Move source code inside a src subdirectory.

This prevents transformers from being importable simply because the CWD
is the root of the git repository, while not being importable from other
directories. That led to inconsistent behavior, especially in examples.

Once you fetch this commit, in your dev environment, you must run:

    $ pip uninstall transformers
    $ pip install -e .

src/transformers/data/__init__.py

@@ -0,0 +1,27 @@
+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+from .metrics import is_sklearn_available
+from .processors import (
+    DataProcessor,
+    InputExample,
+    InputFeatures,
+    SingleSentenceClassificationProcessor,
+    SquadExample,
+    SquadFeatures,
+    SquadV1Processor,
+    SquadV2Processor,
+    glue_convert_examples_to_features,
+    glue_output_modes,
+    glue_processors,
+    glue_tasks_num_labels,
+    squad_convert_examples_to_features,
+    xnli_output_modes,
+    xnli_processors,
+    xnli_tasks_num_labels,
+)
+
+
+if is_sklearn_available():
+    from .metrics import glue_compute_metrics, xnli_compute_metrics

src/transformers/data/metrics/__init__.py

@@ -0,0 +1,89 @@
+# 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.
+
+import logging
+
+
+logger = logging.getLogger(__name__)
+
+try:
+    from scipy.stats import pearsonr, spearmanr
+    from sklearn.metrics import matthews_corrcoef, f1_score
+
+    _has_sklearn = True
+except (AttributeError, ImportError) as e:
+    logger.warning("To use data.metrics please install scikit-learn. See https://scikit-learn.org/stable/index.html")
+    _has_sklearn = False
+
+
+def is_sklearn_available():
+    return _has_sklearn
+
+
+if _has_sklearn:
+
+    def simple_accuracy(preds, labels):
+        return (preds == labels).mean()
+
+    def acc_and_f1(preds, labels):
+        acc = simple_accuracy(preds, labels)
+        f1 = f1_score(y_true=labels, y_pred=preds)
+        return {
+            "acc": acc,
+            "f1": f1,
+            "acc_and_f1": (acc + f1) / 2,
+        }
+
+    def pearson_and_spearman(preds, labels):
+        pearson_corr = pearsonr(preds, labels)[0]
+        spearman_corr = spearmanr(preds, labels)[0]
+        return {
+            "pearson": pearson_corr,
+            "spearmanr": spearman_corr,
+            "corr": (pearson_corr + spearman_corr) / 2,
+        }
+
+    def glue_compute_metrics(task_name, preds, labels):
+        assert len(preds) == len(labels)
+        if task_name == "cola":
+            return {"mcc": matthews_corrcoef(labels, preds)}
+        elif task_name == "sst-2":
+            return {"acc": simple_accuracy(preds, labels)}
+        elif task_name == "mrpc":
+            return acc_and_f1(preds, labels)
+        elif task_name == "sts-b":
+            return pearson_and_spearman(preds, labels)
+        elif task_name == "qqp":
+            return acc_and_f1(preds, labels)
+        elif task_name == "mnli":
+            return {"acc": simple_accuracy(preds, labels)}
+        elif task_name == "mnli-mm":
+            return {"acc": simple_accuracy(preds, labels)}
+        elif task_name == "qnli":
+            return {"acc": simple_accuracy(preds, labels)}
+        elif task_name == "rte":
+            return {"acc": simple_accuracy(preds, labels)}
+        elif task_name == "wnli":
+            return {"acc": simple_accuracy(preds, labels)}
+        else:
+            raise KeyError(task_name)
+
+    def xnli_compute_metrics(task_name, preds, labels):
+        assert len(preds) == len(labels)
+        if task_name == "xnli":
+            return {"acc": simple_accuracy(preds, labels)}
+        else:
+            raise KeyError(task_name)

src/transformers/data/metrics/squad_metrics.py

@@ -0,0 +1,758 @@
+""" Very heavily inspired by the official evaluation script for SQuAD version 2.0 which was
+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 collections
+import json
+import logging
+import math
+import re
+import string
+from io import open
+
+from transformers.tokenization_bert import BasicTokenizer
+
+
+logger = logging.getLogger(__name__)
+
+
+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(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)
+
+    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 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_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 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 _, 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_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 squad_evaluate(examples, preds, no_answer_probs=None, no_answer_probability_threshold=1.0):
+    qas_id_to_has_answer = {example.qas_id: bool(example.answers) for example in examples}
+    has_answer_qids = [qas_id for qas_id, has_answer in qas_id_to_has_answer.items() if has_answer]
+    no_answer_qids = [qas_id for qas_id, has_answer in qas_id_to_has_answer.items() if not has_answer]
+
+    if no_answer_probs is None:
+        no_answer_probs = {k: 0.0 for k in preds}
+
+    exact, f1 = get_raw_scores(examples, preds)
+
+    exact_threshold = apply_no_ans_threshold(
+        exact, no_answer_probs, qas_id_to_has_answer, no_answer_probability_threshold
+    )
+    f1_threshold = apply_no_ans_threshold(f1, no_answer_probs, qas_id_to_has_answer, no_answer_probability_threshold)
+
+    evaluation = make_eval_dict(exact_threshold, f1_threshold)
+
+    if has_answer_qids:
+        has_ans_eval = make_eval_dict(exact_threshold, f1_threshold, qid_list=has_answer_qids)
+        merge_eval(evaluation, has_ans_eval, "HasAns")
+
+    if no_answer_qids:
+        no_ans_eval = make_eval_dict(exact_threshold, f1_threshold, qid_list=no_answer_qids)
+        merge_eval(evaluation, no_ans_eval, "NoAns")
+
+    if no_answer_probs:
+        find_all_best_thresh(evaluation, preds, exact, f1, no_answer_probs, qas_id_to_has_answer)
+
+    return evaluation
+
+
+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
+
+
+def compute_predictions_logits(
+    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,
+    tokenizer,
+):
+    """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 = tokenizer.convert_tokens_to_string(tok_tokens)
+
+                # 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
+
+
+def compute_predictions_log_probs(
+    all_examples,
+    all_features,
+    all_results,
+    n_best_size,
+    max_answer_length,
+    output_prediction_file,
+    output_nbest_file,
+    output_null_log_odds_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_logits[i]
+                    start_index = result.start_top_index[i]
+
+                    j_index = i * end_n_top + j
+
+                    end_log_prob = result.end_logits[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)
+
+            if hasattr(tokenizer, "do_lower_case"):
+                do_lower_case = tokenizer.do_lower_case
+            else:
+                do_lower_case = tokenizer.do_lowercase_and_remove_accent
+
+            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
+
+            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")
+
+    return all_predictions

src/transformers/data/processors/__init__.py

@@ -0,0 +1,8 @@
+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+from .glue import glue_convert_examples_to_features, glue_output_modes, glue_processors, glue_tasks_num_labels
+from .squad import SquadExample, SquadFeatures, SquadV1Processor, SquadV2Processor, squad_convert_examples_to_features
+from .utils import DataProcessor, InputExample, InputFeatures, SingleSentenceClassificationProcessor
+from .xnli import xnli_output_modes, xnli_processors, xnli_tasks_num_labels

src/transformers/data/processors/glue.py

@@ -0,0 +1,551 @@
+# 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.
+""" GLUE processors and helpers """
+
+import logging
+import os
+
+from ...file_utils import is_tf_available
+from .utils import DataProcessor, InputExample, InputFeatures
+
+
+if is_tf_available():
+    import tensorflow as tf
+
+logger = logging.getLogger(__name__)
+
+
+def glue_convert_examples_to_features(
+    examples,
+    tokenizer,
+    max_length=512,
+    task=None,
+    label_list=None,
+    output_mode=None,
+    pad_on_left=False,
+    pad_token=0,
+    pad_token_segment_id=0,
+    mask_padding_with_zero=True,
+):
+    """
+    Loads a data file into a list of ``InputFeatures``
+
+    Args:
+        examples: List of ``InputExamples`` or ``tf.data.Dataset`` containing the examples.
+        tokenizer: Instance of a tokenizer that will tokenize the examples
+        max_length: Maximum example length
+        task: GLUE task
+        label_list: List of labels. Can be obtained from the processor using the ``processor.get_labels()`` method
+        output_mode: String indicating the output mode. Either ``regression`` or ``classification``
+        pad_on_left: If set to ``True``, the examples will be padded on the left rather than on the right (default)
+        pad_token: Padding token
+        pad_token_segment_id: The segment ID for the padding token (It is usually 0, but can vary such as for XLNet where it is 4)
+        mask_padding_with_zero: If set to ``True``, the attention mask will be filled by ``1`` for actual values
+            and by ``0`` for padded values. If set to ``False``, inverts it (``1`` for padded values, ``0`` for
+            actual values)
+
+    Returns:
+        If the ``examples`` input is a ``tf.data.Dataset``, will return a ``tf.data.Dataset``
+        containing the task-specific features. If the input is a list of ``InputExamples``, will return
+        a list of task-specific ``InputFeatures`` which can be fed to the model.
+
+    """
+    is_tf_dataset = False
+    if is_tf_available() and isinstance(examples, tf.data.Dataset):
+        is_tf_dataset = True
+
+    if task is not None:
+        processor = glue_processors[task]()
+        if label_list is None:
+            label_list = processor.get_labels()
+            logger.info("Using label list %s for task %s" % (label_list, task))
+        if output_mode is None:
+            output_mode = glue_output_modes[task]
+            logger.info("Using output mode %s for task %s" % (output_mode, task))
+
+    label_map = {label: i for i, label in enumerate(label_list)}
+
+    features = []
+    for (ex_index, example) in enumerate(examples):
+        if ex_index % 10000 == 0:
+            logger.info("Writing example %d" % (ex_index))
+        if is_tf_dataset:
+            example = processor.get_example_from_tensor_dict(example)
+            example = processor.tfds_map(example)
+
+        inputs = tokenizer.encode_plus(example.text_a, example.text_b, add_special_tokens=True, max_length=max_length,)
+        input_ids, token_type_ids = inputs["input_ids"], inputs["token_type_ids"]
+
+        # The mask has 1 for real tokens and 0 for padding tokens. Only real
+        # tokens are attended to.
+        attention_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
+
+        # Zero-pad up to the sequence length.
+        padding_length = max_length - len(input_ids)
+        if pad_on_left:
+            input_ids = ([pad_token] * padding_length) + input_ids
+            attention_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + attention_mask
+            token_type_ids = ([pad_token_segment_id] * padding_length) + token_type_ids
+        else:
+            input_ids = input_ids + ([pad_token] * padding_length)
+            attention_mask = attention_mask + ([0 if mask_padding_with_zero else 1] * padding_length)
+            token_type_ids = token_type_ids + ([pad_token_segment_id] * padding_length)
+
+        assert len(input_ids) == max_length, "Error with input length {} vs {}".format(len(input_ids), max_length)
+        assert len(attention_mask) == max_length, "Error with input length {} vs {}".format(
+            len(attention_mask), max_length
+        )
+        assert len(token_type_ids) == max_length, "Error with input length {} vs {}".format(
+            len(token_type_ids), max_length
+        )
+
+        if output_mode == "classification":
+            label = label_map[example.label]
+        elif output_mode == "regression":
+            label = float(example.label)
+        else:
+            raise KeyError(output_mode)
+
+        if ex_index < 5:
+            logger.info("*** Example ***")
+            logger.info("guid: %s" % (example.guid))
+            logger.info("input_ids: %s" % " ".join([str(x) for x in input_ids]))
+            logger.info("attention_mask: %s" % " ".join([str(x) for x in attention_mask]))
+            logger.info("token_type_ids: %s" % " ".join([str(x) for x in token_type_ids]))
+            logger.info("label: %s (id = %d)" % (example.label, label))
+
+        features.append(
+            InputFeatures(
+                input_ids=input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids, label=label
+            )
+        )
+
+    if is_tf_available() and is_tf_dataset:
+
+        def gen():
+            for ex in features:
+                yield (
+                    {
+                        "input_ids": ex.input_ids,
+                        "attention_mask": ex.attention_mask,
+                        "token_type_ids": ex.token_type_ids,
+                    },
+                    ex.label,
+                )
+
+        return tf.data.Dataset.from_generator(
+            gen,
+            ({"input_ids": tf.int32, "attention_mask": tf.int32, "token_type_ids": tf.int32}, tf.int64),
+            (
+                {
+                    "input_ids": tf.TensorShape([None]),
+                    "attention_mask": tf.TensorShape([None]),
+                    "token_type_ids": tf.TensorShape([None]),
+                },
+                tf.TensorShape([]),
+            ),
+        )
+
+    return features
+
+
+class MrpcProcessor(DataProcessor):
+    """Processor for the MRPC data set (GLUE version)."""
+
+    def get_example_from_tensor_dict(self, tensor_dict):
+        """See base class."""
+        return InputExample(
+            tensor_dict["idx"].numpy(),
+            tensor_dict["sentence1"].numpy().decode("utf-8"),
+            tensor_dict["sentence2"].numpy().decode("utf-8"),
+            str(tensor_dict["label"].numpy()),
+        )
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {}".format(os.path.join(data_dir, "train.tsv")))
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, i)
+            text_a = line[3]
+            text_b = line[4]
+            label = line[0]
+            examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+class MnliProcessor(DataProcessor):
+    """Processor for the MultiNLI data set (GLUE version)."""
+
+    def get_example_from_tensor_dict(self, tensor_dict):
+        """See base class."""
+        return InputExample(
+            tensor_dict["idx"].numpy(),
+            tensor_dict["premise"].numpy().decode("utf-8"),
+            tensor_dict["hypothesis"].numpy().decode("utf-8"),
+            str(tensor_dict["label"].numpy()),
+        )
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev_matched.tsv")), "dev_matched")
+
+    def get_labels(self):
+        """See base class."""
+        return ["contradiction", "entailment", "neutral"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, line[0])
+            text_a = line[8]
+            text_b = line[9]
+            label = line[-1]
+            examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+class MnliMismatchedProcessor(MnliProcessor):
+    """Processor for the MultiNLI Mismatched data set (GLUE version)."""
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev_mismatched.tsv")), "dev_matched")
+
+
+class ColaProcessor(DataProcessor):
+    """Processor for the CoLA data set (GLUE version)."""
+
+    def get_example_from_tensor_dict(self, tensor_dict):
+        """See base class."""
+        return InputExample(
+            tensor_dict["idx"].numpy(),
+            tensor_dict["sentence"].numpy().decode("utf-8"),
+            None,
+            str(tensor_dict["label"].numpy()),
+        )
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            guid = "%s-%s" % (set_type, i)
+            text_a = line[3]
+            label = line[1]
+            examples.append(InputExample(guid=guid, text_a=text_a, text_b=None, label=label))
+        return examples
+
+
+class Sst2Processor(DataProcessor):
+    """Processor for the SST-2 data set (GLUE version)."""
+
+    def get_example_from_tensor_dict(self, tensor_dict):
+        """See base class."""
+        return InputExample(
+            tensor_dict["idx"].numpy(),
+            tensor_dict["sentence"].numpy().decode("utf-8"),
+            None,
+            str(tensor_dict["label"].numpy()),
+        )
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, i)
+            text_a = line[0]
+            label = line[1]
+            examples.append(InputExample(guid=guid, text_a=text_a, text_b=None, label=label))
+        return examples
+
+
+class StsbProcessor(DataProcessor):
+    """Processor for the STS-B data set (GLUE version)."""
+
+    def get_example_from_tensor_dict(self, tensor_dict):
+        """See base class."""
+        return InputExample(
+            tensor_dict["idx"].numpy(),
+            tensor_dict["sentence1"].numpy().decode("utf-8"),
+            tensor_dict["sentence2"].numpy().decode("utf-8"),
+            str(tensor_dict["label"].numpy()),
+        )
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return [None]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, line[0])
+            text_a = line[7]
+            text_b = line[8]
+            label = line[-1]
+            examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+class QqpProcessor(DataProcessor):
+    """Processor for the QQP data set (GLUE version)."""
+
+    def get_example_from_tensor_dict(self, tensor_dict):
+        """See base class."""
+        return InputExample(
+            tensor_dict["idx"].numpy(),
+            tensor_dict["question1"].numpy().decode("utf-8"),
+            tensor_dict["question2"].numpy().decode("utf-8"),
+            str(tensor_dict["label"].numpy()),
+        )
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, line[0])
+            try:
+                text_a = line[3]
+                text_b = line[4]
+                label = line[5]
+            except IndexError:
+                continue
+            examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+class QnliProcessor(DataProcessor):
+    """Processor for the QNLI data set (GLUE version)."""
+
+    def get_example_from_tensor_dict(self, tensor_dict):
+        """See base class."""
+        return InputExample(
+            tensor_dict["idx"].numpy(),
+            tensor_dict["question"].numpy().decode("utf-8"),
+            tensor_dict["sentence"].numpy().decode("utf-8"),
+            str(tensor_dict["label"].numpy()),
+        )
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev_matched")
+
+    def get_labels(self):
+        """See base class."""
+        return ["entailment", "not_entailment"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, line[0])
+            text_a = line[1]
+            text_b = line[2]
+            label = line[-1]
+            examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+class RteProcessor(DataProcessor):
+    """Processor for the RTE data set (GLUE version)."""
+
+    def get_example_from_tensor_dict(self, tensor_dict):
+        """See base class."""
+        return InputExample(
+            tensor_dict["idx"].numpy(),
+            tensor_dict["sentence1"].numpy().decode("utf-8"),
+            tensor_dict["sentence2"].numpy().decode("utf-8"),
+            str(tensor_dict["label"].numpy()),
+        )
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return ["entailment", "not_entailment"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, line[0])
+            text_a = line[1]
+            text_b = line[2]
+            label = line[-1]
+            examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+class WnliProcessor(DataProcessor):
+    """Processor for the WNLI data set (GLUE version)."""
+
+    def get_example_from_tensor_dict(self, tensor_dict):
+        """See base class."""
+        return InputExample(
+            tensor_dict["idx"].numpy(),
+            tensor_dict["sentence1"].numpy().decode("utf-8"),
+            tensor_dict["sentence2"].numpy().decode("utf-8"),
+            str(tensor_dict["label"].numpy()),
+        )
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, line[0])
+            text_a = line[1]
+            text_b = line[2]
+            label = line[-1]
+            examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+glue_tasks_num_labels = {
+    "cola": 2,
+    "mnli": 3,
+    "mrpc": 2,
+    "sst-2": 2,
+    "sts-b": 1,
+    "qqp": 2,
+    "qnli": 2,
+    "rte": 2,
+    "wnli": 2,
+}
+
+glue_processors = {
+    "cola": ColaProcessor,
+    "mnli": MnliProcessor,
+    "mnli-mm": MnliMismatchedProcessor,
+    "mrpc": MrpcProcessor,
+    "sst-2": Sst2Processor,
+    "sts-b": StsbProcessor,
+    "qqp": QqpProcessor,
+    "qnli": QnliProcessor,
+    "rte": RteProcessor,
+    "wnli": WnliProcessor,
+}
+
+glue_output_modes = {
+    "cola": "classification",
+    "mnli": "classification",
+    "mnli-mm": "classification",
+    "mrpc": "classification",
+    "sst-2": "classification",
+    "sts-b": "regression",
+    "qqp": "classification",
+    "qnli": "classification",
+    "rte": "classification",
+    "wnli": "classification",
+}

src/transformers/data/processors/squad.py

@@ -0,0 +1,697 @@
+import json
+import logging
+import os
+from functools import partial
+from multiprocessing import Pool, cpu_count
+
+import numpy as np
+from tqdm import tqdm
+
+from ...file_utils import is_tf_available, is_torch_available
+from ...tokenization_bert import whitespace_tokenize
+from .utils import DataProcessor
+
+
+if is_torch_available():
+    import torch
+    from torch.utils.data import TensorDataset
+
+if is_tf_available():
+    import tensorflow as tf
+
+logger = logging.getLogger(__name__)
+
+
+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 _check_is_max_context(doc_spans, cur_span_index, position):
+    """Check if this is the 'max context' doc span for the token."""
+    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
+
+
+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
+
+
+def _is_whitespace(c):
+    if c == " " or c == "\t" or c == "\r" or c == "\n" or ord(c) == 0x202F:
+        return True
+    return False
+
+
+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)])
+        cleaned_answer_text = " ".join(whitespace_tokenize(example.answer_text))
+        if actual_text.find(cleaned_answer_text) == -1:
+            logger.warning("Could not find answer: '%s' vs. '%s'", actual_text, cleaned_answer_text)
+            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))
+        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"]:
+            non_padded_ids = encoded_dict["input_ids"][: encoded_dict["input_ids"].index(tokenizer.pad_token_id)]
+        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,
+            )
+        )
+    return features
+
+
+def squad_convert_example_to_features_init(tokenizer_for_convert):
+    global tokenizer
+    tokenizer = tokenizer_for_convert
+
+
+def squad_convert_examples_to_features(
+    examples, tokenizer, max_seq_length, doc_stride, max_query_length, is_training, return_dataset=False, threads=1
+):
+    """
+    Converts a list of examples into a list of features that can be directly given as input to a model.
+    It is model-dependant and takes advantage of many of the tokenizer's features to create the model's inputs.
+
+    Args:
+        examples: list of :class:`~transformers.data.processors.squad.SquadExample`
+        tokenizer: an instance of a child of :class:`~transformers.PreTrainedTokenizer`
+        max_seq_length: The maximum sequence length of the inputs.
+        doc_stride: The stride used when the context is too large and is split across several features.
+        max_query_length: The maximum length of the query.
+        is_training: whether to create features for model evaluation or model training.
+        return_dataset: Default False. Either 'pt' or 'tf'.
+            if 'pt': returns a torch.data.TensorDataset,
+            if 'tf': returns a tf.data.Dataset
+        threads: multiple processing threadsa-smi
+
+
+    Returns:
+        list of :class:`~transformers.data.processors.squad.SquadFeatures`
+
+    Example::
+
+        processor = SquadV2Processor()
+        examples = processor.get_dev_examples(data_dir)
+
+        features = 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,
+        )
+    """
+
+    # 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
+    if return_dataset == "pt":
+        if not is_torch_available():
+            raise ImportError("Pytorch must be installed to return a pytorch dataset.")
+
+        # 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)
+
+        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,
+            )
+
+        return features, dataset
+    elif return_dataset == "tf":
+        if not is_tf_available():
+            raise ImportError("TensorFlow must be installed to return a TensorFlow dataset.")
+
+        def gen():
+            for ex in features:
+                yield (
+                    {
+                        "input_ids": ex.input_ids,
+                        "attention_mask": ex.attention_mask,
+                        "token_type_ids": ex.token_type_ids,
+                    },
+                    {
+                        "start_position": ex.start_position,
+                        "end_position": ex.end_position,
+                        "cls_index": ex.cls_index,
+                        "p_mask": ex.p_mask,
+                    },
+                )
+
+        return tf.data.Dataset.from_generator(
+            gen,
+            (
+                {"input_ids": tf.int32, "attention_mask": tf.int32, "token_type_ids": tf.int32},
+                {"start_position": tf.int64, "end_position": tf.int64, "cls_index": tf.int64, "p_mask": tf.int32},
+            ),
+            (
+                {
+                    "input_ids": tf.TensorShape([None]),
+                    "attention_mask": tf.TensorShape([None]),
+                    "token_type_ids": tf.TensorShape([None]),
+                },
+                {
+                    "start_position": tf.TensorShape([]),
+                    "end_position": tf.TensorShape([]),
+                    "cls_index": tf.TensorShape([]),
+                    "p_mask": tf.TensorShape([None]),
+                },
+            ),
+        )
+
+    return features
+
+
+class SquadProcessor(DataProcessor):
+    """
+    Processor for the SQuAD data set.
+    Overriden by SquadV1Processor and SquadV2Processor, used by the version 1.1 and version 2.0 of SQuAD, respectively.
+    """
+
+    train_file = None
+    dev_file = None
+
+    def _get_example_from_tensor_dict(self, tensor_dict, evaluate=False):
+        if not evaluate:
+            answer = tensor_dict["answers"]["text"][0].numpy().decode("utf-8")
+            answer_start = tensor_dict["answers"]["answer_start"][0].numpy()
+            answers = []
+        else:
+            answers = [
+                {"answer_start": start.numpy(), "text": text.numpy().decode("utf-8")}
+                for start, text in zip(tensor_dict["answers"]["answer_start"], tensor_dict["answers"]["text"])
+            ]
+
+            answer = None
+            answer_start = None
+
+        return SquadExample(
+            qas_id=tensor_dict["id"].numpy().decode("utf-8"),
+            question_text=tensor_dict["question"].numpy().decode("utf-8"),
+            context_text=tensor_dict["context"].numpy().decode("utf-8"),
+            answer_text=answer,
+            start_position_character=answer_start,
+            title=tensor_dict["title"].numpy().decode("utf-8"),
+            answers=answers,
+        )
+
+    def get_examples_from_dataset(self, dataset, evaluate=False):
+        """
+        Creates a list of :class:`~transformers.data.processors.squad.SquadExample` using a TFDS dataset.
+
+        Args:
+            dataset: The tfds dataset loaded from `tensorflow_datasets.load("squad")`
+            evaluate: boolean specifying if in evaluation mode or in training mode
+
+        Returns:
+            List of SquadExample
+
+        Examples::
+
+            import tensorflow_datasets as tfds
+            dataset = tfds.load("squad")
+
+            training_examples = get_examples_from_dataset(dataset, evaluate=False)
+            evaluation_examples = get_examples_from_dataset(dataset, evaluate=True)
+        """
+
+        if evaluate:
+            dataset = dataset["validation"]
+        else:
+            dataset = dataset["train"]
+
+        examples = []
+        for tensor_dict in tqdm(dataset):
+            examples.append(self._get_example_from_tensor_dict(tensor_dict, evaluate=evaluate))
+
+        return examples
+
+    def get_train_examples(self, data_dir, filename=None):
+        """
+        Returns the training examples from the data directory.
+
+        Args:
+            data_dir: Directory containing the data files used for training and evaluating.
+            filename: None by default, specify this if the training file has a different name than the original one
+                which is `train-v1.1.json` and `train-v2.0.json` for squad versions 1.1 and 2.0 respectively.
+
+        """
+        if data_dir is None:
+            data_dir = ""
+
+        if self.train_file is None:
+            raise ValueError("SquadProcessor should be instantiated via SquadV1Processor or SquadV2Processor")
+
+        with open(
+            os.path.join(data_dir, self.train_file if filename is None else filename), "r", encoding="utf-8"
+        ) as reader:
+            input_data = json.load(reader)["data"]
+        return self._create_examples(input_data, "train")
+
+    def get_dev_examples(self, data_dir, filename=None):
+        """
+        Returns the evaluation example from the data directory.
+
+        Args:
+            data_dir: Directory containing the data files used for training and evaluating.
+            filename: None by default, specify this if the evaluation file has a different name than the original one
+                which is `train-v1.1.json` and `train-v2.0.json` for squad versions 1.1 and 2.0 respectively.
+        """
+        if data_dir is None:
+            data_dir = ""
+
+        if self.dev_file is None:
+            raise ValueError("SquadProcessor should be instantiated via SquadV1Processor or SquadV2Processor")
+
+        with open(
+            os.path.join(data_dir, self.dev_file if filename is None else filename), "r", encoding="utf-8"
+        ) as reader:
+            input_data = json.load(reader)["data"]
+        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"]
+            for paragraph in entry["paragraphs"]:
+                context_text = paragraph["context"]
+                for qa in paragraph["qas"]:
+                    qas_id = qa["id"]
+                    question_text = qa["question"]
+                    start_position_character = None
+                    answer_text = None
+                    answers = []
+
+                    if "is_impossible" in qa:
+                        is_impossible = qa["is_impossible"]
+                    else:
+                        is_impossible = False
+
+                    if not is_impossible:
+                        if is_training:
+                            answer = qa["answers"][0]
+                            answer_text = answer["text"]
+                            start_position_character = answer["answer_start"]
+                        else:
+                            answers = qa["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
+
+
+class SquadV1Processor(SquadProcessor):
+    train_file = "train-v1.1.json"
+    dev_file = "dev-v1.1.json"
+
+
+class SquadV2Processor(SquadProcessor):
+    train_file = "train-v2.0.json"
+    dev_file = "dev-v2.0.json"
+
+
+class SquadExample(object):
+    """
+    A single training/test example for the Squad dataset, as loaded from disk.
+
+    Args:
+        qas_id: The example's unique identifier
+        question_text: The question string
+        context_text: The context string
+        answer_text: The answer string
+        start_position_character: The character position of the start of the answer
+        title: The title of the example
+        answers: None by default, this is used during evaluation. Holds answers as well as their start positions.
+        is_impossible: False by default, set to True if the example has no possible answer.
+    """
+
+    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 end 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 SquadFeatures(object):
+    """
+    Single squad example features to be fed to a model.
+    Those features are model-specific and can be crafted from :class:`~transformers.data.processors.squad.SquadExample`
+    using the :method:`~transformers.data.processors.squad.squad_convert_examples_to_features` method.
+
+    Args:
+        input_ids: Indices of input sequence tokens in the vocabulary.
+        attention_mask: Mask to avoid performing attention on padding token indices.
+        token_type_ids: Segment token indices to indicate first and second portions of the inputs.
+        cls_index: the index of the CLS token.
+        p_mask: Mask identifying tokens that can be answers vs. tokens that cannot.
+            Mask with 1 for tokens than cannot be in the answer and 0 for token that can be in an answer
+        example_index: the index of the example
+        unique_id: The unique Feature identifier
+        paragraph_len: The length of the context
+        token_is_max_context: List of booleans identifying which tokens have their maximum context in this feature object.
+            If a token does not have their maximum context in this feature object, it means that another feature object
+            has more information related to that token and should be prioritized over this feature for that token.
+        tokens: list of tokens corresponding to the input ids
+        token_to_orig_map: mapping between the tokens and the original text, needed in order to identify the answer.
+        start_position: start of the answer token index
+        end_position: end of the answer token index
+    """
+
+    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,
+    ):
+        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
+
+
+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

src/transformers/data/processors/utils.py

@@ -0,0 +1,333 @@
+# 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.
+
+import copy
+import csv
+import json
+import logging
+import sys
+
+from ...file_utils import is_tf_available, is_torch_available
+
+
+logger = logging.getLogger(__name__)
+
+
+class InputExample(object):
+    """
+    A single training/test example for simple sequence classification.
+
+    Args:
+        guid: Unique id for the example.
+        text_a: string. The untokenized text of the first sequence. For single
+        sequence tasks, only this sequence must be specified.
+        text_b: (Optional) string. The untokenized text of the second sequence.
+        Only must be specified for sequence pair tasks.
+        label: (Optional) string. The label of the example. This should be
+        specified for train and dev examples, but not for test examples.
+    """
+
+    def __init__(self, guid, text_a, text_b=None, label=None):
+        self.guid = guid
+        self.text_a = text_a
+        self.text_b = text_b
+        self.label = label
+
+    def __repr__(self):
+        return str(self.to_json_string())
+
+    def to_dict(self):
+        """Serializes this instance to a Python dictionary."""
+        output = copy.deepcopy(self.__dict__)
+        return output
+
+    def to_json_string(self):
+        """Serializes this instance to a JSON string."""
+        return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"
+
+
+class InputFeatures(object):
+    """
+    A single set of features of data.
+
+    Args:
+        input_ids: Indices of input sequence tokens in the vocabulary.
+        attention_mask: Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            Usually  ``1`` for tokens that are NOT MASKED, ``0`` for MASKED (padded) tokens.
+        token_type_ids: Segment token indices to indicate first and second portions of the inputs.
+        label: Label corresponding to the input
+    """
+
+    def __init__(self, input_ids, attention_mask=None, token_type_ids=None, label=None):
+        self.input_ids = input_ids
+        self.attention_mask = attention_mask
+        self.token_type_ids = token_type_ids
+        self.label = label
+
+    def __repr__(self):
+        return str(self.to_json_string())
+
+    def to_dict(self):
+        """Serializes this instance to a Python dictionary."""
+        output = copy.deepcopy(self.__dict__)
+        return output
+
+    def to_json_string(self):
+        """Serializes this instance to a JSON string."""
+        return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"
+
+
+class DataProcessor(object):
+    """Base class for data converters for sequence classification data sets."""
+
+    @classmethod
+    def _read_tsv(cls, input_file, quotechar=None):
+        """Reads a tab separated value file."""
+        with open(input_file, "r", encoding="utf-8-sig") as f:
+            reader = csv.reader(f, delimiter="\t", quotechar=quotechar)
+            lines = []
+            for line in reader:
+                if sys.version_info[0] == 2:
+                    line = list(unicode(cell, "utf-8") for cell in line)  # noqa: F821
+                lines.append(line)
+            return lines
+
+
+class SingleSentenceClassificationProcessor(DataProcessor):
+    """ Generic processor for a single sentence classification data set."""
+
+    def __init__(self, labels=None, examples=None, mode="classification", verbose=False):
+        self.labels = [] if labels is None else labels
+        self.examples = [] if examples is None else examples
+        self.mode = mode
+        self.verbose = verbose
+
+    def __len__(self):
+        return len(self.examples)
+
+    def __getitem__(self, idx):
+        if isinstance(idx, slice):
+            return SingleSentenceClassificationProcessor(labels=self.labels, examples=self.examples[idx])
+        return self.examples[idx]
+
+    @classmethod
+    def create_from_csv(
+        cls, file_name, split_name="", column_label=0, column_text=1, column_id=None, skip_first_row=False, **kwargs
+    ):
+        processor = cls(**kwargs)
+        processor.add_examples_from_csv(
+            file_name,
+            split_name=split_name,
+            column_label=column_label,
+            column_text=column_text,
+            column_id=column_id,
+            skip_first_row=skip_first_row,
+            overwrite_labels=True,
+            overwrite_examples=True,
+        )
+        return processor
+
+    @classmethod
+    def create_from_examples(cls, texts_or_text_and_labels, labels=None, **kwargs):
+        processor = cls(**kwargs)
+        processor.add_examples(texts_or_text_and_labels, labels=labels)
+        return processor
+
+    def add_examples_from_csv(
+        self,
+        file_name,
+        split_name="",
+        column_label=0,
+        column_text=1,
+        column_id=None,
+        skip_first_row=False,
+        overwrite_labels=False,
+        overwrite_examples=False,
+    ):
+        lines = self._read_tsv(file_name)
+        if skip_first_row:
+            lines = lines[1:]
+        texts = []
+        labels = []
+        ids = []
+        for (i, line) in enumerate(lines):
+            texts.append(line[column_text])
+            labels.append(line[column_label])
+            if column_id is not None:
+                ids.append(line[column_id])
+            else:
+                guid = "%s-%s" % (split_name, i) if split_name else "%s" % i
+                ids.append(guid)
+
+        return self.add_examples(
+            texts, labels, ids, overwrite_labels=overwrite_labels, overwrite_examples=overwrite_examples
+        )
+
+    def add_examples(
+        self, texts_or_text_and_labels, labels=None, ids=None, overwrite_labels=False, overwrite_examples=False
+    ):
+        assert labels is None or len(texts_or_text_and_labels) == len(labels)
+        assert ids is None or len(texts_or_text_and_labels) == len(ids)
+        if ids is None:
+            ids = [None] * len(texts_or_text_and_labels)
+        if labels is None:
+            labels = [None] * len(texts_or_text_and_labels)
+        examples = []
+        added_labels = set()
+        for (text_or_text_and_label, label, guid) in zip(texts_or_text_and_labels, labels, ids):
+            if isinstance(text_or_text_and_label, (tuple, list)) and label is None:
+                text, label = text_or_text_and_label
+            else:
+                text = text_or_text_and_label
+            added_labels.add(label)
+            examples.append(InputExample(guid=guid, text_a=text, text_b=None, label=label))
+
+        # Update examples
+        if overwrite_examples:
+            self.examples = examples
+        else:
+            self.examples.extend(examples)
+
+        # Update labels
+        if overwrite_labels:
+            self.labels = list(added_labels)
+        else:
+            self.labels = list(set(self.labels).union(added_labels))
+
+        return self.examples
+
+    def get_features(
+        self,
+        tokenizer,
+        max_length=None,
+        pad_on_left=False,
+        pad_token=0,
+        mask_padding_with_zero=True,
+        return_tensors=None,
+    ):
+        """
+        Convert examples in a list of ``InputFeatures``
+
+        Args:
+            tokenizer: Instance of a tokenizer that will tokenize the examples
+            max_length: Maximum example length
+            task: GLUE task
+            label_list: List of labels. Can be obtained from the processor using the ``processor.get_labels()`` method
+            output_mode: String indicating the output mode. Either ``regression`` or ``classification``
+            pad_on_left: If set to ``True``, the examples will be padded on the left rather than on the right (default)
+            pad_token: Padding token
+            mask_padding_with_zero: If set to ``True``, the attention mask will be filled by ``1`` for actual values
+                and by ``0`` for padded values. If set to ``False``, inverts it (``1`` for padded values, ``0`` for
+                actual values)
+
+        Returns:
+            If the ``examples`` input is a ``tf.data.Dataset``, will return a ``tf.data.Dataset``
+            containing the task-specific features. If the input is a list of ``InputExamples``, will return
+            a list of task-specific ``InputFeatures`` which can be fed to the model.
+
+        """
+        if max_length is None:
+            max_length = tokenizer.max_len
+
+        label_map = {label: i for i, label in enumerate(self.labels)}
+
+        all_input_ids = []
+        for (ex_index, example) in enumerate(self.examples):
+            if ex_index % 10000 == 0:
+                logger.info("Tokenizing example %d", ex_index)
+
+            input_ids = tokenizer.encode(
+                example.text_a, add_special_tokens=True, max_length=min(max_length, tokenizer.max_len),
+            )
+            all_input_ids.append(input_ids)
+
+        batch_length = max(len(input_ids) for input_ids in all_input_ids)
+
+        features = []
+        for (ex_index, (input_ids, example)) in enumerate(zip(all_input_ids, self.examples)):
+            if ex_index % 10000 == 0:
+                logger.info("Writing example %d", ex_index)
+            # The mask has 1 for real tokens and 0 for padding tokens. Only real
+            # tokens are attended to.
+            attention_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
+
+            # Zero-pad up to the sequence length.
+            padding_length = batch_length - len(input_ids)
+            if pad_on_left:
+                input_ids = ([pad_token] * padding_length) + input_ids
+                attention_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + attention_mask
+            else:
+                input_ids = input_ids + ([pad_token] * padding_length)
+                attention_mask = attention_mask + ([0 if mask_padding_with_zero else 1] * padding_length)
+
+            assert len(input_ids) == batch_length, "Error with input length {} vs {}".format(
+                len(input_ids), batch_length
+            )
+            assert len(attention_mask) == batch_length, "Error with input length {} vs {}".format(
+                len(attention_mask), batch_length
+            )
+
+            if self.mode == "classification":
+                label = label_map[example.label]
+            elif self.mode == "regression":
+                label = float(example.label)
+            else:
+                raise ValueError(self.mode)
+
+            if ex_index < 5 and self.verbose:
+                logger.info("*** Example ***")
+                logger.info("guid: %s" % (example.guid))
+                logger.info("input_ids: %s" % " ".join([str(x) for x in input_ids]))
+                logger.info("attention_mask: %s" % " ".join([str(x) for x in attention_mask]))
+                logger.info("label: %s (id = %d)" % (example.label, label))
+
+            features.append(InputFeatures(input_ids=input_ids, attention_mask=attention_mask, label=label))
+
+        if return_tensors is None:
+            return features
+        elif return_tensors == "tf":
+            if not is_tf_available():
+                raise ImportError("return_tensors set to 'tf' but TensorFlow 2.0 can't be imported")
+            import tensorflow as tf
+
+            def gen():
+                for ex in features:
+                    yield ({"input_ids": ex.input_ids, "attention_mask": ex.attention_mask}, ex.label)
+
+            dataset = tf.data.Dataset.from_generator(
+                gen,
+                ({"input_ids": tf.int32, "attention_mask": tf.int32}, tf.int64),
+                ({"input_ids": tf.TensorShape([None]), "attention_mask": tf.TensorShape([None])}, tf.TensorShape([])),
+            )
+            return dataset
+        elif return_tensors == "pt":
+            if not is_torch_available():
+                raise ImportError("return_tensors set to 'pt' but PyTorch can't be imported")
+            import torch
+            from torch.utils.data import TensorDataset
+
+            all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
+            all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
+            if self.mode == "classification":
+                all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
+            elif self.mode == "regression":
+                all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
+
+            dataset = TensorDataset(all_input_ids, all_attention_mask, all_labels)
+            return dataset
+        else:
+            raise ValueError("return_tensors should be one of 'tf' or 'pt'")

src/transformers/data/processors/xnli.py

@@ -0,0 +1,86 @@
+# 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.
+""" XNLI utils (dataset loading and evaluation) """
+
+from __future__ import absolute_import, division, print_function
+
+import logging
+import os
+
+from .utils import DataProcessor, InputExample
+
+
+logger = logging.getLogger(__name__)
+
+
+class XnliProcessor(DataProcessor):
+    """Processor for the XNLI dataset.
+    Adapted from https://github.com/google-research/bert/blob/f39e881b169b9d53bea03d2d341b31707a6c052b/run_classifier.py#L207"""
+
+    def __init__(self, language, train_language=None):
+        self.language = language
+        self.train_language = train_language
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        lg = self.language if self.train_language is None else self.train_language
+        lines = self._read_tsv(os.path.join(data_dir, "XNLI-MT-1.0/multinli/multinli.train.{}.tsv".format(lg)))
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % ("train", i)
+            text_a = line[0]
+            text_b = line[1]
+            label = "contradiction" if line[2] == "contradictory" else line[2]
+            assert isinstance(text_a, str) and isinstance(text_b, str) and isinstance(label, str)
+            examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        lines = self._read_tsv(os.path.join(data_dir, "XNLI-1.0/xnli.test.tsv"))
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            language = line[0]
+            if language != self.language:
+                continue
+            guid = "%s-%s" % ("test", i)
+            text_a = line[6]
+            text_b = line[7]
+            label = line[1]
+            assert isinstance(text_a, str) and isinstance(text_b, str) and isinstance(label, str)
+            examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+    def get_labels(self):
+        """See base class."""
+        return ["contradiction", "entailment", "neutral"]
+
+
+xnli_processors = {
+    "xnli": XnliProcessor,
+}
+
+xnli_output_modes = {
+    "xnli": "classification",
+}
+
+xnli_tasks_num_labels = {
+    "xnli": 3,
+}