Automatic Speech Recognition
Transformers
Safetensors
Japanese
whisper
audio
hf-asr-leaderboard
Eval Results
Inference Endpoints
File size: 6,394 Bytes
45ce8e0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
158
159
160
161
162
163
164
165
"""Initialize a student Whisper model from a pre-trained teacher model for teacher-student distillation."""
import argparse
import copy
import logging
import os

import numpy as np
import torch
from transformers import GenerationConfig, WhisperForConditionalGeneration, WhisperProcessor

# https://stackoverflow.com/questions/71692354/facing-ssl-error-with-huggingface-pretrained-models
os.environ['CURL_CA_BUNDLE'] = ''

# disable warning message
os.environ['TOKENIZERS_PARALLELISM'] = 'false'


logger = logging.getLogger(__name__)


def parse_args():
    parser = argparse.ArgumentParser(
        description="Initialise a student Whisper model from a teacher model, copying the relevant layer weights and adjusting the processor as necessary."
    )
    parser.add_argument(
        "--teacher_checkpoint",
        type=str,
        required=True,
        help="The HF Hub ID of the teacher checkpoint.",
    )
    parser.add_argument(
        "--encoder_layers",
        type=int,
        default=None,
        help="Number of encoder layers to use in the student model. Defaults to all layers from the teacher.",
    )
    parser.add_argument(
        "--decoder_layers",
        type=int,
        default=2,
        help="Number of decoder layers to use in the student model. Defaults to 2 layers.",
    )
    parser.add_argument(
        "--save_dir",
        type=str,
        required=True,
        help="Where to save the student weights and processor.",
    )
    args = parser.parse_args()
    return args


def init_student_model_from_teacher(
    teacher_checkpoint,
    save_dir,
    encoder_layers=None,
    decoder_layers=2,
):
    teacher_model = WhisperForConditionalGeneration.from_pretrained(
        teacher_checkpoint,
        low_cpu_mem_usage=True,
    )
    processor = WhisperProcessor.from_pretrained(teacher_checkpoint)
    generation_config = GenerationConfig.from_pretrained(teacher_checkpoint)

    teacher_config = teacher_model.config
    teacher_encoder_layers = teacher_config.encoder_layers
    teacher_decoder_layers = teacher_config.decoder_layers

    student_config = copy.deepcopy(teacher_config)
    student_config.update(
        {
            "encoder_layers": encoder_layers if encoder_layers is not None else teacher_encoder_layers,
            "decoder_layers": decoder_layers,
        }
    )

    encoder_mapping = np.linspace(0, teacher_encoder_layers - 1, student_config.encoder_layers, dtype=int)
    encoder_mapping[-1] = teacher_encoder_layers - 1

    encoder_map = {}
    for student_layer, teacher_layer in enumerate(encoder_mapping):
        encoder_map[teacher_layer] = student_layer

    decoder_mapping = np.linspace(0, teacher_decoder_layers - 1, student_config.decoder_layers, dtype=int)
    decoder_mapping[-1] = teacher_decoder_layers - 1

    decoder_map = {}
    for student_layer, teacher_layer in enumerate(decoder_mapping):
        decoder_map[teacher_layer] = student_layer

    # init the student params from the teacher model
    student_model = WhisperForConditionalGeneration(student_config)
    missing_keys, unexpected_keys = student_model.load_state_dict(teacher_model.state_dict(), strict=False)
    if len(missing_keys) > 0:
        raise RuntimeError(
            "Error(s) in loading state_dict for WhisperForConditionalGeneration. \n"
            f"Missing key(s) in state_dict: {missing_keys}"
        )
    if decoder_layers == teacher_decoder_layers:
        decoder_keys = [key for key in unexpected_keys if "model.decoder.layers" in key]
        if len(decoder_keys) > 0:
            raise RuntimeError(
                "Error(s) in loading state_dict for WhisperForConditionalGeneration. \n"
                f"Unexpected key(s) in state_dict: {decoder_keys}"
            )
    if encoder_layers == teacher_encoder_layers:
        encoder_keys = [key for key in unexpected_keys if "model.encoder.layers" in key]
        if len(encoder_keys) > 0:
            raise RuntimeError(
                "Error(s) in loading state_dict for WhisperForConditionalGeneration. \n"
                f"Unexpected key(s) in state_dict: {encoder_keys}"
            )

    for layer in range(teacher_decoder_layers):
        if layer in decoder_map:
            # re-introduce pre-defined layers from the teacher
            student_model.model.decoder.layers[decoder_map[layer]].load_state_dict(
                teacher_model.model.decoder.layers[layer].state_dict()
            )

    if encoder_layers is not None:
        for layer in range(teacher_encoder_layers):
            if layer in encoder_map:
                # re-introduce pre-defined layers from the teacher
                student_model.model.encoder.layers[encoder_map[layer]].load_state_dict(
                    teacher_model.model.encoder.layers[layer].state_dict()
                )

    # remove the teacher params and model
    del teacher_model

    # save the converted weights and model
    student_model.save_pretrained(save_dir)
    # we also need to correctly save the processor and generation config
    processor.save_pretrained(save_dir)
    generation_config.save_pretrained(save_dir)

    # check we can do a forward pass with the saved model - first load the weights and processor
    logger.info("Checking we can load the saved model...")
    student_model = WhisperForConditionalGeneration.from_pretrained(save_dir, low_cpu_mem_usage=True)
    processor = WhisperProcessor.from_pretrained(save_dir)

    # define some random inputs
    input_features = processor(np.ones(16000), sampling_rate=16000, return_tensors="pt").input_features
    decoder_start_token_id = student_model.config.decoder_start_token_id
    decoder_input_ids = torch.ones((input_features.shape[0], 1), dtype=torch.long) * decoder_start_token_id

    # do a forward pass - outputs will be gibberish for the initialised model so we can't check them
    # but we make can sure the model runs as expected
    logger.info("Checking we can run the converted model forward...")
    _ = student_model(input_features, decoder_input_ids=decoder_input_ids).logits
    logger.info("Conversion successful!")


if __name__ == "__main__":
    args = parse_args()

    init_student_model_from_teacher(
        teacher_checkpoint=args.teacher_checkpoint,
        encoder_layers=args.encoder_layers,
        decoder_layers=args.decoder_layers,
        save_dir=args.save_dir,
    )