src/autotrain/trainers/extractive_question_answering/utils.py

import collections
import json
import os

import numpy as np
from datasets import load_metric
from transformers import EvalPrediction

from autotrain import logger


MODEL_CARD = """
---
library_name: transformers
tags:
- autotrain
- question-answering{base_model}
widget:
- text: "Who loves AutoTrain?"
  context: "Everyone loves AutoTrain"{dataset_tag}
---

# Model Trained Using AutoTrain

- Problem type: Extractive Question Answering

## Validation Metrics

{validation_metrics}

## Usage


```python
import torch

from transformers import AutoModelForQuestionAnswering, AutoTokenizer

model = AutoModelForQuestionAnswering.from_pretrained(...)

tokenizer = AutoTokenizer.from_pretrained(...)

from transformers import BertTokenizer, BertForQuestionAnswering

question, text = "Who loves AutoTrain?", "Everyone loves AutoTrain"

inputs = tokenizer(question, text, return_tensors='pt')

start_positions = torch.tensor([1])

end_positions = torch.tensor([3])

outputs = model(**inputs, start_positions=start_positions, end_positions=end_positions)

loss = outputs.loss

start_scores = outputs.start_logits

end_scores = outputs.end_logits
```
"""

SQUAD_METRIC = load_metric("squad")
SQUAD_V2_METRIC = load_metric("squad_v2")


def postprocess_qa_predictions(
    examples,
    features,
    predictions,
    config,
    version_2_with_negative=False,
    n_best_size=20,
    max_answer_length=30,
    null_score_diff_threshold=0.0,
    output_dir=None,
    prefix=None,
):
    # This function is taken from: https://github.com/huggingface/transformers/blob/dcec4c4387850dff8123d5752aab8c1b5431465b/examples/pytorch/question-answering/run_qa.py#L470
    """
    Post-processes the predictions of a question-answering model to convert them to answers that are substrings of the
    original contexts. This is the base postprocessing functions for models that only return start and end logits.

    Args:
        examples: The non-preprocessed dataset (see the main script for more information).
        features: The processed dataset (see the main script for more information).
        predictions (:obj:`Tuple[np.ndarray, np.ndarray]`):
            The predictions of the model: two arrays containing the start logits and the end logits respectively. Its
            first dimension must match the number of elements of :obj:`features`.
        version_2_with_negative (:obj:`bool`, `optional`, defaults to :obj:`False`):
            Whether or not the underlying dataset contains examples with no answers.
        n_best_size (:obj:`int`, `optional`, defaults to 20):
            The total number of n-best predictions to generate when looking for an answer.
        max_answer_length (:obj:`int`, `optional`, defaults to 30):
            The maximum length of an answer that can be generated. This is needed because the start and end predictions
            are not conditioned on one another.
        null_score_diff_threshold (:obj:`float`, `optional`, defaults to 0):
            The threshold used to select the null answer: if the best answer has a score that is less than the score of
            the null answer minus this threshold, the null answer is selected for this example (note that the score of
            the null answer for an example giving several features is the minimum of the scores for the null answer on
            each feature: all features must be aligned on the fact they `want` to predict a null answer).

            Only useful when :obj:`version_2_with_negative` is :obj:`True`.
        output_dir (:obj:`str`, `optional`):
            If provided, the dictionaries of predictions, n_best predictions (with their scores and logits) and, if
            :obj:`version_2_with_negative=True`, the dictionary of the scores differences between best and null
            answers, are saved in `output_dir`.
        prefix (:obj:`str`, `optional`):
            If provided, the dictionaries mentioned above are saved with `prefix` added to their names.
        log_level (:obj:`int`, `optional`, defaults to ``logging.WARNING``):
            ``logging`` log level (e.g., ``logging.WARNING``)
    """
    if len(predictions) != 2:
        raise ValueError("`predictions` should be a tuple with two elements (start_logits, end_logits).")
    all_start_logits, all_end_logits = predictions

    if len(predictions[0]) != len(features):
        raise ValueError(f"Got {len(predictions[0])} predictions and {len(features)} features.")

    # Build a map example to its corresponding features.
    example_id_to_index = {k: i for i, k in enumerate(examples["id"])}
    features_per_example = collections.defaultdict(list)
    for i, feature in enumerate(features):
        features_per_example[example_id_to_index[feature["example_id"]]].append(i)

    # The dictionaries we have to fill.
    all_predictions = collections.OrderedDict()
    all_nbest_json = collections.OrderedDict()
    if version_2_with_negative:
        scores_diff_json = collections.OrderedDict()

    # Logging.
    logger.info(f"Post-processing {len(examples)} example predictions split into {len(features)} features.")

    # Let's loop over all the examples!
    for example_index, example in enumerate(examples):
        # Those are the indices of the features associated to the current example.
        feature_indices = features_per_example[example_index]

        min_null_prediction = None
        prelim_predictions = []

        # Looping through all the features associated to the current example.
        for feature_index in feature_indices:
            # We grab the predictions of the model for this feature.
            start_logits = all_start_logits[feature_index]
            end_logits = all_end_logits[feature_index]
            # This is what will allow us to map some the positions in our logits to span of texts in the original
            # context.
            offset_mapping = features[feature_index]["offset_mapping"]
            # Optional `token_is_max_context`, if provided we will remove answers that do not have the maximum context
            # available in the current feature.
            token_is_max_context = features[feature_index].get("token_is_max_context", None)

            # Update minimum null prediction.
            feature_null_score = start_logits[0] + end_logits[0]
            if min_null_prediction is None or min_null_prediction["score"] > feature_null_score:
                min_null_prediction = {
                    "offsets": (0, 0),
                    "score": feature_null_score,
                    "start_logit": start_logits[0],
                    "end_logit": end_logits[0],
                }

            # Go through all possibilities for the `n_best_size` greater start and end logits.
            start_indexes = np.argsort(start_logits)[-1 : -n_best_size - 1 : -1].tolist()
            end_indexes = np.argsort(end_logits)[-1 : -n_best_size - 1 : -1].tolist()
            for start_index in start_indexes:
                for end_index in end_indexes:
                    # Don't consider out-of-scope answers, either because the indices are out of bounds or correspond
                    # to part of the input_ids that are not in the context.
                    if (
                        start_index >= len(offset_mapping)
                        or end_index >= len(offset_mapping)
                        or offset_mapping[start_index] is None
                        or len(offset_mapping[start_index]) < 2
                        or offset_mapping[end_index] is None
                        or len(offset_mapping[end_index]) < 2
                    ):
                        continue
                    # Don't consider answers with a length that is either < 0 or > max_answer_length.
                    if end_index < start_index or end_index - start_index + 1 > max_answer_length:
                        continue
                    # Don't consider answer that don't have the maximum context available (if such information is
                    # provided).
                    if token_is_max_context is not None and not token_is_max_context.get(str(start_index), False):
                        continue

                    prelim_predictions.append(
                        {
                            "offsets": (offset_mapping[start_index][0], offset_mapping[end_index][1]),
                            "score": start_logits[start_index] + end_logits[end_index],
                            "start_logit": start_logits[start_index],
                            "end_logit": end_logits[end_index],
                        }
                    )
        if version_2_with_negative and min_null_prediction is not None:
            # Add the minimum null prediction
            prelim_predictions.append(min_null_prediction)
            null_score = min_null_prediction["score"]

        # Only keep the best `n_best_size` predictions.
        predictions = sorted(prelim_predictions, key=lambda x: x["score"], reverse=True)[:n_best_size]

        # Add back the minimum null prediction if it was removed because of its low score.
        if (
            version_2_with_negative
            and min_null_prediction is not None
            and not any(p["offsets"] == (0, 0) for p in predictions)
        ):
            predictions.append(min_null_prediction)

        # Use the offsets to gather the answer text in the original context.
        context = example[config.text_column]
        for pred in predictions:
            offsets = pred.pop("offsets")
            pred["text"] = context[offsets[0] : offsets[1]]

        # In the very rare edge case we have not a single non-null prediction, we create a fake prediction to avoid
        # failure.
        if len(predictions) == 0 or (len(predictions) == 1 and predictions[0]["text"] == ""):
            predictions.insert(0, {"text": "empty", "start_logit": 0.0, "end_logit": 0.0, "score": 0.0})

        # Compute the softmax of all scores (we do it with numpy to stay independent from torch/tf in this file, using
        # the LogSumExp trick).
        scores = np.array([pred.pop("score") for pred in predictions])
        exp_scores = np.exp(scores - np.max(scores))
        probs = exp_scores / exp_scores.sum()

        # Include the probabilities in our predictions.
        for prob, pred in zip(probs, predictions):
            pred["probability"] = prob

        # Pick the best prediction. If the null answer is not possible, this is easy.
        if not version_2_with_negative:
            all_predictions[example["id"]] = predictions[0]["text"]
        else:
            # Otherwise we first need to find the best non-empty prediction.
            i = 0
            while predictions[i]["text"] == "":
                i += 1
            best_non_null_pred = predictions[i]

            # Then we compare to the null prediction using the threshold.
            score_diff = null_score - best_non_null_pred["start_logit"] - best_non_null_pred["end_logit"]
            scores_diff_json[example["id"]] = float(score_diff)  # To be JSON-serializable.
            if score_diff > null_score_diff_threshold:
                all_predictions[example["id"]] = ""
            else:
                all_predictions[example["id"]] = best_non_null_pred["text"]

        # Make `predictions` JSON-serializable by casting np.float back to float.
        all_nbest_json[example["id"]] = [
            {k: (float(v) if isinstance(v, (np.float16, np.float32, np.float64)) else v) for k, v in pred.items()}
            for pred in predictions
        ]

    # If we have an output_dir, let's save all those dicts.
    if output_dir is not None:
        if not os.path.isdir(output_dir):
            raise EnvironmentError(f"{output_dir} is not a directory.")

        prediction_file = os.path.join(
            output_dir, "predictions.json" if prefix is None else f"{prefix}_predictions.json"
        )
        nbest_file = os.path.join(
            output_dir, "nbest_predictions.json" if prefix is None else f"{prefix}_nbest_predictions.json"
        )
        if version_2_with_negative:
            null_odds_file = os.path.join(
                output_dir, "null_odds.json" if prefix is None else f"{prefix}_null_odds.json"
            )

        logger.info(f"Saving predictions to {prediction_file}.")
        with open(prediction_file, "w") as writer:
            writer.write(json.dumps(all_predictions, indent=4) + "\n")
        logger.info(f"Saving nbest_preds to {nbest_file}.")
        with open(nbest_file, "w") as writer:
            writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
        if version_2_with_negative:
            logger.info(f"Saving null_odds to {null_odds_file}.")
            with open(null_odds_file, "w") as writer:
                writer.write(json.dumps(scores_diff_json, indent=4) + "\n")

    return all_predictions


def post_processing_function_qa(examples, features, predictions, version_2_with_negative, config, stage="eval"):
    # Post-processing: we match the start logits and end logits to answers in the original context.
    predictions = postprocess_qa_predictions(
        examples=examples,
        features=features,
        predictions=predictions,
        version_2_with_negative=version_2_with_negative,
        n_best_size=20,
        max_answer_length=30,
        null_score_diff_threshold=0.0,
        output_dir=None,
        prefix=stage,
        config=config,
    )
    # Format the result to the format the metric expects.
    if version_2_with_negative:
        formatted_predictions = [
            {"id": k, "prediction_text": v, "no_answer_probability": 0.0} for k, v in predictions.items()
        ]
    else:
        formatted_predictions = [{"id": k, "prediction_text": v} for k, v in predictions.items()]

    references = [{"id": str(ex["id"]), "answers": ex[config.answer_column]} for ex in examples]
    return EvalPrediction(predictions=formatted_predictions, label_ids=references)


def compute_metrics(pred, eval_dataset, eval_examples, use_v2, config):
    preds, label_ids = post_processing_function_qa(eval_examples, eval_dataset, pred.predictions, use_v2, config)

    if use_v2:
        result = SQUAD_V2_METRIC.compute(predictions=preds, references=label_ids)
    else:
        result = SQUAD_METRIC.compute(predictions=preds, references=label_ids)
    return {k: round(v, 4) for k, v in result.items()}


def create_model_card(config, trainer):
    if config.valid_split is not None:
        eval_scores = trainer.evaluate()
        eval_scores = [f"{k[len('eval_'):]}: {v}" for k, v in eval_scores.items()]
        eval_scores = "\n\n".join(eval_scores)

    else:
        eval_scores = "No validation metrics available"

    if config.data_path == f"{config.project_name}/autotrain-data" or os.path.isdir(config.data_path):
        dataset_tag = ""
    else:
        dataset_tag = f"\ndatasets:\n- {config.data_path}"

    if os.path.isdir(config.model):
        base_model = ""
    else:
        base_model = f"\nbase_model: {config.model}"

    model_card = MODEL_CARD.format(
        dataset_tag=dataset_tag,
        validation_metrics=eval_scores,
        base_model=base_model,
    )
    return model_card


def prepare_qa_validation_features(examples, tokenizer, config):
    # Some of the questions have lots of whitespace on the left, which is not useful and will make the
    # truncation of the context fail (the tokenized question will take a lots of space). So we remove that
    # left whitespace
    pad_on_right = tokenizer.padding_side == "right"
    examples[config.question_column] = [q.lstrip() for q in examples[config.question_column]]

    # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
    # in one example possible giving several features when a context is long, each of those features having a
    # context that overlaps a bit the context of the previous feature.
    tokenized_examples = tokenizer(
        examples[config.question_column if pad_on_right else config.text_column],
        examples[config.text_column if pad_on_right else config.question_column],
        truncation="only_second" if pad_on_right else "only_first",
        max_length=config.max_seq_length,
        stride=config.max_doc_stride,
        return_overflowing_tokens=True,
        return_offsets_mapping=True,
        padding="max_length",
    )

    # Since one example might give us several features if it has a long context, we need a map from a feature to
    # its corresponding example. This key gives us just that.
    sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")

    # For evaluation, we will need to convert our predictions to substrings of the context, so we keep the
    # corresponding example_id and we will store the offset mappings.
    tokenized_examples["example_id"] = []

    for i in range(len(tokenized_examples["input_ids"])):
        # Grab the sequence corresponding to that example (to know what is the context and what is the question).
        sequence_ids = tokenized_examples.sequence_ids(i)
        context_index = 1 if pad_on_right else 0

        # One example can give several spans, this is the index of the example containing this span of text.
        sample_index = sample_mapping[i]
        tokenized_examples["example_id"].append(examples["id"][sample_index])

        # Set to None the offset_mapping that are not part of the context so it's easy to determine if a token
        # position is part of the context or not.
        tokenized_examples["offset_mapping"][i] = [
            (o if sequence_ids[k] == context_index else None)
            for k, o in enumerate(tokenized_examples["offset_mapping"][i])
        ]

    return tokenized_examples