_to_word_offset = []
        prev_is_whitespace = True

        # Split on whitespace so that different tokens may be attributed to their original position.
        for c in self.context_text:
            if _is_whitespace(c):
                prev_is_whitespace = True
            else:
                if prev_is_whitespace:
                    doc_tokens.append(c)
                else:
                    doc_tokens[-1] += c
                prev_is_whitespace = False
            char_to_word_offset.append(len(doc_tokens) - 1)

        self.doc_tokens = doc_tokens
        self.char_to_word_offset = char_to_word_offset

        # Start and end positions only has a value during evaluation.
        if start_position_character is not None and not is_impossible:
            self.start_position = char_to_word_offset[start_position_character]
            self.end_position = char_to_word_offset[
                min(start_position_character + len(answer_text) - 1, len(char_to_word_offset) - 1)
            ]


class SquadProcessor(DataProcessor):
    def get_train_examples(self, filename=None):
        #with open(os.path.join(filename), "r", encoding="utf-8") as reader:
        #    input_data = json.load(reader)
        input_data = readGZip(filename)
        return self._create_examples(input_data, "train")

    def get_dev_examples(self, filename=None):
        input_data = readGZip(filename)
        return self._create_examples(input_data, "dev")

    def _create_examples(self, input_data, set_type):
        is_training = set_type == "train"
        examples = []
        for entry in tqdm(input_data):
            title = entry["title"]
            
            context_text = entry["context"]
            #for qa in paragraph["qas"]:
            
            qas_id = entry["question_id"]
            question_text = entry["question"]
            start_position_character = None
            answer_text = None
            answers = []

            if "is_impossible" in entry:
                is_impossible = entry["is_impossible"]
            else:
                is_impossible = False

            if not is_impossible:
                if is_training:
                    answer = entry["answers"][0]
                    answer_text = answer["text"]
                    start_position_character = answer["answer_start"]
                else:
                    answers = entry["answers"]

            example = SquadExample(
                qas_id=qas_id,
                question_text=question_text,
                context_text=context_text,
                answer_text=answer_text,
                start_position_character=start_position_character,
                title=title,
                is_impossible=is_impossible,
                answers=answers,
            )

            examples.append(example)
        
        return examples


def get_raw_scores(examples, preds):
    """
    Computes the exact and f1 scores from the examples and the model predictions
    """
    exact_scores = {}
    f1_scores = {}
    
    for example in examples:
        qas_id = example.qas_id
        gold_answers = [answer["text"] for answer in example.answers if normalize_answer(answer["text"])]

        if not gold_answers:
            # For unanswerable questions, only correct answer is empty string
            gold_answers = [""]

        if qas_id not in preds:
            print("Missing prediction for %s" % qas_id)
            continue

        prediction = preds[qas_id]
        exact_scores[qas_id] = max(compute_exact(a, prediction) for a in gold_answers)
        f1_scores[qas_id] = max(compute_f1(a, prediction) for a in gold_answers)

    qid_list = [k for k in exact_scores]
    total = len(qid_list)
    
    return collections.OrderedDict(
        [
            ("exact", 100.0 * sum(exact_scores[k] for k in qid_list) / total),
            ("f1", 100.0 * sum(f1_scores[k] for k in qid_list) / total),
            ("total", total),
        ]
    )


def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False, cache=True):
    # Load data features from cache or dataset file
    processor = SquadProcessor()
    if evaluate:
        examples = processor.get_dev_examples(args.predict_file)
    else:
        examples = processor.get_train_examples(args.train_file)

    features, dataset = squad_convert_examples_to_features(
        examples=examples,
        tokenizer=tokenizer,
        max_seq_length=args.max_seq_length,
        doc_stride=args.doc_stride,
        max_query_length=args.max_query_length,
        is_training=not evaluate,
        threads=args.threads,
    )

    if output_examples:
        return dataset, examples, features

    return dataset


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

    # Required parameters
    parser.add_argument(
        "--model_type",
        default='bert',
        type=str,
        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()),
    )
    parser.add_argument(
        "--model_name_or_path",
        default="bert-base-uncased",
        type=str,
        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS),
    )
    parser.add_argument(
        "--output_dir",
        default='stage3',
        type=str,
        help="The output directory where the model checkpoints and predictions will be written.",
    )
    parser.add_argument(
        "--train_file",
        default=None,
        type=str,
        help="The input training file. If a data dir is specified, will look for the file there"
        + "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
    )
    parser.add_argument(
        "--resource_dir",
        type=str,
        default='data/',
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )   
    parser.add_argument(
        "--predict_file",
        default=None,
        type=str,
        help="The input evaluation file. If a data dir is specified, will look for the file there"
        + "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
    )
    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="/tmp/",
        type=str,
        help="Where do you want to store the pre-trained models downloaded from s3",
    )
    parser.add_argument(
        "--version_2_with_negative",
        action="store_true",
        help="If true, the SQuAD examples contain some that do not have an answer.",
    )
    parser.add_argument(
        "--null_score_diff_threshold",
        type=float,
        default=0.0,
        help="If null_score - best_non_null is greater than the threshold predict null.",
    )
    parser.add_argument(
        "--max_seq_length",
        default=384,
        type=int,
        help="The maximum total input sequence length after WordPiece tokenization. Sequences "
        "longer than this will be truncated, and sequences shorter than this will be padded.",
    )
    parser.add_argument(
        "--doc_stride",
        default=128,
        type=int,
        help="When splitting up a long document into chunks, how much stride to take between chunks.",
    )
    parser.add_argument(
        "--max_query_length",
        default=64,
        type=int,
        help="The maximum number of tokens for the question. Questions longer than this will "
        "be truncated to this length.",
    )
    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("--do_stage3", action="store_true", help="Whether to run eval on the dev set.")    
    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.")
    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("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
    parser.add_argument(
        "--n_best_size",
        default=20,
        type=int,
        help="The total number of n-best predictions to generate in the nbest_predictions.json output file.",
    )
    parser.add_argument(
        "--max_answer_length",
        default=30,
        type=int,
        help="The maximum length of an answer that can be generated. This is needed because the start "
        "and end predictions are not conditioned on one another.",
    )
    parser.add_argument(
        "--verbose_logging",
        action="store_true",
        help="If true, all of the warnings related to data processing will be printed. "
        "A number of warnings are expected for a normal SQuAD evaluation.",
    )
    parser.add_argument(
        "--lang_id",
        default=0,
        type=int,
        help="language id of input for language-specific xlm models (see tokenization_xlm.PRETRAINED_INIT_CONFIGURATION)",
    )
    parser.add_argument("--request_path", type=str, default='request_tok', help="Request directory.")
    parser.add_argument("--logging_steps", type=int, default=500, help="Log every X updates steps.")
    parser.add_argument("--save_steps", type=int, default=500, help="Save checkpoint every X updates steps.")
    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("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus")
    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("--threads", type=int, default=1, help="multiple threads for converting example to features")
    args = parser.parse_args()

    if args.doc_stride >= args.max_seq_length - args.max_query_length:
        logger.warning(
            "WARNING - You've set a doc stride which may be superior to the document length in some "
            "examples. This could result in errors when building features from the examples. Please reduce the doc "
            "stride or increase the maximum length to ensure the features are correctly built."
        )


    assert(args.local_rank == -1)
    # Setup CUDA, GPU & distributed training
    device = torch.device("cuda")
    args.n_gpu = torch.cuda.device_count()

    args.device = device
    args.output_dir = os.path.join(args.output_dir, datetime.now().strftime('%Y_%m_%d_%H_%M_%S'))
    # 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,
    )
    # Set seed
    set_seed(args)

    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,
        cache_dir=args.cache_dir if args.cache_dir else None,
    )
    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,
        cache_dir=args.cache_dir if args.cache_dir else None,
    )
    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,
    )

    model.to(args.device)

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

    # Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
    # Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
    # remove the need for this code, but it is still valid.
    if args.fp16:
        try:
            import apex
            apex.amp.register_half_function(torch, "einsum")
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")

    # Training
    if args.do_train:
        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False)
        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)

    # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory    
    if args.do_eval and args.local_rank in [-1, 0]:
        results = {}
        logger.info("Loading checkpoint %s for evaluation", args.model_name_or_path)
        checkpoints = []

        logger.info("Evaluate the following checkpoints: %s", args.model_name_or_path)

        model = model_class.from_pretrained(args.model_name_or_path)
        model.to(args.device)

        # Evaluate
        with open(args.predict_file, 'r') as f:
            full_split = json.load(f)
        
        key2idx = {}
        for step, d in enumerate(full_split):
            key2idx[d['question_id']] = step            

        prediction = evaluate_simplified(full_split, args, model, tokenizer)
        for k, step in key2idx.items():
            full_split[step]['pred'] = prediction.get(k, 'None')

        with open('passage_only_predictions.json', 'w') as f:
            json.dump(full_split, f, indent=2)

    if args.do_stage3 and args.local_rank in [-1, 0]:
        logger.info("Loading checkpoint %s for evaluation", args.model_name_or_path)
        model = model_class.from_pretrained(args.model_name_or_path)
        model.to(args.device)
        
        with open(args.request_path, 'r') as f:
            requests = json.load(f)

        #evaluate(args, model, tokenizer, prefix=global_step)
        with open(args.predict_file, 'r') as f:
            data = json.load(f)

        full_split = []
        key2idx = {}
        for step, d in enumerate(data):
            if isinstance(d['pred'], str):
                continue

            table_id = d['table_id']
            
            node = d['pred']

            context = 'Title : {} . {}'.format(node[0], requests[node[1]])
            
            full_split.append({'context': context, 'title': table_id, 
                              'question': d['question'], 'question_id': d['question_id'],
                              'answers': [{'answer_start': None, 'text': None}]})

            key2idx[d['question_id']] = step

        prediction = evaluate_simplified(full_split, args, model, tokenizer)
                
        for k, step in key2idx.items():
            data[step]['pred'] = prediction.get(k, 'None')

        #for _ in data:
        #    assert isinstance(_['pred'], str), "there are some unprocessed stage3 examples"

        with open('predictions.json', 'w') as f:
            json.dump(data, f, indent=2)


if __name__ == "__main__":
    main()