forked from open-compass/VLMEvalKit
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathinference.py
188 lines (153 loc) · 6.2 KB
/
inference.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
import torch
import torch.distributed as dist
from vlmeval.config import supported_VLM
from vlmeval.utils import track_progress_rich
from vlmeval.smp import *
FAIL_MSG = 'Failed to obtain answer via API.'
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--data', type=str, nargs='+', required=True)
parser.add_argument('--model', type=str, nargs='+', required=True)
parser.add_argument('--nproc', type=int, default=4, required=True)
parser.add_argument('--verbose', action='store_true')
args = parser.parse_args()
return args
# Only API model is accepted
def infer_data_api(model, work_dir, model_name, dataset, index_set=None, api_nproc=4, ignore_failed=False):
rank, world_size = get_rank_and_world_size()
assert rank == 0 and world_size == 1
dataset_name = dataset.dataset_name
data = dataset.data
if index_set is not None:
data = data[data['index'].isin(index_set)]
model = supported_VLM[model_name]() if isinstance(model, str) else model
assert getattr(model, 'is_api', False)
if hasattr(model, 'set_dump_image'):
model.set_dump_image(dataset.dump_image)
lt, indices = len(data), list(data['index'])
structs = []
for i in range(lt):
item = data.iloc[i]
if hasattr(model, 'use_custom_prompt') and model.use_custom_prompt(dataset_name):
assert hasattr(model, 'build_prompt')
struct = model.build_prompt(item, dataset=dataset_name)
else:
struct = dataset.build_prompt(item)
structs.append(struct)
# structs = [dataset.build_prompt(data.iloc[i]) for i in range(lt)]
out_file = f'{work_dir}/{model_name}_{dataset_name}_supp.pkl'
res = {}
if osp.exists(out_file):
res = load(out_file)
if ignore_failed:
res = {k: v for k, v in res.items() if FAIL_MSG not in v}
structs = [s for i, s in zip(indices, structs) if i not in res]
indices = [i for i in indices if i not in res]
gen_func = model.generate
structs = [dict(message=struct, dataset=dataset_name) for struct in structs]
if len(structs):
track_progress_rich(gen_func, structs, nproc=api_nproc, chunksize=api_nproc, save=out_file, keys=indices)
res = load(out_file)
if index_set is not None:
res = {k: v for k, v in res.items() if k in index_set}
os.remove(out_file)
return res
def infer_data(model, model_name, work_dir, dataset, out_file, verbose=False, api_nproc=4):
dataset_name = dataset.dataset_name
prev_file = f'{work_dir}/{model_name}_{dataset_name}_PREV.pkl'
res = load(prev_file) if osp.exists(prev_file) else {}
if osp.exists(out_file):
res.update(load(out_file))
rank, world_size = get_rank_and_world_size()
sheet_indices = list(range(rank, len(dataset), world_size))
lt = len(sheet_indices)
data = dataset.data.iloc[sheet_indices]
data_indices = [i for i in data['index']]
# If finished, will exit without building the model
all_finished = True
for i in range(lt):
idx = data.iloc[i]['index']
if idx not in res:
all_finished = False
if all_finished:
res = {k: res[k] for k in data_indices}
dump(res, out_file)
return
# Data need to be inferred
data = data[~data['index'].isin(res)]
lt = len(data)
model = supported_VLM[model_name]() if isinstance(model, str) else model
is_api = getattr(model, 'is_api', False)
if is_api:
lt, indices = len(data), list(data['index'])
supp = infer_data_api(
model=model,
work_dir=work_dir,
model_name=model_name,
dataset=dataset,
index_set=set(indices),
api_nproc=api_nproc)
for idx in indices:
assert idx in supp
res.update(supp)
res = {k: res[k] for k in data_indices}
dump(res, out_file)
return model
else:
model.set_dump_image(dataset.dump_image)
for i in tqdm(range(lt)):
idx = data.iloc[i]['index']
if idx in res:
continue
if hasattr(model, 'use_custom_prompt') and model.use_custom_prompt(dataset_name):
struct = model.build_prompt(data.iloc[i], dataset=dataset_name)
else:
struct = dataset.build_prompt(data.iloc[i])
response = model.generate(message=struct, dataset=dataset_name)
torch.cuda.empty_cache()
if verbose:
print(response, flush=True)
res[idx] = response
if (i + 1) % 10 == 0:
dump(res, out_file)
res = {k: res[k] for k in data_indices}
dump(res, out_file)
return model
# A wrapper for infer_data, do the pre & post processing
def infer_data_job(model, work_dir, model_name, dataset, verbose=False, api_nproc=4, ignore_failed=False):
rank, world_size = get_rank_and_world_size()
dataset_name = dataset.dataset_name
result_file = osp.join(work_dir, f'{model_name}_{dataset_name}.xlsx')
prev_file = f'{work_dir}/{model_name}_{dataset_name}_PREV.pkl'
if osp.exists(result_file):
if rank == 0:
data = load(result_file)
results = {k: v for k, v in zip(data['index'], data['prediction'])}
if not ignore_failed:
results = {k: v for k, v in results.items() if FAIL_MSG not in str(v)}
dump(results, prev_file)
if world_size > 1:
dist.barrier()
tmpl = osp.join(work_dir, '{}' + f'{world_size}_{dataset_name}.pkl')
out_file = tmpl.format(rank)
model = infer_data(
model=model, work_dir=work_dir, model_name=model_name, dataset=dataset,
out_file=out_file, verbose=verbose, api_nproc=api_nproc)
if world_size > 1:
dist.barrier()
if rank == 0:
data_all = {}
for i in range(world_size):
data_all.update(load(tmpl.format(i)))
data = dataset.data
for x in data['index']:
assert x in data_all
data['prediction'] = [str(data_all[x]) for x in data['index']]
if 'image' in data:
data.pop('image')
dump(data, result_file)
for i in range(world_size):
os.remove(tmpl.format(i))
if world_size > 1:
dist.barrier()
return model