-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathevaluation_aqa_dataset.py
390 lines (355 loc) · 16.7 KB
/
evaluation_aqa_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
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
"""
这个文件用于进行带mask和reference的模型的evaluation.
v2: 加入了confident的支持。
"""
import argparse
import os
import random
import numpy as np
import torch
import time
import json
import jsonlines
import torch.backends.cudnn as cudnn
import csv
from tqdm import tqdm
from PIL import Image
import cv2
from torch.utils.data import DataLoader
from minigpt4.common.config import Config
from minigpt4.common.dist_utils import get_rank
from minigpt4.common.registry import registry
from minigpt4.conversation.conversation import StoppingCriteriaSub
from minigpt4.datasets.datasets.anomaly_detection import AnomalyDetectionDataset
from transformers import StoppingCriteria, StoppingCriteriaList
from datetime import datetime
# imports modules for registration
from minigpt4.processors.transform import Expand2square
from minigpt4.datasets.builders import *
from minigpt4.models import *
from minigpt4.processors import *
from minigpt4.runners import *
import minigpt4.tasks as tasks
from minigpt4.common.utils import disable_torch_init
from minigpt4.datasets.data_utils import prepare_sample
def parse_args():
parser = argparse.ArgumentParser(description="Demo")
parser.add_argument("--cfg-path", required=True, help="path to configuration file.")
parser.add_argument("--gpu-id", type=int, default=0, help="specify the gpu to load the model.")
parser.add_argument("--task_type", type=str, default='aqa', choices=['aqa', 'roi', 'al', '1cls', 'adroi', 'shot'])
parser.add_argument("--split", type=str, default='eval', choices=['eval', 'test', 'train','eval_un','eval_fewshot','visa','mvtec'])
parser.add_argument("--ckpt", type=int, default=-1)
parser.add_argument("--bs", type=int, default=1)
parser.add_argument("--round_index", type=int, default=14)
parser.add_argument("--k_shot", type=int, default=0)
parser.add_argument("--start", type=int, default=0)
parser.add_argument(
"--options",
nargs="+",
help="override some settings in the used config, the key-value pair "
"in xxx=yyy format will be merged into config file (deprecate), "
"change to --cfg-options instead.",
)
args = parser.parse_args()
return args
def setup_seeds(config):
seed = config.run_cfg.seed + get_rank()
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
cudnn.benchmark = False
cudnn.deterministic = True
def get_context_emb(img_embeds, questions, model):
batch_size = img_embeds.shape[0]
texts = questions.split('<ImageHere>')
assert img_embeds.shape[1] == len(texts) - 1
text_embeds = []
for ind, text in enumerate(texts):
text_token = model.llama_tokenizer(
text, return_tensors="pt", add_special_tokens=(ind==0)).to(img_embeds.device)
text_embed = model.llama_model.model.embed_tokens(text_token.input_ids).expand(batch_size, -1, -1)
text_embeds.append(text_embed)
# print('text:', text_embed.shape)
if ind < img_embeds.shape[1]:
# print('roi:', img_embeds[:, ind].shape)
text_embeds.append(img_embeds[:, ind])
q_embed = torch.cat(text_embeds, dim=1)
return q_embed
def anomaly_map_handler(output):
if 've_anomaly_maps' not in output:
return []
anomaly_maps = output['ve_anomaly_maps']
bs = anomaly_maps.shape[0]
anomaly_map_list = []
for ind in range(bs):
_tmp = anomaly_maps[ind].detach().cpu().expand(3, -1, -1)
_tmp = _tmp.permute(1, 2, 0).numpy() * 255.
anomaly_map_list.append(_tmp.astype(np.uint8))
return anomaly_map_list
ROOTS = {
'train': "./data/TrainADDataset",
'eval': "./data/EvalADDataset",
'test': "./data/EvalADDataset",
'visa': "./data/EvalADDataset",
'mvtec': "./data/EvalADDataset",
}
ANNO_FILES = {
'aqa': {
'train': 'AQA_train.jsonl',
'test': 'AQA_test.jsonl',
'eval': 'AQA_eval.jsonl'
},
'roi': {
'train': 'AQA_train.jsonl',
'test': 'AQA_test.jsonl',
'eval': 'AQA_eval.jsonl'
},
'al': {
'eval': 'val_coco.json'
},
'ad': {
'eval': 'DC_MVTEC_test_normal.jsonl',
# 'eval': 'DC_VISA_test_normal.jsonl',
},
'ad_few': {
'eval': 'DC_VISA_test_normal.jsonl',
# 'eval': 'DC_VISA_test_normal.jsonl',
},
'adroi': {
'eval': 'DC_MVTEC_test_normal.jsonl'
},
'1cls': {
'visa': 'DC_VISA_test_normal.jsonl',
'mvtec': 'DC_MVTEC_test_normal.jsonl'
},
'shot': {
'visa': 'DC_VISA_test_normal.jsonl',
'mvtec': 'DC_MVTEC_test_normal.jsonl'
},
}
def build_dataset(args, ds_cfg):
if args.task_type in ['aqa']:
return AQADataset(
LocImageTrainProcessor(identity=True),
None,
ROOTS[args.split],
ann_paths=[os.path.join(ROOTS[args.split], ANNO_FILES[args.task_type][args.split])],
add_noa=True,
with_ve=ds_cfg[args.task_type].get('with_ve', False),
stage='test'
)
elif args.task_type in ['roi']:
return ROIDataset(
LocImageTrainProcessor(identity=True),
None,
ROOTS[args.split],
ann_paths=[os.path.join(ROOTS[args.split], ANNO_FILES[args.task_type][args.split])],
with_ve=ds_cfg[args.task_type].get('with_ve', False),
stage='test'
)
elif args.task_type in ['al']:
return AlignDataset(
LocImageTrainProcessor(identity=True),
None,
ROOTS[args.split],
ve_root="/mnt/vdb1/datasets/aprilgan_processresults",
ann_paths=[os.path.join(ROOTS[args.split], ANNO_FILES[args.task_type][args.split])],
with_mask=ds_cfg['align'].get('with_mask', False),
with_gt_seg=ds_cfg['align'].get('with_gt_seg', False),
)
elif args.task_type in ['ad']:
return AnomalyDetectionDataset(
LocImageTrainProcessor(identity=True),
None,
ROOTS[args.split],
ve_root="/mnt/vdb1/datasets/aprilgan_processresults",
ann_paths=[os.path.join(ROOTS[args.split], ANNO_FILES[args.task_type][args.split])],
with_mask=ds_cfg['anomaly_detection'].get('with_mask', False),
is_preload=True,
stage='test'
)
elif args.task_type in ['ad_few']:
return AnomalyDetectionDataset(
LocImageTrainProcessor(identity=True),
None,
ROOTS[args.split],
ve_root="/mnt/vdb1/datasets/aprilgan_processresults",
ann_paths=[os.path.join(ROOTS[args.split], ANNO_FILES[args.task_type][args.split])],
with_mask=ds_cfg['anomaly_detection'].get('with_mask', False),
is_preload=True,
stage='test'
)
elif args.task_type in ['1cls','shot']:
return AnomalyDetectionDataset(
LocImageTrainProcessor(identity=True),
None,
ROOTS[args.split],
ve_root="/mnt/vdb1/datasets/aprilgan_processresults",
ann_paths=[ANNO_FILES[args.task_type][args.split]],
img_size=ds_cfg['anomaly_detection'].get("img_size", 224),
crop_size=ds_cfg['anomaly_detection'].get("crop_size", 224),
with_mask=ds_cfg['anomaly_detection'].get('with_mask', False),
is_preload=True,
stage='test'
)
elif args.task_type in ['adroi']:
print("Num RoIs:", ds_cfg['adroi'].get("num_rois", 3))
return ADRoIDataset(
LocImageTrainProcessor(identity=True),
None,
ROOTS[args.split],
ve_root="/mnt/vdb1/datasets/aprilgan_processresults",
ann_paths=[os.path.join(ROOTS[args.split], ANNO_FILES[args.task_type][args.split])],
is_preload=True,
num_rois=ds_cfg['adroi'].get("num_rois", 3),
stage='test'
)
else:
raise NotImplementedError(f"Not implement for task type {args.task_type}")
if __name__ == '__main__':
# ========================================
# Model Initialization
# ========================================
print('Initializing Chat')
args = parse_args()
cfg = Config(args)
disable_torch_init()
model_config = cfg.model_cfg
round_index = args.round_index
model_config.round_index = round_index
k_shot = args.k_shot
model_config.k_shot = k_shot
if args.ckpt != -1:
old_ckpt = model_config.ckpt
ckpt_path = old_ckpt.split('/')
ckpt_path[-1] = f'checkpoint_{args.ckpt}.pth'
model_config.ckpt = '/'.join(ckpt_path)
#
model_config.device_8bit = args.gpu_id
model_cls = registry.get_model_class(model_config.arch)
model = model_cls.from_config(model_config).to('cuda:{}'.format(args.gpu_id))
# print('aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',model_config)
# print(model_config{'round_index'})
# print('round_indexround_indexround_indexround_indexround_indexround_indexround_indexround_indexround_indexround_indexround_indexround_indexround_index',round_index)
# print(model_config['round_index'])
# round_index = model_config.round_index
# ========================================
# Data Preparation
# ========================================
system="Give the following image: <Img>ImageContent</Img>. You will be able to see the image once I provide it to you. Please answer my questions."
stop_words_ids = [torch.tensor([835]).cuda(),
torch.tensor([2277, 29937]).cuda()] # '###' can be encoded in two different ways.
stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops=stop_words_ids)])
dataloader = DataLoader(
build_dataset(args, cfg.datasets_cfg),
batch_size=1 if args.task_type == 'al' else args.bs,
num_workers=4,
pin_memory=True,
)
# 保存路径
ckpt_name, _ = os.path.splitext(os.path.basename(model_config.ckpt))
num_ckpt = int(ckpt_name.split('_')[-1])
prefix = f"results_ckpt{num_ckpt}_training={args.task_type}_split={args.split}_kshot={k_shot}_roundindex={round_index}_{datetime.now().strftime('%Y%m%d_%H%M')}"
save_path = os.path.join('/', *model_config.ckpt.split('/')[1:-1], f"{prefix}.jsonl")
# if model_config.get("use_ve", False):
# anomaly_map_dir = os.path.join(os.path.join('/', *model_config.ckpt.split('/')[1:-1],'result', prefix))
# os.makedirs(anomaly_map_dir)
# print(f"Anomaly Maps will be saved to {anomaly_map_dir}")
print(f"Results will be saved to {save_path}")
generate_kwargs = {
'max_new_tokens': 90,
'stopping_criteria': stopping_criteria,
# 'num_beams': 1,
'do_sample': True,
'use_cache':True,
'min_length': 1,
# 'top_p': 0.9,
'top_p': 0.01,
# 'repetition_penalty': 1.0,
# 'length_penalty': 1,
'temperature': 1.0,
}
model.eval()
results = []
use_global_feat = model_config.get("use_global_feat", False)
append_global_feat = model_config.get("append_global_feat", False)
get_mem=lambda x:max(x,sum([torch.cuda.memory_allocated(i) for i in range(1)]))
max_mem = 0.0
all_time = 0.0
with jsonlines.open(save_path, 'w') as writer:
for testid, data_sample in tqdm(enumerate(iter(dataloader)), total=len(dataloader), desc='Evaluation AQA Task:'):
if testid >=args.start:
with torch.no_grad():
samples = prepare_sample(data_sample)
batch_size = samples['image'].shape[0]
# 这里应该不加上###Human: ... ###Assistant:，这个应该是model的工作
# if args.task_type == 'aqa':
# samples['question'] = ["Which part of images has defects? (A) <Img><ImageHere></Img> (B) <Img><ImageHere></Img> (C) <Img><ImageHere></Img> (D) all parts are normal. "]
# elif args.task_type in ['roi', 'ad', 'adroi','ad_few']:
# pass
# elif args.task_type == 'al':
# samples['question'] = [
# "###Human: <Img><ImageHere></Img> Describe this image in detail and find out anomaly defects if they exist. ###Assistant: "
# ] * batch_size
# else:
# raise NotImplementedError(f"Not implement for task type {args.task_type}")
max_mem = get_mem(max_mem)
t1 = time.time()
outputs = model.generate(samples, **generate_kwargs)
t2 = time.time()
max_mem = get_mem(max_mem)
all_time += (t2 - t1)
if isinstance(outputs, dict):
token_ids = outputs['token_ids']
anomaly_maps = anomaly_map_handler(outputs)
else:
token_ids = outputs
token_ids = torch.clamp(token_ids,1,40000)
output_text = model.llama_tokenizer.batch_decode(token_ids, add_special_tokens=False)
for ind, text in enumerate(output_text):
if args.task_type in ['aqa']:
item = {
'image_id': data_sample['image_id'][ind].item(),
'output': text.split("###")[0],
'question': samples['question'] if len(samples['question']) == 1 else samples['question'][ind],
'options': samples['options'][ind].detach().cpu().numpy().tolist(),
'answer': data_sample['answer'][ind].item(), # gt.
'is_anomaly': data_sample['is_anomaly'][ind].item(), # gt
}
elif args.task_type in ['roi']:
item = {
'image_id': data_sample['image_id'][ind].item(),
'output': text.split("###")[0],
'question': samples['question'] if len(samples['question']) == 1 else samples['question'][ind],
'options': samples['options'][ind].detach().cpu().numpy().tolist(),
# 'answer': data_sample['answer'][ind].item(), # gt.
'is_anomaly': data_sample['is_anomaly'][ind].item(), # gt
}
elif args.task_type in ['al', 'ad', 'adroi', 'ad_few','1cls','shot']:
imagepath = samples['img_path'][ind]
item = {
'image_id': data_sample['image_id'][ind].item(),
'image_path': '/'.join(imagepath.split('/')[-5:]),
# 'origin_output': model.llama_tokenizer.decode(token_ids[ind], add_special_tokens=False), # 说明batch_decode和decode没有本质区别。
# 'question': samples['question'] if len(samples['question']) == 1 else samples['question'][ind],
'is_anomaly': data_sample['is_anomaly'][ind].item(), # gt
}
if anomaly_maps is not None:
# anomaly_map_path = os.path.join(anomaly_map_dir, f"{data_sample['image_id'][ind].item():04d}.png")
# os.makedirs(anomaly_map_dir+'/'.join(samples['img_path'][ind].split('/')[-5:-1]),exist_ok=True)
# print(anomaly_map_dir+samples['img_path'][ind],anomaly_map_dir+'/'.join(samples['img_path'][ind].split('/')[-5:-1]))
# cv2.imwrite(anomaly_map_dir+'/'.join(samples['img_path'][ind].split('/')[-5:]), anomaly_maps[ind])
if 'Yes' in text.split("###")[0] and data_sample['is_anomaly'][ind]==True:
item['error'] = '0'
elif 'No' in text.split("###")[0] and data_sample['is_anomaly'][ind].item()==False:
item['error'] = '0'
else:
item['error'] = '1'
item['output'] = text.split("###")[0]
item['anomaly_score'] = str(round((anomaly_maps[ind].max())/255.0,4))
else:
raise NotImplementedError(f"Not implement for task type {args.task_type}")
writer.write(item)
print("CUDA Memory:", torch.cuda.max_memory_allocated() / (1024 * 1024))
print("Max Memory: ", max_mem / (1024 * 1024))
print("Mean Time: ", all_time / len(dataloader))