kserve/python/custom_tokenizer/data_processing.py

# Copyright 2023 The KServe Authors.
#
# 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.

# Modifications copyright (C) 2019 NVIDIA Corp.

import tokenization
import collections
import numpy as np
import six
import math


def convert_doc_tokens(paragraph_text):
    """Return the list of tokens from the doc text"""

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

    doc_tokens = []
    prev_is_whitespace = True
    for c in paragraph_text:
        if is_whitespace(c):
            prev_is_whitespace = True
        else:
            if prev_is_whitespace:
                doc_tokens.append(c)
            else:
                doc_tokens[-1] += c
            prev_is_whitespace = False

    return doc_tokens


def _check_is_max_context(doc_spans, cur_span_index, position):
    """Check if this is the 'max context' doc span for the token."""

    # Because of the sliding window approach taken to scoring documents, a single
    # token can appear in multiple documents. E.g.
    #  Doc: the man went to the store and bought a gallon of milk
    #  Span A: the man went to the
    #  Span B: to the store and bought
    #  Span C: and bought a gallon of
    #  ...
    #
    # Now the word 'bought' will have two scores from spans B and C. We only
    # want to consider the score with "maximum context", which we define as
    # the *minimum* of its left and right context (the *sum* of left and
    # right context will always be the same, of course).
    #
    # In the example the maximum context for 'bought' would be span C since
    # it has 1 left context and 3 right context, while span B has 4 left context
    # and 0 right context.
    best_score = None
    best_span_index = None
    for span_index, doc_span in enumerate(doc_spans):
        end = doc_span.start + doc_span.length - 1
        if position < doc_span.start:
            continue
        if position > end:
            continue
        num_left_context = position - doc_span.start
        num_right_context = end - position
        score = min(num_left_context, num_right_context) + 0.01 * doc_span.length
        if best_score is None or score > best_score:
            best_score = score
            best_span_index = span_index

    return cur_span_index == best_span_index


def convert_examples_to_features(
    doc_tokens, question_text, tokenizer, max_seq_length, doc_stride, max_query_length
):
    """Loads a data file into a list of `InputBatch`s."""

    query_tokens = tokenizer.tokenize(question_text)

    if len(query_tokens) > max_query_length:
        query_tokens = query_tokens[0:max_query_length]

    tok_to_orig_index = []
    orig_to_tok_index = []
    all_doc_tokens = []
    for i, token in enumerate(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)

    # The -3 accounts for [CLS], [SEP] and [SEP]
    max_tokens_for_doc = max_seq_length - len(query_tokens) - 3

    # We can have documents that are longer than the maximum sequence length.
    # To deal with this we do a sliding window approach, where we take chunks
    # of the up to our max length with a stride of `doc_stride`.
    _DocSpan = collections.namedtuple(  # pylint: disable=invalid-name
        "DocSpan", ["start", "length"]
    )
    doc_spans = []
    start_offset = 0
    while start_offset < len(all_doc_tokens):
        length = len(all_doc_tokens) - start_offset
        if length > max_tokens_for_doc:
            length = max_tokens_for_doc
        doc_spans.append(_DocSpan(start=start_offset, length=length))
        if start_offset + length == len(all_doc_tokens):
            break
        start_offset += min(length, doc_stride)

    for doc_span_index, doc_span in enumerate(doc_spans):
        tokens = []
        token_to_orig_map = {}
        token_is_max_context = {}
        segment_ids = []
        tokens.append("[CLS]")
        segment_ids.append(0)
        for token in query_tokens:
            tokens.append(token)
            segment_ids.append(0)
        tokens.append("[SEP]")
        segment_ids.append(0)

        for i in range(doc_span.length):
            split_token_index = doc_span.start + i
            token_to_orig_map[len(tokens)] = tok_to_orig_index[split_token_index]

            is_max_context = _check_is_max_context(
                doc_spans, doc_span_index, split_token_index
            )
            token_is_max_context[len(tokens)] = is_max_context
            tokens.append(all_doc_tokens[split_token_index])
            segment_ids.append(1)
        tokens.append("[SEP]")
        segment_ids.append(1)

    input_ids = tokenizer.convert_tokens_to_ids(tokens)

    # The mask has 1 for real tokens and 0 for padding tokens. Only real
    # tokens are attended to.
    input_mask = [1] * len(input_ids)

    # Zero-pad up to the sequence length.
    while len(input_ids) < max_seq_length:
        input_ids.append(0)
        input_mask.append(0)
        segment_ids.append(0)

    assert len(input_ids) == max_seq_length
    assert len(input_mask) == max_seq_length
    assert len(segment_ids) == max_seq_length

    def create_int_feature(values):
        feature = np.asarray(values, dtype=np.int32, order=None)
        return feature

    features = collections.OrderedDict()
    features["input_ids"] = create_int_feature(input_ids)
    features["input_mask"] = create_int_feature(input_mask)
    features["segment_ids"] = create_int_feature(segment_ids)
    features["tokens"] = tokens
    features["token_to_orig_map"] = token_to_orig_map
    features["token_is_max_context"] = token_is_max_context
    return features


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 get_final_text(pred_text, orig_text, do_lower_case):
    """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 = tokenization.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:
        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):
        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 six.iteritems(tok_ns_to_s_map):
        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:
        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:
        return orig_text

    output_text = orig_text[orig_start_position : (orig_end_position + 1)]
    return output_text


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 get_predictions(
    doc_tokens, features, start_logits, end_logits, n_best_size, max_answer_length
):
    _PrelimPrediction = collections.namedtuple(  # pylint: disable=invalid-name
        "PrelimPrediction",
        ["feature_index", "start_index", "end_index", "start_logit", "end_logit"],
    )

    prediction = ""
    scores_diff_json = 0.0

    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 mull 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

    start_indexes = _get_best_indexes(start_logits, n_best_size)
    end_indexes = _get_best_indexes(end_logits, n_best_size)

    # if we could have irrelevant answers, get the min score of irrelevant
    version_2_with_negative = True
    if version_2_with_negative:
        feature_null_score = start_logits[0] + end_logits[0]
        if feature_null_score < score_null:
            score_null = feature_null_score
            min_null_feature_index = 0
            null_start_logit = start_logits[0]
            null_end_logit = 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(features["tokens"]):
                continue
            if end_index >= len(features["tokens"]):
                continue
            if start_index not in features["token_to_orig_map"]:
                continue
            if end_index not in features["token_to_orig_map"]:
                continue
            if not features["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=0,
                    start_index=start_index,
                    end_index=end_index,
                    start_logit=start_logits[start_index],
                    end_logit=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

        if pred.start_index > 0:  # this is a non-null prediction
            tok_tokens = features["tokens"][pred.start_index : (pred.end_index + 1)]
            orig_doc_start = features["token_to_orig_map"][pred.start_index]
            orig_doc_end = features["token_to_orig_map"][pred.end_index]
            orig_tokens = doc_tokens[orig_doc_start : (orig_doc_end + 1)]
            tok_text = " ".join(tok_tokens)

            # De-tokenize WordPieces that have been split off.
            tok_text = tok_text.replace(" ##", "")
            tok_text = tok_text.replace("##", "")

            # Clean whitespace
            tok_text = tok_text.strip()
            tok_text = " ".join(tok_text.split())
            orig_text = " ".join(orig_tokens)

            final_text = get_final_text(tok_text, orig_text, True)
            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 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

    null_score_diff_threshold = 0.0
    if not version_2_with_negative:
        prediction = 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 = score_diff
        if score_diff > null_score_diff_threshold:
            prediction = ""
        else:
            prediction = best_non_null_entry.text

    return prediction, nbest_json, scores_diff_json