-
Notifications
You must be signed in to change notification settings - Fork 0
/
evaluate.py
160 lines (135 loc) · 6.2 KB
/
evaluate.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
#!/usr/bin/env python3 -u
# Copyright 2022 The OFA-Sys Team.
# All rights reserved.
# This source code is licensed under the Apache 2.0 license
# found in the LICENSE file in the root directory.
import logging
import os
import sys
import numpy as np
import torch
from fairseq import distributed_utils, options, tasks, utils
from fairseq.dataclass.utils import convert_namespace_to_omegaconf
from fairseq.logging import progress_bar
from fairseq.utils import reset_logging
from omegaconf import DictConfig
from utils import checkpoint_utils
from utils.eval_utils import eval_step, merge_results
from utils.zero_shot_utils import zero_shot_step
logging.basicConfig(
format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
datefmt="%Y-%m-%d %H:%M:%S",
level=os.environ.get("LOGLEVEL", "INFO").upper(),
stream=sys.stdout,
)
logger = logging.getLogger("ofa.evaluate")
def apply_half(t):
if t.dtype is torch.float32:
return t.to(dtype=torch.half)
return t
def main(cfg: DictConfig, **kwargs):
utils.import_user_module(cfg.common)
reset_logging()
logger.info(cfg)
assert (
cfg.dataset.max_tokens is not None or cfg.dataset.batch_size is not None
), "Must specify batch size either with --max-tokens or --batch-size"
# Fix seed for stochastic decoding
if cfg.common.seed is not None and not cfg.generation.no_seed_provided:
np.random.seed(cfg.common.seed)
utils.set_torch_seed(cfg.common.seed)
use_fp16 = cfg.common.fp16
use_cuda = torch.cuda.is_available() and not cfg.common.cpu
if use_cuda:
torch.cuda.set_device(cfg.distributed_training.device_id)
# Load ensemble
overrides = eval(cfg.common_eval.model_overrides)
# Deal with beam-search / all-candidate VQA eval
if cfg.task._name == "vqa_gen":
overrides['val_inference_type'] = "beamsearch" if kwargs['beam_search_vqa_eval'] else "allcand"
logger.info("loading model(s) from {}".format(cfg.common_eval.path))
if kwargs["zero_shot"]:
task = tasks.setup_task(cfg.task)
models, saved_cfg = checkpoint_utils.load_model_ensemble(
utils.split_paths(cfg.common_eval.path),
arg_overrides=overrides,
task=task,
suffix=cfg.checkpoint.checkpoint_suffix,
strict=(cfg.checkpoint.checkpoint_shard_count == 1),
num_shards=cfg.checkpoint.checkpoint_shard_count,
)
else:
models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
utils.split_paths(cfg.common_eval.path),
arg_overrides=overrides,
suffix=cfg.checkpoint.checkpoint_suffix,
strict=(cfg.checkpoint.checkpoint_shard_count == 1),
num_shards=cfg.checkpoint.checkpoint_shard_count,
)
# loading the dataset should happen after the checkpoint has been loaded so we can give it the saved task config
task.load_dataset(cfg.dataset.gen_subset, task_cfg=saved_cfg.task)
# Move models to GPU
for model, ckpt_path in zip(models, utils.split_paths(cfg.common_eval.path)):
if kwargs['ema_eval']:
logger.info("loading EMA weights from {}".format(ckpt_path))
model.load_state_dict(checkpoint_utils.load_ema_from_checkpoint(ckpt_path)['model'])
model.eval()
if use_fp16:
model.half()
if use_cuda and not cfg.distributed_training.pipeline_model_parallel:
model.cuda()
model.prepare_for_inference_(cfg)
# Load dataset (possibly sharded)
itr = task.get_batch_iterator(
dataset=task.dataset(cfg.dataset.gen_subset),
max_tokens=cfg.dataset.max_tokens,
max_sentences=cfg.dataset.batch_size,
max_positions=utils.resolve_max_positions(
task.max_positions(), *[m.max_positions() for m in models]
),
ignore_invalid_inputs=cfg.dataset.skip_invalid_size_inputs_valid_test,
required_batch_size_multiple=cfg.dataset.required_batch_size_multiple,
seed=cfg.common.seed,
num_shards=cfg.distributed_training.distributed_world_size,
shard_id=cfg.distributed_training.distributed_rank,
num_workers=cfg.dataset.num_workers,
data_buffer_size=cfg.dataset.data_buffer_size,
).next_epoch_itr(shuffle=False)
progress = progress_bar.progress_bar(
itr,
log_format=cfg.common.log_format,
log_interval=cfg.common.log_interval,
default_log_format=("tqdm" if not cfg.common.no_progress_bar else "simple"),
)
# Initialize generator
generator = task.build_generator(models, cfg.generation)
results = []
score_sum = torch.FloatTensor([0]).cuda()
score_cnt = torch.FloatTensor([0]).cuda()
for sample in progress:
if "net_input" not in sample:
continue
sample = utils.move_to_cuda(sample) if use_cuda else sample
sample = utils.apply_to_sample(apply_half, sample) if cfg.common.fp16 else sample
with torch.no_grad():
if kwargs["zero_shot"]:
result, scores = zero_shot_step(task, generator, models, sample)
else:
result, scores = eval_step(task, generator, models, sample, **kwargs)
results += result
score_sum += sum(scores) if scores is not None else 0
score_cnt += len(scores) if scores is not None else 0
progress.log({"sentences": sample["nsentences"]})
merge_results(task, cfg, logger, score_cnt, score_sum, results)
def cli_main():
parser = options.get_generation_parser()
parser.add_argument("--ema-eval", action='store_true', help="Use EMA weights to make evaluation.")
parser.add_argument("--beam-search-vqa-eval", action='store_true', help="Use beam search for vqa evaluation (faster inference speed but sub-optimal result), if not specified, we compute scores for each answer in the candidate set, which is slower but can obtain best result.")
parser.add_argument("--zero-shot", action='store_true')
args = options.parse_args_and_arch(parser)
cfg = convert_namespace_to_omegaconf(args)
distributed_utils.call_main(
cfg, main, ema_eval=args.ema_eval, beam_search_vqa_eval=args.beam_search_vqa_eval, zero_shot=args.zero_shot
)
if __name__ == "__main__":
cli_main()