model_merging/data.py at master · jacob-morrison/model_merging · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
"""Code for loading data, focusing on GLUE."""
import numpy as np
import tensorflow as tf
import tensorflow_datasets as tfds
from transformers.data.processors import glue as hf_glue


_glue_processors = hf_glue.glue_processors
_glue_output_modes = hf_glue.glue_output_modes


# Despite what the paper says, STS-B starts at 0, not 1.
_STSB_MIN = 0
_STSB_MAX = 5
# Corresponds to rounding to nearest 0.2 increment.
_STSB_NUM_BINS = 5 * (_STSB_MAX - _STSB_MIN)


def _to_tfds_task_name(task, split):
    if task == "sts-b":
        task = "stsb"
    elif task == "sst-2":
        task = "sst2"
    elif task == "mnli" and split != "train":
        task = "mnli_matched"
    elif task == "mnli-mm" and split != "train":
        task = "mnli_mismatched"
    return task


def _convert_dataset_to_features(
    dataset,
    tokenizer,
    max_length,
    task,
):
    """Note that this is only for single examples; won't work with batched inputs."""
    pad_token = tokenizer.pad_token_id
    # NOTE: Not sure if this is correct, but it matches up for BERT. RoBERTa does
    # not appear to use token types.
    pad_token_segment_id = tokenizer.pad_token_type_id

    processor = _glue_processors[task]()
    output_mode = _glue_output_modes[task]
    # print('output mode:')
    # print(output_mode)

    if task == "sts-b":
        # STS-B regression.
        stsb_bins = np.linspace(_STSB_MIN, _STSB_MAX, num=_STSB_NUM_BINS + 1)
        # print('stsb bins:')
        # print(stsb_bins)
        stsb_bins = stsb_bins[1:-1]
        # print(stsb_bins)
    else:
        label_list = processor.get_labels()
        label_map = {label: i for i, label in enumerate(label_list)}

    def py_map_fn(keys, *values):
        example = {tf.compat.as_str(k.numpy()): v for k, v in zip(keys, values)}
        example = processor.get_example_from_tensor_dict(example)
        example = processor.tfds_map(example)

        # print(example.text_a)
        # print(example.text_b)
        # print()
        inputs = tokenizer.encode_plus(
            example.text_a,
            example.text_b,
            add_special_tokens=True,
            max_length=max_length,
            return_token_type_ids=True,
            truncation=True,
        )
        input_ids = inputs["input_ids"]
        token_type_ids = inputs["token_type_ids"]
        attention_mask = inputs["attention_mask"]

        input_ids = tf.constant(input_ids, dtype=tf.int32)
        token_type_ids = tf.constant(token_type_ids, dtype=tf.int32)
        attention_mask = tf.constant(attention_mask, dtype=tf.int32)

        if output_mode == "classification":
            label = label_map[example.label]
            label = tf.constant(label, dtype=tf.int64)
        else:
            # print('label:')
            label = float(example.label)
            # print(label)
            assert 0.0 <= label <= 5.0, f"Out of range STS-B label {label}."
            label = np.digitize(label, stsb_bins)
            # print(label)
            label = tf.constant(label, dtype=tf.int64)
            # print(label)
        # print(example.text_a)
        # print(example.text_b)
        # print(output_mode)
        # print(input_ids)
        # print(token_type_ids)
        # print(label)
        # print()
        # print(None.length)
        return input_ids, token_type_ids, attention_mask, label

    def map_fn(example):
        input_ids, token_type_ids, attention_mask, label = tf.py_function(
            func=py_map_fn,
            inp=[list(example.keys()), *example.values()],
            Tout=[tf.int32, tf.int32, tf.int32, tf.int64],
        )
        return input_ids, token_type_ids, attention_mask, label

    def pad_fn(input_ids, token_type_ids, attention_mask, label):
        # Zero-pad up to the sequence length.
        padding_length = max_length - tf.shape(input_ids)[-1]

        input_ids = tf.concat(
            [input_ids, pad_token * tf.ones(padding_length, dtype=tf.int32)], axis=-1
        )
        token_type_ids = tf.concat(
            [
                token_type_ids,
                pad_token_segment_id * tf.ones(padding_length, dtype=tf.int32),
            ],
            axis=-1,
        )
        attention_mask = tf.concat(
            [
                attention_mask,
                0 * tf.zeros(padding_length, dtype=tf.int32)
            ],
            axis=-1,
        )

        tf_example = {
            # Ensure the shape is known as this is often needed for downstream steps.
            "input_ids": tf.reshape(input_ids, [max_length]),
            "token_type_ids": tf.reshape(token_type_ids, [max_length]),
            "attention_mask": tf.reshape(attention_mask, [max_length]),
        }
        return tf_example, label

    dataset = dataset.map(map_fn)
    dataset = dataset.map(pad_fn)
    return dataset


def load_glue_dataset(task: str, split: str, tokenizer, max_length: int):
    tfds_task = _to_tfds_task_name(task, split)
    ds = tfds.load(f"glue/{tfds_task}", split=split)
    ds = _convert_dataset_to_features(
        ds,
        tokenizer,
        max_length,
        task,
    )
    return ds