File size: 14,868 Bytes
1938217
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
import copy
import torch
import transformers
import tokenizers

from typing import Dict, Sequence

from ola.constants import IGNORE_INDEX, DEFAULT_SPEECH_TOKEN, IMAGE_TOKEN_INDEX
from ola import conversation as conversation_lib
from ola.model import *
from ola.arguments import DataArguments
from ola.constants import SPEECH_TOKEN_INDEX

from packaging import version

IS_TOKENIZER_GREATER_THAN_0_14 = version.parse(tokenizers.__version__) >= version.parse('0.14')


def tokenizer_speech_token(prompt, tokenizer, speech_token_index=SPEECH_TOKEN_INDEX, return_tensors=None):
    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split('<speech>')]

    def insert_separator(X, sep):
        return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]

    input_ids = []
    offset = 0
    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
        offset = 1
        input_ids.append(prompt_chunks[0][0])

    for x in insert_separator(prompt_chunks, [speech_token_index] * (offset + 1)):
        input_ids.extend(x[offset:])

    if return_tensors is not None:
        if return_tensors == 'pt':
            return torch.tensor(input_ids, dtype=torch.long)
        raise ValueError(f'Unsupported tensor type: {return_tensors}')
    return input_ids


def preprocess_multimodal(
    sources: Sequence[str],
    data_args: DataArguments
) -> Dict:
    is_multimodal = data_args.is_multimodal
    if not is_multimodal:
        return sources

    for source in sources:
        for sentence in source:
            if DEFAULT_SPEECH_TOKEN in sentence['value']:
                sentence['value'] = sentence['value'].replace(DEFAULT_SPEECH_TOKEN, '').strip()
                sentence['value'] = DEFAULT_SPEECH_TOKEN + '\n' + sentence['value']
                sentence['value'] = sentence['value'].strip()

    return sources

def tokenizer_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):
    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split('<image>')]

    def insert_separator(X, sep):
        return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]

    input_ids = []
    offset = 0
    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
        offset = 1
        input_ids.append(prompt_chunks[0][0])

    for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
        input_ids.extend(x[offset:])

    if return_tensors is not None:
        if return_tensors == 'pt':
            return torch.tensor(input_ids, dtype=torch.long)
        raise ValueError(f'Unsupported tensor type: {return_tensors}')
    return input_ids

def tokenizer_speech_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, speech_token_idx=SPEECH_TOKEN_INDEX, return_tensors=None):
    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split('<speech><image>')]

    def insert_separator(X, sep):
        return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]

    input_ids = []
    offset = 0
    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
        offset = 1
        input_ids.append(prompt_chunks[0][0])

    for x in insert_separator(prompt_chunks, [speech_token_idx, image_token_index] * (offset + 1)):
        input_ids.extend(x[offset:])

    if return_tensors is not None:
        if return_tensors == 'pt':
            return torch.tensor(input_ids, dtype=torch.long)
        raise ValueError(f'Unsupported tensor type: {return_tensors}')
    return input_ids

def tokenizer_speech_question_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, speech_token_idx=SPEECH_TOKEN_INDEX, return_tensors=None):
    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split("<image>\nUser's question in speech: <speech>\n")]

    def insert_separator(X, sep):
        return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]

    input_ids = []
    offset = 0
    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
        offset = 1
        input_ids.append(prompt_chunks[0][0])

    nl_tokens = tokenizer("\n").input_ids[0]
    special_chunks = [image_token_index, nl_tokens]
    special_chunks.extend(tokenizer("User's question in speech: ").input_ids)
    special_chunks.extend([speech_token_idx, nl_tokens])

    for x in insert_separator(prompt_chunks, special_chunks):
        input_ids.extend(x[offset:])

    if return_tensors is not None:
        if return_tensors == 'pt':
            return torch.tensor(input_ids, dtype=torch.long)
        raise ValueError(f'Unsupported tensor type: {return_tensors}')
    return input_ids

def preprocess_llama_2(
    sources,
    tokenizer: transformers.PreTrainedTokenizer,
    has_speech: bool = False
) -> Dict:
    conv = conversation_lib.default_conversation.copy()
    roles = {"human": conv.roles[0], "gpt": conv.roles[1]}

    # Apply prompt templates
    conversations = []
    for i, source in enumerate(sources):
        if roles[source[0]["from"]] != conv.roles[0]:
            # Skip the first one if it is not from human
            source = source[1:]

        conv.messages = []
        for j, sentence in enumerate(source):
            role = roles[sentence["from"]]
            assert role == conv.roles[j % 2], f"{i}"
            conv.append_message(role, sentence["value"])
        conversations.append(conv.get_prompt())

    # Tokenize conversations

    if has_speech:
        input_ids = torch.stack([tokenizer_speech_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
    else:
        input_ids = tokenizer(
            conversations,
            return_tensors="pt",
            padding="longest",
            max_length=tokenizer.model_max_length,
            truncation=True,
        ).input_ids

    targets = input_ids.clone()

    assert conv.sep_style == conversation_lib.SeparatorStyle.LLAMA_2

    # Mask targets
    sep = "[/INST] "
    for conversation, target in zip(conversations, targets):
        total_len = int(target.ne(tokenizer.pad_token_id).sum())

        rounds = conversation.split(conv.sep2)
        cur_len = 1
        target[:cur_len] = IGNORE_INDEX
        for i, rou in enumerate(rounds):
            if rou == "":
                break

            parts = rou.split(sep)
            if len(parts) != 2:
                break
            parts[0] += sep

            if has_speech:
                round_len = len(tokenizer_speech_token(rou, tokenizer))
                instruction_len = len(tokenizer_speech_token(parts[0], tokenizer)) - 2
            else:
                round_len = len(tokenizer(rou).input_ids)
                instruction_len = len(tokenizer(parts[0]).input_ids) - 2

            target[cur_len : cur_len + instruction_len] = IGNORE_INDEX

            cur_len += round_len
        target[cur_len:] = IGNORE_INDEX

        if cur_len < tokenizer.model_max_length:
            if cur_len != total_len:
                target[:] = IGNORE_INDEX
                print(
                    f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
                    f" (ignored)"
                )

    return dict(
        input_ids=input_ids,
        labels=targets,
    )


def preprocess_llama_3(
    sources,
    tokenizer: transformers.PreTrainedTokenizer,
    has_speech: bool = False
) -> Dict:
    conv = conversation_lib.default_conversation.copy()
    roles = {"human": conv.roles[0], "gpt": conv.roles[1]}

    # Apply prompt templates
    conversations = []
    for i, source in enumerate(sources):
        if roles[source[0]["from"]] != conv.roles[0]:
            # Skip the first one if it is not from human
            source = source[1:]

        assert len(source) == 2, "now only support single-turn conversation"

        conv.messages = []
        for j, sentence in enumerate(source):
            role = roles[sentence["from"]]
            assert role == conv.roles[j % 2], f"{i}"
            conv.append_message(role, sentence["value"])
        conversations.append(conv.get_prompt())

    # Tokenize conversations

    if has_speech:
        input_ids = torch.stack([tokenizer_speech_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
    else:
        input_ids = tokenizer(
            conversations,
            return_tensors="pt",
            padding="longest",
            max_length=tokenizer.model_max_length,
            truncation=True,
        ).input_ids

    targets = input_ids.clone()

    assert conv.sep_style == conversation_lib.SeparatorStyle.LLAMA_3

    # Mask targets
    sep = "<|start_header_id|>" + conv.roles[1] + "<|end_header_id|>\n\n"
    for conversation, target in zip(conversations, targets):
        total_len = int(target.ne(tokenizer.pad_token_id).sum())

        cur_len = 1
        target[:cur_len] = IGNORE_INDEX
        parts = conversation.split(sep)
        parts[0] += sep

        if has_speech:
            conversation_len = len(tokenizer_speech_token(conversation, tokenizer))
            instruction_len = len(tokenizer_speech_token(parts[0], tokenizer)) - 1
        else:
            conversation_len = len(tokenizer(conversation).input_ids)
            instruction_len = len(tokenizer(parts[0]).input_ids) - 1

        target[cur_len : cur_len + instruction_len] = IGNORE_INDEX
        cur_len += conversation_len
        target[cur_len:] = IGNORE_INDEX

        # if cur_len < tokenizer.model_max_length:
        #     if cur_len != total_len:
        #         target[:] = IGNORE_INDEX
        #         print(
        #             f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
        #             f" (ignored)"
        #         )

    return dict(
        input_ids=input_ids,
        labels=targets,
    )


def preprocess_v1(
    sources,
    tokenizer: transformers.PreTrainedTokenizer,
    has_speech: bool = False
) -> Dict:
    conv = conversation_lib.default_conversation.copy()
    roles = {"human": conv.roles[0], "gpt": conv.roles[1]}

    # Apply prompt templates
    conversations = []
    for i, source in enumerate(sources):
        if roles[source[0]["from"]] != conv.roles[0]:
            # Skip the first one if it is not from human
            source = source[1:]

        conv.messages = []
        for j, sentence in enumerate(source):
            role = roles[sentence["from"]]
            assert role == conv.roles[j % 2], f"{i}"
            conv.append_message(role, sentence["value"])
        conversations.append(conv.get_prompt())

    # Tokenize conversations

    if has_speech:
        input_ids = torch.stack([tokenizer_speech_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
    else:
        input_ids = tokenizer(
            conversations,
            return_tensors="pt",
            padding="longest",
            max_length=tokenizer.model_max_length,
            truncation=True,
        ).input_ids

    targets = input_ids.clone()

    assert conv.sep_style == conversation_lib.SeparatorStyle.TWO

    # Mask targets
    sep = conv.sep + conv.roles[1] + ": "
    for conversation, target in zip(conversations, targets):
        total_len = int(target.ne(tokenizer.pad_token_id).sum())

        rounds = conversation.split(conv.sep2)
        cur_len = 1
        target[:cur_len] = IGNORE_INDEX
        for i, rou in enumerate(rounds):
            if rou == "":
                break

            parts = rou.split(sep)
            if len(parts) != 2:
                break
            parts[0] += sep

            if has_speech:
                round_len = len(tokenizer_speech_token(rou, tokenizer))
                instruction_len = len(tokenizer_speech_token(parts[0], tokenizer)) - 2
            else:
                round_len = len(tokenizer(rou).input_ids)
                instruction_len = len(tokenizer(parts[0]).input_ids) - 2

            # FIXME: tokenizer bug
            if i != 0 and not tokenizer.legacy and IS_TOKENIZER_GREATER_THAN_0_14:
                round_len -= 1
                instruction_len -= 1

            target[cur_len : cur_len + instruction_len] = IGNORE_INDEX

            cur_len += round_len
        target[cur_len:] = IGNORE_INDEX

        if cur_len < tokenizer.model_max_length:
            if cur_len != total_len:
                target[:] = IGNORE_INDEX
                print(
                    f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
                    f" (ignored)"
                )

    return dict(
        input_ids=input_ids,
        labels=targets,
    )


def preprocess_plain(
    sources: Sequence[str],
    tokenizer: transformers.PreTrainedTokenizer,
) -> Dict:
    # add end signal and concatenate together
    conversations = []
    for source in sources:
        assert len(source) == 2
        assert DEFAULT_SPEECH_TOKEN in source[0]['value']
        source[0]['value'] = DEFAULT_SPEECH_TOKEN
        conversation = source[0]['value'] + source[1]['value'] + conversation_lib.default_conversation.sep
        conversations.append(conversation)
    # tokenize conversations
    input_ids = [tokenizer_speech_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations]
    targets = copy.deepcopy(input_ids)
    for target, source in zip(targets, sources):
        tokenized_len = len(tokenizer_speech_token(source[0]['value'], tokenizer))
        target[:tokenized_len] = IGNORE_INDEX

    return dict(input_ids=input_ids, labels=targets)


def preprocess(
    sources: Sequence[str],
    tokenizer: transformers.PreTrainedTokenizer,
    has_speech: bool = False
) -> Dict:
    """
    Given a list of sources, each is a conversation list. This transform:
    1. Add signal '### ' at the beginning each sentence, with end signal '\n';
    2. Concatenate conversations together;
    3. Tokenize the concatenated conversation;
    4. Make a deepcopy as the target. Mask human words with IGNORE_INDEX.
    """
    if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.PLAIN:
        return preprocess_plain(sources, tokenizer)
    if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.LLAMA_2:
        return preprocess_llama_2(sources, tokenizer, has_speech=has_speech)
    if conversation_lib.default_conversation.version.startswith("v1"):
        return preprocess_v1(sources, tokenizer, has_speech=has_speech)
    if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.LLAMA_3:
        return preprocess_llama_3(sources, tokenizer, has_speech=has_speech)
    raise NotImplementedError