-
Notifications
You must be signed in to change notification settings - Fork 48
/
dataset.py
278 lines (224 loc) · 10.8 KB
/
dataset.py
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
import json
import random
import re
import pandas as pd
import numpy as np
from PIL import Image
from torch.utils.data import Dataset, DataLoader
import torch
from model.vision_utils import get_img_process
import os
os.environ["TOKENIZERS_PARALLELISM"] = "false"
class PretrainDataset(Dataset):
def __init__(self, json_path, tokenizer, vision_model=None, max_length=1024,
prompt_max_len=512,
answer_max_len=256,
image_special_token='<' * 25 + '>' * 25):
super().__init__()
with open(json_path, 'r', encoding='utf-8') as f:
self.data = json.load(f)
self.max_length = max_length
self.prompt_max_len = prompt_max_len
self.answer_max_len = answer_max_len
self.tokenizer = tokenizer
self.vision_model, self.preprocess = vision_model
self.padding = 0
self.bos_id = self.tokenizer('<s>assistant').data['input_ids']
self.dataset_path = './dataset/pretrain_images/'
self.image_special_token = image_special_token
def __len__(self):
return len(self.data)
def find_sublist_index(self, main_list, sub_list) -> int:
last_index = -1
for i in range(len(main_list) - len(sub_list) + 1):
if main_list[i:i + len(sub_list)] == sub_list:
last_index = i
return last_index
def safe_eval(self, s):
try:
res = eval(s)
except Exception as e:
return []
return res
def __getitem__(self, index: int):
sample = self.data[index]
image_name = sample['image']
conversation = sample['conversations']
# minimind-v的image的特殊占位符，对应每张图切分成M个token，和get_img_process中的数量对应
messages = []
# 遍历 conversation 列表
for i in range(0, len(conversation), 2):
# 检查是否有配对的问题和回答
if i + 1 < len(conversation):
q = conversation[i]['value'].replace('<image>', self.image_special_token)
a = conversation[i + 1]['value']
if q and a:
messages.append({"role": "user", "content": q})
messages.append({"role": "assistant", "content": a})
new_prompt = self.tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
input_id = self.tokenizer(new_prompt).data['input_ids'][:self.max_length]
question_length = self.find_sublist_index(input_id, self.bos_id) + len(self.bos_id)
padding_len = self.max_length - len(input_id)
input_id = input_id + [self.padding] * padding_len
mask_len = len(input_id) - question_length - padding_len
loss_mask = [0] * question_length + [1] * (mask_len) + [0] * padding_len
input_id = np.array(input_id)
X = np.array(input_id[:-1]).astype(np.int64)
Y = np.array(input_id[1:]).astype(np.int64)
loss_mask = np.array(loss_mask[1:]).astype(np.int64)
X_tensor = torch.from_numpy(X)
Y_tensor = torch.from_numpy(Y)
loss_mask_tensor = torch.from_numpy(loss_mask)
image = Image.open(f'{self.dataset_path}{image_name}')
image_encoders = get_img_process(image, self.preprocess)
return X_tensor, Y_tensor, loss_mask_tensor, image_encoders
class SFTDataset(Dataset):
def __init__(self, json_path, tokenizer, vision_model=None, max_length=1024,
prompt_max_len=512,
answer_max_len=256,
image_special_token='<' * 25 + '>' * 25):
super().__init__()
with open(json_path, 'r', encoding='utf-8') as f:
self.data = json.load(f)
self.max_length = max_length
self.prompt_max_len = prompt_max_len
self.answer_max_len = answer_max_len
self.tokenizer = tokenizer
self.vision_model, self.preprocess = vision_model
self.padding = 0
self.bos_id = self.tokenizer('<s>assistant').data['input_ids']
self.dataset_path = './dataset/sft_images/'
self.image_special_token = image_special_token
def __len__(self):
return len(self.data)
def find_sublist_index(self, main_list, sub_list) -> int:
last_index = -1
for i in range(len(main_list) - len(sub_list) + 1):
if main_list[i:i + len(sub_list)] == sub_list:
last_index = i
return last_index
def safe_eval(self, s):
try:
res = eval(s)
except Exception as e:
return []
return res
def __getitem__(self, index: int):
sample = self.data[index]
image_name = 'COCO_train2014_' + sample['image']
conversation = sample['conversations']
# minimind-v的image的特殊占位符，对应每张图切分成M个token，和get_img_process中的数量对应
messages = []
# 遍历 conversation 列表
# for i in range(0, len(conversation), 2):
for i in range(0, 1):
# 检查是否有配对的问题和回答
if i + 1 < len(conversation):
q = conversation[i]['value'].replace('<image>', self.image_special_token)
a = conversation[i + 1]['value']
if q and a:
messages.append({"role": "user", "content": q})
messages.append({"role": "assistant", "content": a})
new_prompt = self.tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
input_id = self.tokenizer(new_prompt).data['input_ids'][:self.max_length]
question_length = self.find_sublist_index(input_id, self.bos_id) + len(self.bos_id)
padding_len = self.max_length - len(input_id)
input_id = input_id + [self.padding] * padding_len
mask_len = len(input_id) - question_length - padding_len
loss_mask = [0] * question_length + [1] * (mask_len) + [0] * padding_len
input_id = np.array(input_id)
X = np.array(input_id[:-1]).astype(np.int64)
Y = np.array(input_id[1:]).astype(np.int64)
loss_mask = np.array(loss_mask[1:]).astype(np.int64)
X_tensor = torch.from_numpy(X)
Y_tensor = torch.from_numpy(Y)
loss_mask_tensor = torch.from_numpy(loss_mask)
image = Image.open(f'{self.dataset_path}{image_name}')
image_encoders = get_img_process(image, self.preprocess)
return X_tensor, Y_tensor, loss_mask_tensor, image_encoders
class SFTDataset_multi(Dataset):
def __init__(self, json_path, tokenizer, vision_model=None, max_length=1024,
prompt_max_len=512,
answer_max_len=256,
image_special_token='<' * 25 + '>' * 25):
super().__init__()
with open(json_path, 'r', encoding='utf-8') as f:
self.data = json.load(f)
self.max_length = max_length
self.prompt_max_len = prompt_max_len
self.answer_max_len = answer_max_len
self.tokenizer = tokenizer
self.vision_model, self.preprocess = vision_model
self.padding = 0
self.bos_id = self.tokenizer('<s>assistant').data['input_ids']
self.dataset_path = './dataset/sft_multi_images/'
self.image_special_token = image_special_token
def __len__(self):
return len(self.data)
def find_sublist_index(self, main_list, sub_list) -> int:
last_index = -1
for i in range(len(main_list) - len(sub_list) + 1):
if main_list[i:i + len(sub_list)] == sub_list:
last_index = i
return last_index
def safe_eval(self, s):
try:
res = eval(s)
except Exception as e:
return []
return res
def __getitem__(self, index: int):
sample = self.data[index]
image_names = sample['image'].split(', ')
conversation = sample['conversations']
messages = []
for i in range(0, 1):
if i + 1 < len(conversation):
q = conversation[i]['value'].replace('<image>', self.image_special_token)
a = conversation[i + 1]['value']
if q and a:
messages.append({"role": "user", "content": q})
messages.append({"role": "assistant", "content": a})
# print(messages) # [{'role': 'user', 'content': '<<<<<<<<<<<<<<<<<<<<<<<<<>>>>>>>>>>>>>>>>>>>>>>>>><<<<<<<<<<<<<<<<<<<<<<<<<>>>>>>>>>>>>>>>>>>>>>>>>>\nAre the two images below that resemble each other described by the same term? You must choose your answer from the Choice List. Choice_List: True, False.'}, {'role': 'assistant', 'content': 'False'}][{'role': 'user', 'content': '<<<<<<<<<<<<<<<<<<<<<<<<<>>>>>>>>>>>>>>>>>>>>>>>>><<<<<<<<<<<<<<<<<<<<<<<<<>>>>>>>>>>>>>>>>>>>>>>>>><<<<<<<<<<<<<<<<<<<<<<<<<>>>>>>>>>>>>>>>>>>>>>>>>><<<<<<<<<<<<<<<<<<<<<<<<<>>>>>>>>>>>>>>>>>>>>>>>>>\nDo these four pictures fall into the same category? You must choose your answer from the Choice List. Choice_List: True, False.'}, {'role': 'assistant', 'content': 'False'}]
new_prompt = self.tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
input_id = self.tokenizer(new_prompt).data['input_ids'][:self.max_length]
# print(len(input_id)) # 165 or 263 or 259
question_length = self.find_sublist_index(input_id, self.bos_id) + len(self.bos_id)
padding_len = self.max_length - len(input_id)
input_id = input_id + [self.padding] * padding_len
mask_len = len(input_id) - question_length - padding_len
loss_mask = [0] * question_length + [1] * (mask_len) + [0] * padding_len
input_id = np.array(input_id)
X = np.array(input_id[:-1]).astype(np.int64)
Y = np.array(input_id[1:]).astype(np.int64)
loss_mask = np.array(loss_mask[1:]).astype(np.int64)
X_tensor = torch.from_numpy(X)
Y_tensor = torch.from_numpy(Y)
loss_mask_tensor = torch.from_numpy(loss_mask)
# 读取多张图像
image_encoders = []
for image_name in image_names:
image = Image.open(f'{self.dataset_path}{image_name.strip()}') # 去掉可能的空格
image_encoders.append(get_img_process(image, self.preprocess)['pixel_values'])
# 确定目标形状
max_images = 2 # 根据你的需求设置 一次性最大输入几张图片
target_shape = (max_images, 3, 224, 224)
# 创建填充张量
padded_image_encoders = torch.zeros(target_shape, dtype=torch.float32)
# 填充图像编码
for i, img_enc in enumerate(image_encoders):
if i < max_images:
padded_image_encoders[i] = img_enc.squeeze(0)
return X_tensor, Y_tensor, loss_mask_tensor, padded_image_encoders