Cleanup

Improve global visibility on the run_squad script, remove unused files and fixes related to XLNet.

examples/run_squad.py

@@ -27,8 +27,7 @@ import glob
 import timeit
 import numpy as np
 import torch
-from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
-                              TensorDataset)
+from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler, TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
 
 try:
@@ -48,14 +47,6 @@ from transformers import (WEIGHTS_NAME, BertConfig,
 
 from transformers import AdamW, get_linear_schedule_with_warmup, squad_convert_examples_to_features
 
-from utils_squad import (convert_examples_to_features as old_convert, read_squad_examples as old_read, RawResult, write_predictions,
-                         RawResultExtended, write_predictions_extended)
-
-# The follwing import is the official SQuAD evaluation script (2.0).
-# You can remove it from the dependencies if you are using this script outside of the library
-# We've added it here for automated tests (see examples/test_examples.py file)
-from utils_squad_evaluate import EVAL_OPTS, main as evaluate_on_squad
-
 logger = logging.getLogger(__name__)
 
 ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) \
@@ -98,14 +89,16 @@ def train(args, train_dataset, model, tokenizer):
     optimizer_grouped_parameters = [
         {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
         {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
-        ]
+    ]
     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
     scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
+
     if args.fp16:
         try:
             from apex import amp
         except ImportError:
             raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+        
         model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
 
     # multi-gpu training (should be after apex fp16 initialization)
@@ -133,20 +126,26 @@ def train(args, train_dataset, model, tokenizer):
     model.zero_grad()
     train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
     set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
+    
     for _ in train_iterator:
         epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
         for step, batch in enumerate(epoch_iterator):
             model.train()
             batch = tuple(t.to(args.device) for t in batch)
-            inputs = {'input_ids':       batch[0],
-                      'attention_mask':  batch[1],
-                      'start_positions': batch[3],
-                      'end_positions':   batch[4]}
+
+            inputs = {
+                'input_ids':       batch[0],
+                'attention_mask':  batch[1],
+                'start_positions': batch[3],
+                'end_positions':   batch[4]
+            }
+
             if args.model_type != 'distilbert':
                 inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
+
             if args.model_type in ['xlnet', 'xlm']:
-                inputs.update({'cls_index': batch[5],
-                               'p_mask':       batch[6]})
+                inputs.update({'cls_index': batch[5], 'p_mask': batch[6]})
+
             outputs = model(**inputs)
             loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
 
@@ -173,8 +172,8 @@ def train(args, train_dataset, model, tokenizer):
                 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:
-                    # Log metrics
                     if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
                         results = evaluate(args, model, tokenizer)
                         for key, value in results.items():
@@ -183,8 +182,8 @@ def train(args, train_dataset, model, tokenizer):
                     tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
                     logging_loss = tr_loss
 
+                # Save model checkpoint
                 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)
@@ -213,6 +212,7 @@ def evaluate(args, model, tokenizer, prefix=""):
         os.makedirs(args.output_dir)
 
     args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+
     # Note that DistributedSampler samples randomly
     eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset)
     eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
@@ -225,11 +225,14 @@ def evaluate(args, model, tokenizer, prefix=""):
     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],
@@ -238,10 +241,13 @@ def evaluate(args, model, tokenizer, prefix=""):
             
             if args.model_type != 'distilbert':
                 inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]  # XLM don't use segment_ids
+
             example_indices = batch[3]
+            
+            # XLNet and XLM use more arguments for their predictions
             if args.model_type in ['xlnet', 'xlm']:
-                inputs.update({'cls_index': batch[4],
-                               'p_mask':    batch[5]})
+                inputs.update({'cls_index': batch[4], 'p_mask': batch[5]})
+
             outputs = model(**inputs)
 
         for i, example_index in enumerate(example_indices):
@@ -250,11 +256,13 @@ def evaluate(args, model, tokenizer, prefix=""):
 
             output = [to_list(output[i]) for output in outputs]
 
+            # Some models (XLNet, XLM) use 5 arguments for their predictions, while the other "simpler"
+            # models only use two.
             if len(output) >= 5:
                 start_logits = output[0]
                 start_top_index = output[1]
                 end_logits = output[2]
-                end_top_index = output[3],
+                end_top_index = output[3]
                 cls_logits = output[4]
 
                 result = SquadResult(
@@ -278,16 +286,17 @@ def evaluate(args, model, tokenizer, prefix=""):
     # Compute predictions
     output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
     output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
+
     if args.version_2_with_negative:
         output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
     else:
         output_null_log_odds_file = None
 
+    # XLNet and XLM use a more complex post-processing procedure
     if args.model_type in ['xlnet', 'xlm']:
-        # XLNet uses a more complex post-processing procedure
         predictions = compute_predictions_log_probs(examples, features, all_results, args.n_best_size,
                         args.max_answer_length, output_prediction_file,
-                        output_nbest_file, output_null_log_odds_file, args.predict_file,
+                        output_nbest_file, output_null_log_odds_file,
                         model.config.start_n_top, model.config.end_n_top,
                         args.version_2_with_negative, tokenizer, args.verbose_logging)
     else:
@@ -296,6 +305,7 @@ def evaluate(args, model, tokenizer, prefix=""):
                         output_nbest_file, output_null_log_odds_file, args.verbose_logging,
                         args.version_2_with_negative, args.null_score_diff_threshold)
 
+    # Compute the F1 and exact scores.
     results = squad_evaluate(examples, predictions)
     return results
 
@@ -308,7 +318,10 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
     cached_features_file = os.path.join(input_dir, 'cached_{}_{}_{}'.format(
         'dev' if evaluate else 'train',
         list(filter(None, args.model_name_or_path.split('/'))).pop(),
-        str(args.max_seq_length)))
+        str(args.max_seq_length))
+    )
+
+    # Init features and dataset from cache if it exists
     if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
         logger.info("Loading features from cached file %s", cached_features_file)
         features_and_dataset = torch.load(cached_features_file)
@@ -341,7 +354,6 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
             return_dataset='pt'
         )
 
-
         if args.local_rank in [-1, 0]:
             logger.info("Saving features into cached file %s", cached_features_file)
             torch.save({"features": features, "dataset": dataset}, cached_features_file)
@@ -452,6 +464,11 @@ def main():
     parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
     args = parser.parse_args()
 
+    args.predict_file = os.path.join(args.output_dir, 'predictions_{}_{}.txt'.format(
+        list(filter(None, args.model_name_or_path.split('/'))).pop(),
+        str(args.max_seq_length))
+    )
+
     if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
         raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
 

examples/utils_squad.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.
-""" Load SQuAD dataset. """
-
-from __future__ import absolute_import, division, print_function
-
-import json
-import logging
-import math
-import collections
-from io import open
-from tqdm import tqdm
-
-from transformers.tokenization_bert import BasicTokenizer, whitespace_tokenize
-
-# Required by XLNet evaluation method to compute optimal threshold (see write_predictions_extended() method)
-from utils_squad_evaluate import find_all_best_thresh_v2, make_qid_to_has_ans, get_raw_scores
-
-logger = logging.getLogger(__name__)
-
-
-class SquadExample(object):
-    """
-    A single training/test example for the Squad dataset.
-    For examples without an answer, the start and end position are -1.
-    """
-
-    def __init__(self,
-                 qas_id,
-                 question_text,
-                 doc_tokens,
-                 orig_answer_text=None,
-                 start_position=None,
-                 end_position=None,
-                 is_impossible=None):
-        self.qas_id = qas_id
-        self.question_text = question_text
-        self.doc_tokens = doc_tokens
-        self.orig_answer_text = orig_answer_text
-        self.start_position = start_position
-        self.end_position = end_position
-        self.is_impossible = is_impossible
-
-    def __str__(self):
-        return self.__repr__()
-
-    def __repr__(self):
-        s = ""
-        s += "qas_id: %s" % (self.qas_id)
-        s += ", question_text: %s" % (
-            self.question_text)
-        s += ", doc_tokens: [%s]" % (" ".join(self.doc_tokens))
-        if self.start_position:
-            s += ", start_position: %d" % (self.start_position)
-        if self.end_position:
-            s += ", end_position: %d" % (self.end_position)
-        if self.is_impossible:
-            s += ", is_impossible: %r" % (self.is_impossible)
-        return s
-
-
-class InputFeatures(object):
-    """A single set of features of data."""
-
-    def __init__(self,
-                 unique_id,
-                 example_index,
-                 doc_span_index,
-                 tokens,
-                 token_to_orig_map,
-                 token_is_max_context,
-                 input_ids,
-                 input_mask,
-                 segment_ids,
-                 cls_index,
-                 p_mask,
-                 paragraph_len,
-                 start_position=None,
-                 end_position=None,
-                 is_impossible=None):
-        self.unique_id = unique_id
-        self.example_index = example_index
-        self.doc_span_index = doc_span_index
-        self.tokens = tokens
-        self.token_to_orig_map = token_to_orig_map
-        self.token_is_max_context = token_is_max_context
-        self.input_ids = input_ids
-        self.input_mask = input_mask
-        self.segment_ids = segment_ids
-        self.cls_index = cls_index
-        self.p_mask = p_mask
-        self.paragraph_len = paragraph_len
-        self.start_position = start_position
-        self.end_position = end_position
-        self.is_impossible = is_impossible
-
-
-def read_squad_examples(input_file, is_training, version_2_with_negative):
-    """Read a SQuAD json file into a list of SquadExample."""
-    with open(input_file, "r", encoding='utf-8') as reader:
-        input_data = json.load(reader)["data"]
-
-    def is_whitespace(c):
-        if c == " " or c == "\t" or c == "\r" or c == "\n" or ord(c) == 0x202F:
-            return True
-        return False
-
-    examples = []
-    for entry in input_data:
-        for paragraph in entry["paragraphs"]:
-            paragraph_text = paragraph["context"]
-            doc_tokens = []
-            char_to_word_offset = []
-            prev_is_whitespace = True
-            for c in paragraph_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)
-
-            for qa in paragraph["qas"]:
-                qas_id = qa["id"]
-                question_text = qa["question"]
-                start_position = None
-                end_position = None
-                orig_answer_text = None
-                is_impossible = False
-                if is_training:
-                    if version_2_with_negative:
-                        is_impossible = qa["is_impossible"]
-                    if (len(qa["answers"]) != 1) and (not is_impossible):
-                        raise ValueError(
-                            "For training, each question should have exactly 1 answer.")
-                    if not is_impossible:
-                        answer = qa["answers"][0]
-                        orig_answer_text = answer["text"]
-                        answer_offset = answer["answer_start"]
-                        answer_length = len(orig_answer_text)
-                        start_position = char_to_word_offset[answer_offset]
-                        end_position = char_to_word_offset[answer_offset + answer_length - 1]
-                        # Only add answers where the text can be exactly recovered from the
-                        # document. If this CAN'T happen it's likely due to weird Unicode
-                        # stuff so we will just skip the example.
-                        #
-                        # Note that this means for training mode, every example is NOT
-                        # guaranteed to be preserved.
-                        actual_text = " ".join(doc_tokens[start_position:(end_position + 1)])
-                        cleaned_answer_text = " ".join(
-                            whitespace_tokenize(orig_answer_text))
-                        if actual_text.find(cleaned_answer_text) == -1:
-                            logger.warning("Could not find answer: '%s' vs. '%s'",
-                                           actual_text, cleaned_answer_text)
-                            continue
-                    else:
-                        start_position = -1
-                        end_position = -1
-                        orig_answer_text = ""
-
-                example = SquadExample(
-                    qas_id=qas_id,
-                    question_text=question_text,
-                    doc_tokens=doc_tokens,
-                    orig_answer_text=orig_answer_text,
-                    start_position=start_position,
-                    end_position=end_position,
-                    is_impossible=is_impossible)
-                examples.append(example)
-    return examples
-
-
-def convert_examples_to_features(examples, tokenizer, max_seq_length,
-                                 doc_stride, max_query_length, is_training,
-                                 cls_token_at_end=False,
-                                 cls_token='[CLS]', sep_token='[SEP]', pad_token=0,
-                                 sequence_a_segment_id=0, sequence_b_segment_id=1,
-                                 cls_token_segment_id=0, pad_token_segment_id=0,
-                                 mask_padding_with_zero=True,
-                                 sequence_a_is_doc=False):
-    """Loads a data file into a list of `InputBatch`s."""
-
-    unique_id = 1000000000
-    # cnt_pos, cnt_neg = 0, 0
-    # max_N, max_M = 1024, 1024
-    # f = np.zeros((max_N, max_M), dtype=np.float32)
-
-    features = []
-    for (example_index, example) in enumerate(tqdm(examples)):
-
-        # if example_index % 100 == 0:
-        #     logger.info('Converting %s/%s pos %s neg %s', example_index, len(examples), cnt_pos, cnt_neg)
-
-        query_tokens = tokenizer.tokenize(example.question_text)
-
-        if len(query_tokens) > max_query_length:
-            query_tokens = query_tokens[0:max_query_length]
-
-        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)
-
-        tok_start_position = None
-        tok_end_position = None
-        if is_training and example.is_impossible:
-            tok_start_position = -1
-            tok_end_position = -1
-        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.orig_answer_text)
-
-        # The -3 accounts for [CLS], [SEP] and [SEP]
-        max_tokens_for_doc = max_seq_length - len(query_tokens) - 3
-        assert max_tokens_for_doc > 0
-
-        # We can have documents that are longer than the maximum sequence length.
-        # To deal with this we do a sliding window approach, where we take chunks
-        # of the up to our max length with a stride of `doc_stride`.
-        _DocSpan = collections.namedtuple(  # pylint: disable=invalid-name
-            "DocSpan", ["start", "length"])
-        doc_spans = []
-        start_offset = 0
-        while start_offset < len(all_doc_tokens):
-            length = len(all_doc_tokens) - start_offset
-            if length > max_tokens_for_doc:
-                length = max_tokens_for_doc
-            doc_spans.append(_DocSpan(start=start_offset, length=length))
-            if start_offset + length == len(all_doc_tokens):
-                break
-            start_offset += min(length, doc_stride)
-
-        for (doc_span_index, doc_span) in enumerate(doc_spans):
-            tokens = []
-            token_to_orig_map = {}
-            token_is_max_context = {}
-            segment_ids = []
-
-            # 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 = []
-
-            # CLS token at the beginning
-            if not cls_token_at_end:
-                tokens.append(cls_token)
-                segment_ids.append(cls_token_segment_id)
-                p_mask.append(0)
-                cls_index = 0
-
-            # XLNet: P SEP Q SEP CLS
-            # Others: CLS Q SEP P SEP
-            if not sequence_a_is_doc:
-                # Query
-                tokens += query_tokens
-                segment_ids += [sequence_a_segment_id] * len(query_tokens)
-                p_mask += [1] * len(query_tokens)
-
-                # SEP token
-                tokens.append(sep_token)
-                segment_ids.append(sequence_a_segment_id)
-                p_mask.append(1)
-
-            # Paragraph
-            for i in range(doc_span.length):
-                split_token_index = doc_span.start + i
-                token_to_orig_map[len(tokens)] = tok_to_orig_index[split_token_index]
-
-                is_max_context = _check_is_max_context(doc_spans, doc_span_index,
-                                                       split_token_index)
-                token_is_max_context[len(tokens)] = is_max_context
-                tokens.append(all_doc_tokens[split_token_index])
-                if not sequence_a_is_doc:
-                    segment_ids.append(sequence_b_segment_id)
-                else:
-                    segment_ids.append(sequence_a_segment_id)
-                p_mask.append(0)
-            paragraph_len = doc_span.length
-
-            if sequence_a_is_doc:
-                # SEP token
-                tokens.append(sep_token)
-                segment_ids.append(sequence_a_segment_id)
-                p_mask.append(1)
-
-                tokens += query_tokens
-                segment_ids += [sequence_b_segment_id] * len(query_tokens)
-                p_mask += [1] * len(query_tokens)
-
-            # SEP token
-            tokens.append(sep_token)
-            segment_ids.append(sequence_b_segment_id)
-            p_mask.append(1)
-
-            # CLS token at the end
-            if cls_token_at_end:
-                tokens.append(cls_token)
-                segment_ids.append(cls_token_segment_id)
-                p_mask.append(0)
-                cls_index = len(tokens) - 1  # Index of classification token
-
-            input_ids = tokenizer.convert_tokens_to_ids(tokens)
-
-            # The mask has 1 for real tokens and 0 for padding tokens. Only real
-            # tokens are attended to.
-            input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
-
-            # Zero-pad up to the sequence length.
-            while len(input_ids) < max_seq_length:
-                input_ids.append(pad_token)
-                input_mask.append(0 if mask_padding_with_zero else 1)
-                segment_ids.append(pad_token_segment_id)
-                p_mask.append(1)
-
-            assert len(input_ids) == max_seq_length
-            assert len(input_mask) == max_seq_length
-            assert len(segment_ids) == max_seq_length
-
-            span_is_impossible = example.is_impossible
-            start_position = None
-            end_position = None
-            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 = doc_span.start
-                doc_end = doc_span.start + doc_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 = 0
-                    end_position = 0
-                    span_is_impossible = True
-                else:
-                    if sequence_a_is_doc:
-                        doc_offset = 0
-                    else:
-                        doc_offset = len(query_tokens) + 2
-                    start_position = tok_start_position - doc_start + doc_offset
-                    end_position = tok_end_position - doc_start + doc_offset
-
-            if is_training and span_is_impossible:
-                start_position = cls_index
-                end_position = cls_index
-
-            if example_index < 20:
-                logger.info("*** Example ***")
-                logger.info("unique_id: %s" % (unique_id))
-                logger.info("example_index: %s" % (example_index))
-                logger.info("doc_span_index: %s" % (doc_span_index))
-                logger.info("tokens: %s" % " ".join(tokens))
-                logger.info("token_to_orig_map: %s" % " ".join([
-                    "%d:%d" % (x, y) for (x, y) in token_to_orig_map.items()]))
-                logger.info("token_is_max_context: %s" % " ".join([
-                    "%d:%s" % (x, y) for (x, y) in token_is_max_context.items()
-                ]))
-                logger.info("input_ids: %s" % " ".join([str(x) for x in input_ids]))
-                logger.info(
-                    "input_mask: %s" % " ".join([str(x) for x in input_mask]))
-                logger.info(
-                    "segment_ids: %s" % " ".join([str(x) for x in segment_ids]))
-                if is_training and span_is_impossible:
-                    logger.info("impossible example")
-                if is_training and not span_is_impossible:
-                    answer_text = " ".join(tokens[start_position:(end_position + 1)])
-                    logger.info("start_position: %d" % (start_position))
-                    logger.info("end_position: %d" % (end_position))
-                    logger.info(
-                        "answer: %s" % (answer_text))
-
-            features.append(
-                InputFeatures(
-                    unique_id=unique_id,
-                    example_index=example_index,
-                    doc_span_index=doc_span_index,
-                    tokens=tokens,
-                    token_to_orig_map=token_to_orig_map,
-                    token_is_max_context=token_is_max_context,
-                    input_ids=input_ids,
-                    input_mask=input_mask,
-                    segment_ids=segment_ids,
-                    cls_index=cls_index,
-                    p_mask=p_mask,
-                    paragraph_len=paragraph_len,
-                    start_position=start_position,
-                    end_position=end_position,
-                    is_impossible=span_is_impossible))
-            unique_id += 1
-
-    return features
-
-
-def _improve_answer_span(doc_tokens, input_start, input_end, tokenizer,
-                         orig_answer_text):
-    """Returns tokenized answer spans that better match the annotated answer."""
-
-    # The SQuAD annotations are character based. We first project them to
-    # whitespace-tokenized words. But then after WordPiece tokenization, we can
-    # often find a "better match". For example:
-    #
-    #   Question: What year was John Smith born?
-    #   Context: The leader was John Smith (1895-1943).
-    #   Answer: 1895
-    #
-    # The original whitespace-tokenized answer will be "(1895-1943).". However
-    # after tokenization, our tokens will be "( 1895 - 1943 ) .". So we can match
-    # the exact answer, 1895.
-    #
-    # However, this is not always possible. Consider the following:
-    #
-    #   Question: What country is the top exporter of electornics?
-    #   Context: The Japanese electronics industry is the lagest in the world.
-    #   Answer: Japan
-    #
-    # In this case, the annotator chose "Japan" as a character sub-span of
-    # the word "Japanese". Since our WordPiece tokenizer does not split
-    # "Japanese", we just use "Japanese" as the annotation. This is fairly rare
-    # in SQuAD, but does happen.
-    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 _check_is_max_context(doc_spans, cur_span_index, position):
-    """Check if this is the 'max context' doc span for the token."""
-
-    # Because of the sliding window approach taken to scoring documents, a single
-    # token can appear in multiple documents. E.g.
-    #  Doc: the man went to the store and bought a gallon of milk
-    #  Span A: the man went to the
-    #  Span B: to the store and bought
-    #  Span C: and bought a gallon of
-    #  ...
-    #
-    # Now the word 'bought' will have two scores from spans B and C. We only
-    # want to consider the score with "maximum context", which we define as
-    # the *minimum* of its left and right context (the *sum* of left and
-    # right context will always be the same, of course).
-    #
-    # In the example the maximum context for 'bought' would be span C since
-    # it has 1 left context and 3 right context, while span B has 4 left context
-    # and 0 right context.
-    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
-
-
-RawResult = collections.namedtuple("RawResult",
-                                   ["unique_id", "start_logits", "end_logits"])
-
-def write_predictions(all_examples, all_features, all_results, n_best_size,
-                      max_answer_length, do_lower_case, output_prediction_file,
-                      output_nbest_file, output_null_log_odds_file, verbose_logging,
-                      version_2_with_negative, null_score_diff_threshold):
-    """Write final predictions to the json file and log-odds of null if needed."""
-    logger.info("Writing predictions to: %s" % (output_prediction_file))
-    logger.info("Writing nbest to: %s" % (output_nbest_file))
-
-    example_index_to_features = collections.defaultdict(list)
-    for feature in all_features:
-        example_index_to_features[feature.example_index].append(feature)
-
-    unique_id_to_result = {}
-    for result in all_results:
-        unique_id_to_result[result.unique_id] = result
-
-    _PrelimPrediction = collections.namedtuple(  # pylint: disable=invalid-name
-        "PrelimPrediction",
-        ["feature_index", "start_index", "end_index", "start_logit", "end_logit"])
-
-    all_predictions = collections.OrderedDict()
-    all_nbest_json = collections.OrderedDict()
-    scores_diff_json = collections.OrderedDict()
-
-    for (example_index, example) in enumerate(all_examples):
-        features = example_index_to_features[example_index]
-
-        prelim_predictions = []
-        # keep track of the minimum score of null start+end of position 0
-        score_null = 1000000  # large and positive
-        min_null_feature_index = 0  # the paragraph slice with min null score
-        null_start_logit = 0  # the start logit at the slice with min null score
-        null_end_logit = 0  # the end logit at the slice with min null score
-        for (feature_index, feature) in enumerate(features):
-            result = unique_id_to_result[feature.unique_id]
-            start_indexes = _get_best_indexes(result.start_logits, n_best_size)
-            end_indexes = _get_best_indexes(result.end_logits, n_best_size)
-            # if we could have irrelevant answers, get the min score of irrelevant
-            if version_2_with_negative:
-                feature_null_score = result.start_logits[0] + result.end_logits[0]
-                if feature_null_score < score_null:
-                    score_null = feature_null_score
-                    min_null_feature_index = feature_index
-                    null_start_logit = result.start_logits[0]
-                    null_end_logit = result.end_logits[0]
-            for start_index in start_indexes:
-                for end_index in end_indexes:
-                    # We could hypothetically create invalid predictions, e.g., predict
-                    # that the start of the span is in the question. We throw out all
-                    # invalid predictions.
-                    if start_index >= len(feature.tokens):
-                        continue
-                    if end_index >= len(feature.tokens):
-                        continue
-                    if start_index not in feature.token_to_orig_map:
-                        continue
-                    if end_index not in feature.token_to_orig_map:
-                        continue
-                    if not feature.token_is_max_context.get(start_index, False):
-                        continue
-                    if end_index < start_index:
-                        continue
-                    length = end_index - start_index + 1
-                    if length > max_answer_length:
-                        continue
-                    prelim_predictions.append(
-                        _PrelimPrediction(
-                            feature_index=feature_index,
-                            start_index=start_index,
-                            end_index=end_index,
-                            start_logit=result.start_logits[start_index],
-                            end_logit=result.end_logits[end_index]))
-        if version_2_with_negative:
-            prelim_predictions.append(
-                _PrelimPrediction(
-                    feature_index=min_null_feature_index,
-                    start_index=0,
-                    end_index=0,
-                    start_logit=null_start_logit,
-                    end_logit=null_end_logit))
-        prelim_predictions = sorted(
-            prelim_predictions,
-            key=lambda x: (x.start_logit + x.end_logit),
-            reverse=True)
-
-        _NbestPrediction = collections.namedtuple(  # pylint: disable=invalid-name
-            "NbestPrediction", ["text", "start_logit", "end_logit"])
-
-        seen_predictions = {}
-        nbest = []
-        for pred in prelim_predictions:
-            if len(nbest) >= n_best_size:
-                break
-            feature = features[pred.feature_index]
-            if pred.start_index > 0:  # this is a non-null prediction
-                tok_tokens = feature.tokens[pred.start_index:(pred.end_index + 1)]
-                orig_doc_start = feature.token_to_orig_map[pred.start_index]
-                orig_doc_end = feature.token_to_orig_map[pred.end_index]
-                orig_tokens = example.doc_tokens[orig_doc_start:(orig_doc_end + 1)]
-                tok_text = " ".join(tok_tokens)
-
-                # De-tokenize WordPieces that have been split off.
-                tok_text = tok_text.replace(" ##", "")
-                tok_text = tok_text.replace("##", "")
-
-                # Clean whitespace
-                tok_text = tok_text.strip()
-                tok_text = " ".join(tok_text.split())
-                orig_text = " ".join(orig_tokens)
-
-                final_text = get_final_text(tok_text, orig_text, do_lower_case, verbose_logging)
-                if final_text in seen_predictions:
-                    continue
-
-                seen_predictions[final_text] = True
-            else:
-                final_text = ""
-                seen_predictions[final_text] = True
-
-            nbest.append(
-                _NbestPrediction(
-                    text=final_text,
-                    start_logit=pred.start_logit,
-                    end_logit=pred.end_logit))
-        # if we didn't include the empty option in the n-best, include it
-        if version_2_with_negative:
-            if "" not in seen_predictions:
-                nbest.append(
-                    _NbestPrediction(
-                        text="",
-                        start_logit=null_start_logit,
-                        end_logit=null_end_logit))
-                
-            # In very rare edge cases we could only have single null prediction.
-            # So we just create a nonce prediction in this case to avoid failure.
-            if len(nbest)==1:
-                nbest.insert(0,
-                    _NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
-
-        # In very rare edge cases we could have no valid predictions. So we
-        # just create a nonce prediction in this case to avoid failure.
-        if not nbest:
-            nbest.append(
-                _NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
-
-        assert len(nbest) >= 1
-
-        total_scores = []
-        best_non_null_entry = None
-        for entry in nbest:
-            total_scores.append(entry.start_logit + entry.end_logit)
-            if not best_non_null_entry:
-                if entry.text:
-                    best_non_null_entry = entry
-
-        probs = _compute_softmax(total_scores)
-
-        nbest_json = []
-        for (i, entry) in enumerate(nbest):
-            output = collections.OrderedDict()
-            output["text"] = entry.text
-            output["probability"] = probs[i]
-            output["start_logit"] = entry.start_logit
-            output["end_logit"] = entry.end_logit
-            nbest_json.append(output)
-
-        assert len(nbest_json) >= 1
-
-        if not version_2_with_negative:
-            all_predictions[example.qas_id] = nbest_json[0]["text"]
-        else:
-            # predict "" iff the null score - the score of best non-null > threshold
-            score_diff = score_null - best_non_null_entry.start_logit - (
-                best_non_null_entry.end_logit)
-            scores_diff_json[example.qas_id] = score_diff
-            if score_diff > null_score_diff_threshold:
-                all_predictions[example.qas_id] = ""
-            else:
-                all_predictions[example.qas_id] = best_non_null_entry.text
-        all_nbest_json[example.qas_id] = nbest_json
-
-    with open(output_prediction_file, "w") as writer:
-        writer.write(json.dumps(all_predictions, indent=4) + "\n")
-
-    with open(output_nbest_file, "w") as writer:
-        writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
-
-    if version_2_with_negative:
-        with open(output_null_log_odds_file, "w") as writer:
-            writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
-
-    return all_predictions
-
-
-# For XLNet (and XLM which uses the same head)
-RawResultExtended = collections.namedtuple("RawResultExtended",
-    ["unique_id", "start_top_log_probs", "start_top_index",
-     "end_top_log_probs", "end_top_index", "cls_logits"])
-
-
-def write_predictions_extended(all_examples, all_features, all_results, n_best_size,
-                                max_answer_length, output_prediction_file,
-                                output_nbest_file,
-                                output_null_log_odds_file, orig_data_file,
-                                start_n_top, end_n_top, version_2_with_negative,
-                                tokenizer, verbose_logging):
-    """ XLNet write prediction logic (more complex than Bert's).
-        Write final predictions to the json file and log-odds of null if needed.
-
-        Requires utils_squad_evaluate.py
-    """
-    _PrelimPrediction = collections.namedtuple(  # pylint: disable=invalid-name
-        "PrelimPrediction",
-        ["feature_index", "start_index", "end_index",
-        "start_log_prob", "end_log_prob"])
-
-    _NbestPrediction = collections.namedtuple(  # pylint: disable=invalid-name
-        "NbestPrediction", ["text", "start_log_prob", "end_log_prob"])
-
-    logger.info("Writing predictions to: %s", output_prediction_file)
-    # logger.info("Writing nbest to: %s" % (output_nbest_file))
-
-    example_index_to_features = collections.defaultdict(list)
-    for feature in all_features:
-        example_index_to_features[feature.example_index].append(feature)
-
-    unique_id_to_result = {}
-    for result in all_results:
-        unique_id_to_result[result.unique_id] = result
-
-    all_predictions = collections.OrderedDict()
-    all_nbest_json = collections.OrderedDict()
-    scores_diff_json = collections.OrderedDict()
-
-    for (example_index, example) in enumerate(all_examples):
-        features = example_index_to_features[example_index]
-
-        prelim_predictions = []
-        # keep track of the minimum score of null start+end of position 0
-        score_null = 1000000  # large and positive
-
-        for (feature_index, feature) in enumerate(features):
-            result = unique_id_to_result[feature.unique_id]
-
-            cur_null_score = result.cls_logits
-
-            # if we could have irrelevant answers, get the min score of irrelevant
-            score_null = min(score_null, cur_null_score)
-
-            for i in range(start_n_top):
-                for j in range(end_n_top):
-                    start_log_prob = result.start_top_log_probs[i]
-                    start_index = result.start_top_index[i]
-
-                    j_index = i * end_n_top + j
-
-                    end_log_prob = result.end_top_log_probs[j_index]
-                    end_index = result.end_top_index[j_index]
-
-                    # We could hypothetically create invalid predictions, e.g., predict
-                    # that the start of the span is in the question. We throw out all
-                    # invalid predictions.
-                    if start_index >= feature.paragraph_len - 1:
-                        continue
-                    if end_index >= feature.paragraph_len - 1:
-                        continue
-
-                    if not feature.token_is_max_context.get(start_index, False):
-                        continue
-                    if end_index < start_index:
-                        continue
-                    length = end_index - start_index + 1
-                    if length > max_answer_length:
-                        continue
-
-                    prelim_predictions.append(
-                        _PrelimPrediction(
-                            feature_index=feature_index,
-                            start_index=start_index,
-                            end_index=end_index,
-                            start_log_prob=start_log_prob,
-                            end_log_prob=end_log_prob))
-
-        prelim_predictions = sorted(
-            prelim_predictions,
-            key=lambda x: (x.start_log_prob + x.end_log_prob),
-            reverse=True)
-
-        seen_predictions = {}
-        nbest = []
-        for pred in prelim_predictions:
-            if len(nbest) >= n_best_size:
-                break
-            feature = features[pred.feature_index]
-
-            # XLNet un-tokenizer
-            # Let's keep it simple for now and see if we need all this later.
-            # 
-            # tok_start_to_orig_index = feature.tok_start_to_orig_index
-            # tok_end_to_orig_index = feature.tok_end_to_orig_index
-            # start_orig_pos = tok_start_to_orig_index[pred.start_index]
-            # end_orig_pos = tok_end_to_orig_index[pred.end_index]
-            # paragraph_text = example.paragraph_text
-            # final_text = paragraph_text[start_orig_pos: end_orig_pos + 1].strip()
-
-            # Previously used Bert untokenizer
-            tok_tokens = feature.tokens[pred.start_index:(pred.end_index + 1)]
-            orig_doc_start = feature.token_to_orig_map[pred.start_index]
-            orig_doc_end = feature.token_to_orig_map[pred.end_index]
-            orig_tokens = example.doc_tokens[orig_doc_start:(orig_doc_end + 1)]
-            tok_text = tokenizer.convert_tokens_to_string(tok_tokens)
-
-            # Clean whitespace
-            tok_text = tok_text.strip()
-            tok_text = " ".join(tok_text.split())
-            orig_text = " ".join(orig_tokens)
-
-            final_text = get_final_text(tok_text, orig_text, tokenizer.do_lower_case,
-                                        verbose_logging)
-
-            if final_text in seen_predictions:
-                continue
-
-            seen_predictions[final_text] = True
-
-            nbest.append(
-                _NbestPrediction(
-                    text=final_text,
-                    start_log_prob=pred.start_log_prob,
-                    end_log_prob=pred.end_log_prob))
-
-        # In very rare edge cases we could have no valid predictions. So we
-        # just create a nonce prediction in this case to avoid failure.
-        if not nbest:
-            nbest.append(
-                _NbestPrediction(text="", start_log_prob=-1e6,
-                end_log_prob=-1e6))
-
-        total_scores = []
-        best_non_null_entry = None
-        for entry in nbest:
-            total_scores.append(entry.start_log_prob + entry.end_log_prob)
-            if not best_non_null_entry:
-                best_non_null_entry = entry
-
-        probs = _compute_softmax(total_scores)
-
-        nbest_json = []
-        for (i, entry) in enumerate(nbest):
-            output = collections.OrderedDict()
-            output["text"] = entry.text
-            output["probability"] = probs[i]
-            output["start_log_prob"] = entry.start_log_prob
-            output["end_log_prob"] = entry.end_log_prob
-            nbest_json.append(output)
-
-        assert len(nbest_json) >= 1
-        assert best_non_null_entry is not None
-
-        score_diff = score_null
-        scores_diff_json[example.qas_id] = score_diff
-        # note(zhiliny): always predict best_non_null_entry
-        # and the evaluation script will search for the best threshold
-        all_predictions[example.qas_id] = best_non_null_entry.text
-
-        all_nbest_json[example.qas_id] = nbest_json
-
-    with open(output_prediction_file, "w") as writer:
-        writer.write(json.dumps(all_predictions, indent=4) + "\n")
-
-    with open(output_nbest_file, "w") as writer:
-        writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
-
-    if version_2_with_negative:
-        with open(output_null_log_odds_file, "w") as writer:
-            writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
-
-    with open(orig_data_file, "r", encoding='utf-8') as reader:
-        orig_data = json.load(reader)["data"]
-
-    qid_to_has_ans = make_qid_to_has_ans(orig_data)
-    has_ans_qids = [k for k, v in qid_to_has_ans.items() if v]
-    no_ans_qids = [k for k, v in qid_to_has_ans.items() if not v]
-    exact_raw, f1_raw = get_raw_scores(orig_data, all_predictions)
-    out_eval = {}
-
-    find_all_best_thresh_v2(out_eval, all_predictions, exact_raw, f1_raw, scores_diff_json, qid_to_has_ans)
-
-    return out_eval
-
-
-def get_final_text(pred_text, orig_text, do_lower_case, verbose_logging=False):
-    """Project the tokenized prediction back to the original text."""
-
-    # When we created the data, we kept track of the alignment between original
-    # (whitespace tokenized) tokens and our WordPiece tokenized tokens. So
-    # now `orig_text` contains the span of our original text corresponding to the
-    # span that we predicted.
-    #
-    # However, `orig_text` may contain extra characters that we don't want in
-    # our prediction.
-    #
-    # For example, let's say:
-    #   pred_text = steve smith
-    #   orig_text = Steve Smith's
-    #
-    # We don't want to return `orig_text` because it contains the extra "'s".
-    #
-    # We don't want to return `pred_text` because it's already been normalized
-    # (the SQuAD eval script also does punctuation stripping/lower casing but
-    # our tokenizer does additional normalization like stripping accent
-    # characters).
-    #
-    # What we really want to return is "Steve Smith".
-    #
-    # Therefore, we have to apply a semi-complicated alignment heuristic between
-    # `pred_text` and `orig_text` to get a character-to-character alignment. This
-    # can fail in certain cases in which case we just return `orig_text`.
-
-    def _strip_spaces(text):
-        ns_chars = []
-        ns_to_s_map = collections.OrderedDict()
-        for (i, c) in enumerate(text):
-            if c == " ":
-                continue
-            ns_to_s_map[len(ns_chars)] = i
-            ns_chars.append(c)
-        ns_text = "".join(ns_chars)
-        return (ns_text, ns_to_s_map)
-
-    # We first tokenize `orig_text`, strip whitespace from the result
-    # and `pred_text`, and check if they are the same length. If they are
-    # NOT the same length, the heuristic has failed. If they are the same
-    # length, we assume the characters are one-to-one aligned.
-    tokenizer = BasicTokenizer(do_lower_case=do_lower_case)
-
-    tok_text = " ".join(tokenizer.tokenize(orig_text))
-
-    start_position = tok_text.find(pred_text)
-    if start_position == -1:
-        if verbose_logging:
-            logger.info(
-                "Unable to find text: '%s' in '%s'" % (pred_text, orig_text))
-        return orig_text
-    end_position = start_position + len(pred_text) - 1
-
-    (orig_ns_text, orig_ns_to_s_map) = _strip_spaces(orig_text)
-    (tok_ns_text, tok_ns_to_s_map) = _strip_spaces(tok_text)
-
-    if len(orig_ns_text) != len(tok_ns_text):
-        if verbose_logging:
-            logger.info("Length not equal after stripping spaces: '%s' vs '%s'",
-                        orig_ns_text, tok_ns_text)
-        return orig_text
-
-    # We then project the characters in `pred_text` back to `orig_text` using
-    # the character-to-character alignment.
-    tok_s_to_ns_map = {}
-    for (i, tok_index) in tok_ns_to_s_map.items():
-        tok_s_to_ns_map[tok_index] = i
-
-    orig_start_position = None
-    if start_position in tok_s_to_ns_map:
-        ns_start_position = tok_s_to_ns_map[start_position]
-        if ns_start_position in orig_ns_to_s_map:
-            orig_start_position = orig_ns_to_s_map[ns_start_position]
-
-    if orig_start_position is None:
-        if verbose_logging:
-            logger.info("Couldn't map start position")
-        return orig_text
-
-    orig_end_position = None
-    if end_position in tok_s_to_ns_map:
-        ns_end_position = tok_s_to_ns_map[end_position]
-        if ns_end_position in orig_ns_to_s_map:
-            orig_end_position = orig_ns_to_s_map[ns_end_position]
-
-    if orig_end_position is None:
-        if verbose_logging:
-            logger.info("Couldn't map end position")
-        return orig_text
-
-    output_text = orig_text[orig_start_position:(orig_end_position + 1)]
-    return output_text
-
-
-def _get_best_indexes(logits, n_best_size):
-    """Get the n-best logits from a list."""
-    index_and_score = sorted(enumerate(logits), key=lambda x: x[1], reverse=True)
-
-    best_indexes = []
-    for i in range(len(index_and_score)):
-        if i >= n_best_size:
-            break
-        best_indexes.append(index_and_score[i][0])
-    return best_indexes
-
-
-def _compute_softmax(scores):
-    """Compute softmax probability over raw logits."""
-    if not scores:
-        return []
-
-    max_score = None
-    for score in scores:
-        if max_score is None or score > max_score:
-            max_score = score
-
-    exp_scores = []
-    total_sum = 0.0
-    for score in scores:
-        x = math.exp(score - max_score)
-        exp_scores.append(x)
-        total_sum += x
-
-    probs = []
-    for score in exp_scores:
-        probs.append(score / total_sum)
-    return probs

examples/utils_squad_evaluate.py

-""" Official evaluation script for SQuAD version 2.0.
-    Modified by XLNet authors to update `find_best_threshold` scripts for SQuAD V2.0
-
-In addition to basic functionality, we also compute additional statistics and
-plot precision-recall curves if an additional na_prob.json file is provided.
-This file is expected to map question ID's to the model's predicted probability
-that a question is unanswerable.
-"""
-import argparse
-import collections
-import json
-import numpy as np
-import os
-import re
-import string
-import sys
-
-class EVAL_OPTS():
-  def __init__(self, data_file, pred_file, out_file="",
-               na_prob_file="na_prob.json", na_prob_thresh=1.0,
-               out_image_dir=None, verbose=False):
-    self.data_file = data_file
-    self.pred_file = pred_file
-    self.out_file = out_file
-    self.na_prob_file = na_prob_file
-    self.na_prob_thresh = na_prob_thresh
-    self.out_image_dir = out_image_dir
-    self.verbose = verbose
-
-OPTS = None
-
-def parse_args():
-  parser = argparse.ArgumentParser('Official evaluation script for SQuAD version 2.0.')
-  parser.add_argument('data_file', metavar='data.json', help='Input data JSON file.')
-  parser.add_argument('pred_file', metavar='pred.json', help='Model predictions.')
-  parser.add_argument('--out-file', '-o', metavar='eval.json',
-                      help='Write accuracy metrics to file (default is stdout).')
-  parser.add_argument('--na-prob-file', '-n', metavar='na_prob.json',
-                      help='Model estimates of probability of no answer.')
-  parser.add_argument('--na-prob-thresh', '-t', type=float, default=1.0,
-                      help='Predict "" if no-answer probability exceeds this (default = 1.0).')
-  parser.add_argument('--out-image-dir', '-p', metavar='out_images', default=None,
-                      help='Save precision-recall curves to directory.')
-  parser.add_argument('--verbose', '-v', action='store_true')
-  if len(sys.argv) == 1:
-    parser.print_help()
-    sys.exit(1)
-  return parser.parse_args()
-
-def make_qid_to_has_ans(dataset):
-  qid_to_has_ans = {}
-  for article in dataset:
-    for p in article['paragraphs']:
-      for qa in p['qas']:
-        qid_to_has_ans[qa['id']] = bool(qa['answers'])
-  return qid_to_has_ans
-
-def normalize_answer(s):
-  """Lower text and remove punctuation, articles and extra whitespace."""
-  def remove_articles(text):
-    regex = re.compile(r'\b(a|an|the)\b', re.UNICODE)
-    return re.sub(regex, ' ', text)
-  def white_space_fix(text):
-    return ' '.join(text.split())
-  def remove_punc(text):
-    exclude = set(string.punctuation)
-    return ''.join(ch for ch in text if ch not in exclude)
-  def lower(text):
-    return text.lower()
-  return white_space_fix(remove_articles(remove_punc(lower(s))))
-
-def get_tokens(s):
-  if not s: return []
-  return normalize_answer(s).split()
-
-def compute_exact(a_gold, a_pred):
-  return int(normalize_answer(a_gold) == normalize_answer(a_pred))
-
-def compute_f1(a_gold, a_pred):
-  gold_toks = get_tokens(a_gold)
-  pred_toks = get_tokens(a_pred)
-  common = collections.Counter(gold_toks) & collections.Counter(pred_toks)
-  num_same = sum(common.values())
-  if len(gold_toks) == 0 or len(pred_toks) == 0:
-    # If either is no-answer, then F1 is 1 if they agree, 0 otherwise
-    return int(gold_toks == pred_toks)
-  if num_same == 0:
-    return 0
-  precision = 1.0 * num_same / len(pred_toks)
-  recall = 1.0 * num_same / len(gold_toks)
-  f1 = (2 * precision * recall) / (precision + recall)
-  return f1
-
-def get_raw_scores(dataset, preds):
-  exact_scores = {}
-  f1_scores = {}
-  for article in dataset:
-    for p in article['paragraphs']:
-      for qa in p['qas']:
-        qid = qa['id']
-        gold_answers = [a['text'] for a in qa['answers']
-                        if normalize_answer(a['text'])]
-        if not gold_answers:
-          # For unanswerable questions, only correct answer is empty string
-          gold_answers = ['']
-        if qid not in preds:
-          print('Missing prediction for %s' % qid)
-          continue
-        a_pred = preds[qid]
-        # Take max over all gold answers
-        exact_scores[qid] = max(compute_exact(a, a_pred) for a in gold_answers)
-        f1_scores[qid] = max(compute_f1(a, a_pred) for a in gold_answers)
-  return exact_scores, f1_scores
-
-def apply_no_ans_threshold(scores, na_probs, qid_to_has_ans, na_prob_thresh):
-  new_scores = {}
-  for qid, s in scores.items():
-    pred_na = na_probs[qid] > na_prob_thresh
-    if pred_na:
-      new_scores[qid] = float(not qid_to_has_ans[qid])
-    else:
-      new_scores[qid] = s
-  return new_scores
-
-def make_eval_dict(exact_scores, f1_scores, qid_list=None):
-  if not qid_list:
-    total = len(exact_scores)
-    return collections.OrderedDict([
-        ('exact', 100.0 * sum(exact_scores.values()) / total),
-        ('f1', 100.0 * sum(f1_scores.values()) / total),
-        ('total', total),
-    ])
-  else:
-    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 merge_eval(main_eval, new_eval, prefix):
-  for k in new_eval:
-    main_eval['%s_%s' % (prefix, k)] = new_eval[k]
-
-def plot_pr_curve(precisions, recalls, out_image, title):
-  plt.step(recalls, precisions, color='b', alpha=0.2, where='post')
-  plt.fill_between(recalls, precisions, step='post', alpha=0.2, color='b')
-  plt.xlabel('Recall')
-  plt.ylabel('Precision')
-  plt.xlim([0.0, 1.05])
-  plt.ylim([0.0, 1.05])
-  plt.title(title)
-  plt.savefig(out_image)
-  plt.clf()
-
-def make_precision_recall_eval(scores, na_probs, num_true_pos, qid_to_has_ans,
-                               out_image=None, title=None):
-  qid_list = sorted(na_probs, key=lambda k: na_probs[k])
-  true_pos = 0.0
-  cur_p = 1.0
-  cur_r = 0.0
-  precisions = [1.0]
-  recalls = [0.0]
-  avg_prec = 0.0
-  for i, qid in enumerate(qid_list):
-    if qid_to_has_ans[qid]:
-      true_pos += scores[qid]
-    cur_p = true_pos / float(i+1)
-    cur_r = true_pos / float(num_true_pos)
-    if i == len(qid_list) - 1 or na_probs[qid] != na_probs[qid_list[i+1]]:
-      # i.e., if we can put a threshold after this point
-      avg_prec += cur_p * (cur_r - recalls[-1])
-      precisions.append(cur_p)
-      recalls.append(cur_r)
-  if out_image:
-    plot_pr_curve(precisions, recalls, out_image, title)
-  return {'ap': 100.0 * avg_prec}
-
-def run_precision_recall_analysis(main_eval, exact_raw, f1_raw, na_probs, 
-                                  qid_to_has_ans, out_image_dir):
-  if out_image_dir and not os.path.exists(out_image_dir):
-    os.makedirs(out_image_dir)
-  num_true_pos = sum(1 for v in qid_to_has_ans.values() if v)
-  if num_true_pos == 0:
-    return
-  pr_exact = make_precision_recall_eval(
-      exact_raw, na_probs, num_true_pos, qid_to_has_ans,
-      out_image=os.path.join(out_image_dir, 'pr_exact.png'),
-      title='Precision-Recall curve for Exact Match score')
-  pr_f1 = make_precision_recall_eval(
-      f1_raw, na_probs, num_true_pos, qid_to_has_ans,
-      out_image=os.path.join(out_image_dir, 'pr_f1.png'),
-      title='Precision-Recall curve for F1 score')
-  oracle_scores = {k: float(v) for k, v in qid_to_has_ans.items()}
-  pr_oracle = make_precision_recall_eval(
-      oracle_scores, na_probs, num_true_pos, qid_to_has_ans,
-      out_image=os.path.join(out_image_dir, 'pr_oracle.png'),
-      title='Oracle Precision-Recall curve (binary task of HasAns vs. NoAns)')
-  merge_eval(main_eval, pr_exact, 'pr_exact')
-  merge_eval(main_eval, pr_f1, 'pr_f1')
-  merge_eval(main_eval, pr_oracle, 'pr_oracle')
-
-def histogram_na_prob(na_probs, qid_list, image_dir, name):
-  if not qid_list:
-    return
-  x = [na_probs[k] for k in qid_list]
-  weights = np.ones_like(x) / float(len(x))
-  plt.hist(x, weights=weights, bins=20, range=(0.0, 1.0))
-  plt.xlabel('Model probability of no-answer')
-  plt.ylabel('Proportion of dataset')
-  plt.title('Histogram of no-answer probability: %s' % name)
-  plt.savefig(os.path.join(image_dir, 'na_prob_hist_%s.png' % name))
-  plt.clf()
-
-def find_best_thresh(preds, scores, na_probs, qid_to_has_ans):
-  num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
-  cur_score = num_no_ans
-  best_score = cur_score
-  best_thresh = 0.0
-  qid_list = sorted(na_probs, key=lambda k: na_probs[k])
-  for i, qid in enumerate(qid_list):
-    if qid not in scores: continue
-    if qid_to_has_ans[qid]:
-      diff = scores[qid]
-    else:
-      if preds[qid]:
-        diff = -1
-      else:
-        diff = 0
-    cur_score += diff
-    if cur_score > best_score:
-      best_score = cur_score
-      best_thresh = na_probs[qid]
-  return 100.0 * best_score / len(scores), best_thresh
-
-def find_best_thresh_v2(preds, scores, na_probs, qid_to_has_ans):
-  num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
-  cur_score = num_no_ans
-  best_score = cur_score
-  best_thresh = 0.0
-  qid_list = sorted(na_probs, key=lambda k: na_probs[k])
-  for i, qid in enumerate(qid_list):
-    if qid not in scores: continue
-    if qid_to_has_ans[qid]:
-      diff = scores[qid]
-    else:
-      if preds[qid]:
-        diff = -1
-      else:
-        diff = 0
-    cur_score += diff
-    if cur_score > best_score:
-      best_score = cur_score
-      best_thresh = na_probs[qid]
-
-  has_ans_score, has_ans_cnt = 0, 0
-  for qid in qid_list:
-    if not qid_to_has_ans[qid]: continue
-    has_ans_cnt += 1
-
-    if qid not in scores: continue
-    has_ans_score += scores[qid]
-
-  return 100.0 * best_score / len(scores), best_thresh, 1.0 * has_ans_score / has_ans_cnt
-
-def find_all_best_thresh(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
-  best_exact, exact_thresh = find_best_thresh(preds, exact_raw, na_probs, qid_to_has_ans)
-  best_f1, f1_thresh = find_best_thresh(preds, f1_raw, na_probs, qid_to_has_ans)
-  main_eval['best_exact'] = best_exact
-  main_eval['best_exact_thresh'] = exact_thresh
-  main_eval['best_f1'] = best_f1
-  main_eval['best_f1_thresh'] = f1_thresh
-
-def find_all_best_thresh_v2(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
-  best_exact, exact_thresh, has_ans_exact = find_best_thresh_v2(preds, exact_raw, na_probs, qid_to_has_ans)
-  best_f1, f1_thresh, has_ans_f1 = find_best_thresh_v2(preds, f1_raw, na_probs, qid_to_has_ans)
-  main_eval['best_exact'] = best_exact
-  main_eval['best_exact_thresh'] = exact_thresh
-  main_eval['best_f1'] = best_f1
-  main_eval['best_f1_thresh'] = f1_thresh
-  main_eval['has_ans_exact'] = has_ans_exact
-  main_eval['has_ans_f1'] = has_ans_f1
-
-def main(OPTS):
-  with open(OPTS.data_file) as f:
-    dataset_json = json.load(f)
-    dataset = dataset_json['data']
-  with open(OPTS.pred_file) as f:
-    preds = json.load(f)
-  if OPTS.na_prob_file:
-    with open(OPTS.na_prob_file) as f:
-      na_probs = json.load(f)
-  else:
-    na_probs = {k: 0.0 for k in preds}
-  qid_to_has_ans = make_qid_to_has_ans(dataset)  # maps qid to True/False
-  has_ans_qids = [k for k, v in qid_to_has_ans.items() if v]
-  no_ans_qids = [k for k, v in qid_to_has_ans.items() if not v]
-  exact_raw, f1_raw = get_raw_scores(dataset, preds)
-  exact_thresh = apply_no_ans_threshold(exact_raw, na_probs, qid_to_has_ans,
-                                        OPTS.na_prob_thresh)
-  f1_thresh = apply_no_ans_threshold(f1_raw, na_probs, qid_to_has_ans,
-                                     OPTS.na_prob_thresh)
-  out_eval = make_eval_dict(exact_thresh, f1_thresh)
-  if has_ans_qids:
-    has_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=has_ans_qids)
-    merge_eval(out_eval, has_ans_eval, 'HasAns')
-  if no_ans_qids:
-    no_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=no_ans_qids)
-    merge_eval(out_eval, no_ans_eval, 'NoAns')
-  if OPTS.na_prob_file:
-    find_all_best_thresh(out_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans)
-  if OPTS.na_prob_file and OPTS.out_image_dir:
-    run_precision_recall_analysis(out_eval, exact_raw, f1_raw, na_probs, 
-                                  qid_to_has_ans, OPTS.out_image_dir)
-    histogram_na_prob(na_probs, has_ans_qids, OPTS.out_image_dir, 'hasAns')
-    histogram_na_prob(na_probs, no_ans_qids, OPTS.out_image_dir, 'noAns')
-  if OPTS.out_file:
-    with open(OPTS.out_file, 'w') as f:
-      json.dump(out_eval, f)
-  else:
-    print(json.dumps(out_eval, indent=2))
-  return out_eval
-
-if __name__ == '__main__':
-  OPTS = parse_args()
-  if OPTS.out_image_dir:
-    import matplotlib
-    matplotlib.use('Agg')
-    import matplotlib.pyplot as plt 
-  main(OPTS)

transformers/data/metrics/squad_metrics.py

@@ -578,7 +578,6 @@ def compute_predictions_log_probs(
     output_prediction_file,
     output_nbest_file,
     output_null_log_odds_file,
-    orig_data_file,
     start_n_top,
     end_n_top,
     version_2_with_negative,
@@ -756,15 +755,4 @@ def compute_predictions_log_probs(
         with open(output_null_log_odds_file, "w") as writer:
             writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
 
-    with open(orig_data_file, "r", encoding='utf-8') as reader:
-        orig_data = json.load(reader)["data"]
-
-    qid_to_has_ans = make_qid_to_has_ans(orig_data)
-    has_ans_qids = [k for k, v in qid_to_has_ans.items() if v]
-    no_ans_qids = [k for k, v in qid_to_has_ans.items() if not v]
-    exact_raw, f1_raw = get_raw_scores(orig_data, all_predictions)
-    out_eval = {}
-
-    find_all_best_thresh_v2(out_eval, all_predictions, exact_raw, f1_raw, scores_diff_json, qid_to_has_ans)
-
-    return out_eval
+    return all_predictions

transformers/data/processors/squad.py

@@ -9,7 +9,7 @@ from ...tokenization_bert import BasicTokenizer, whitespace_tokenize
 from .utils import DataProcessor, InputExample, InputFeatures
 from ...file_utils import is_tf_available, is_torch_available
 
-if is_torch_available:
+if is_torch_available():
     import torch
     from torch.utils.data import TensorDataset