Repository navigation
Expand file tree
/
Copy pathrun.py
More file actions
275 lines (211 loc) · 9.74 KB
/
Copy pathrun.py
File metadata and controls
275 lines (211 loc) · 9.74 KB
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
import os
import ast
import json
import random
import shutil
import argparse
from collections import defaultdict
from pathlib import Path
from omegaconf import OmegaConf, DictConfig, ListConfig
import torch
import lightning.pytorch as pl
import torch.distributed as dist
from src.utils.utils import instantiate_from_config, get_timestamp, get_metric_statistics
from src.utils.log import setup_logger
logger = setup_logger(__name__)
start_time = get_timestamp()
def do_test(model: pl.LightningModule, trainer: pl.Trainer, ckpt_path: str, data_module: pl.LightningDataModule, args):
results = defaultdict(list)
if ckpt_path == "best":
state_dict = torch.load(trainer.checkpoint_callback.best_model_path, weights_only=False)["state_dict"]
elif ckpt_path == "last":
state_dict = torch.load(trainer.checkpoint_callback.last_model_path, weights_only=False)["state_dict"]
else:
state_dict = torch.load(ckpt_path, weights_only=False)["state_dict"]
logger.info(f"Loading ckpt from {ckpt_path}")
logger.info(model.load_state_dict(state_dict=state_dict, strict=False))
for i in range(args.test_times):
pl.seed_everything(args.seed + i)
res = trainer.test(model=model, datamodule=data_module)[0]
for k, v in res.items():
results[k].append(v)
statistics = {k: get_metric_statistics(v, args.test_times) for k, v in results.items()}
test_result_in_text = f"Test results: {results}\nTest statistics with {args.test_times} replications: {statistics}"
model.text_logger.log(test_result_in_text)
model.sample_logs["test_result"] = test_result_in_text
model.json_logger.log(model.sample_logs)
return results, statistics
def instantiate_callbacks(callback_configs: ListConfig):
callbacks = []
for callback_cfg in callback_configs:
callbacks.append(instantiate_from_config(callback_cfg))
return callbacks
def _preprocess_config(config, args, unknown_args):
def set_config_key_value(inplace_dict, key_path, value):
flag = False
def bfs_set_config_key_value(inplace_dict, key, value):
nonlocal flag
if key in inplace_dict.keys():
inplace_dict[key] = value
flag = True
for v in inplace_dict.values():
if isinstance(v, (DictConfig, dict)):
bfs_set_config_key_value(inplace_dict=v, key=key, value=value)
elif isinstance(v, ListConfig):
for item in v:
if isinstance(item, (DictConfig, dict)):
bfs_set_config_key_value(inplace_dict=item, key=key, value=value)
keys = key_path.split(".") # dataset.a.b = 1
len_keys = len(keys)
if len_keys == 1:
bfs_set_config_key_value(inplace_dict, key=key_path, value=value)
if flag:
return
else:
raise ValueError(f"{key_path} is not found in config")
for key_idx in range(len_keys - 1): #
cur_key = keys[key_idx]
if isinstance(inplace_dict, (DictConfig, dict)) and cur_key not in inplace_dict:
raise ValueError(
f"Config key path '{key_path}' is invalid: missing intermediate key '{cur_key}'. "
f"Available keys at this level: {list(inplace_dict.keys())}"
)
inplace_dict = inplace_dict[cur_key]
if isinstance(inplace_dict, ListConfig):
for item in inplace_dict:
for sub_key_idx in range(key_idx + 1, len_keys - 1):
item = item[keys[sub_key_idx]]
item[keys[-1]] = value
return
inplace_dict[keys[-1]] = value
is_test = False
if p := args.test_ckpt_path:
# load test model config
config = OmegaConf.load(Path(p).parent.parent / "hparams.yaml").all_config
is_test = True
elif p := args.load_ckpt_path:
# load pretrained ckpt config
# config.model = OmegaConf.load(Path(p).parent.parent / 'hparams.yaml').all_config.model
pass
# set unknown args to config
for unknown in unknown_args:
k, v = unknown.split("=")
v = v.strip("'")
vlower = v.lower()
if vlower == "none" or vlower == "~":
v = None
else:
try:
v = json.loads(vlower)
except json.decoder.JSONDecodeError:
pass # v = v, the str itself
set_config_key_value(config, k, v)
# devices
if (devices := args.devices) is not None:
if devices == "all":
devices = ",".join([str(i) for i in range(torch.cuda.device_count())])
config.trainer.devices = [int(rank) for rank in devices.split(",")]
if is_test:
return config
# ++ begin of training configuration ++#
# set project name and signature for logging
if args.no_log:
config.trainer.logger = False
else:
config.trainer.logger.save_dir = f"logs/{args.model}"
config.trainer.logger.name = f"{args.dataset}-{config.data_module.dataset_name}"
config.trainer.logger.version = (
start_time
+ "_"
+ str(random.randint(100000, 999999))
+ (f"_{args.log_suffix}" if args.log_suffix != "" else "")
)
# batch size for ddp
total_bs = config.dataloader.batch_size
num_devices = len(config.trainer.devices)
bs_per_device = total_bs // num_devices
real_bs = bs_per_device * num_devices
if real_bs != total_bs:
logger.warning(f"real batch size is {real_bs}")
config.dataloader.batch_size = bs_per_device
# epoch scaling
epoch_scaling = config.data_module.get("epoch_scaling")
if epoch_scaling is not None and epoch_scaling != 1:
config.trainer.max_epochs = int(config.trainer.max_epochs / epoch_scaling)
logger.info(
f"Training epoch length is scaled by {epoch_scaling}, thus the num of epochs is decreased to {config.trainer.max_epochs}"
)
# customize anything here
config = preprocess_config_hook(config)
return config
def preprocess_config_hook(config):
return config
def get_processed_args_and_config():
args, unknown_args = get_args()
OmegaConf.register_new_resolver("eval", ast.literal_eval)
# load trainer config
trainer_config = OmegaConf.load(f"src/configs/trainer/{args.trainer}.yaml")
OmegaConf.resolve(trainer_config)
# load model config
model_config = OmegaConf.load(f"src/configs/models/{args.model}.yaml")
OmegaConf.resolve(model_config)
config = OmegaConf.merge(trainer_config, model_config)
# load dataset config
dataset_config = OmegaConf.load(f"src/configs/datasets/{args.dataset}.yaml")
OmegaConf.resolve(dataset_config)
config = OmegaConf.merge(config, DictConfig(dataset_config))
config = _preprocess_config(config, args, unknown_args)
# merge args into config
config = OmegaConf.merge(
config,
OmegaConf.create({"args": vars(args), "unkown_args": {x.split("=")[0]: x.split("=")[1] for x in unknown_args}}),
)
if (not dist.is_initialized()) or dist.get_rank() == 0:
logger.info(f"running with config: {config}")
return args, config
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument("--model", type=str, default="colar")
parser.add_argument("--dataset", type=str, default="qsa")
parser.add_argument("--trainer", type=str, default="default")
parser.add_argument("--devices", type=str, default="0")
parser.add_argument("--no_log", help="disable training log", action="store_true")
parser.add_argument("--log_suffix", type=str, help="add suffix to log dir", default="")
parser.add_argument("--resume_ckpt_path", type=str, help="resume training from ckpt", default=None)
parser.add_argument("--load_ckpt_path", type=str, help="load ckpt as initialization", default=None)
parser.add_argument("--workspace_path", type=str, help="assign the path of user workspace directory", default="/workspace/images-ks3-starfs/workspace/wenhui")
parser.add_argument("--do_test", help="test after training", action="store_true")
parser.add_argument("--test_ckpt_path", default="")
parser.add_argument("--test_times", type=int, default=1)
parser.add_argument("--seed", type=int, default=0)
args, unknown_args = parser.parse_known_args()
return args, unknown_args
def main():
args, config = get_processed_args_and_config()
pl.seed_everything(args.seed)
os.environ["TOKENIZERS_PARALLELISM"] = "false"
data_module: pl.LightningDataModule = instantiate_from_config(
config.data_module, extra_kwargs={"all_config": config}
)
model: pl.LightningModule = instantiate_from_config(config.model, extra_kwargs={"all_config": config})
if p := args.load_ckpt_path:
logger.info(model.load_state_dict(state_dict=torch.load(p, map_location="cpu", weights_only=False)["state_dict"], strict=False))
trainer: pl.Trainer = instantiate_from_config(
config.trainer, extra_kwargs={"callbacks": instantiate_callbacks(config.callbacks)}
)
# test only
if p := args.test_ckpt_path:
print(do_test(model=model, trainer=trainer, ckpt_path=p, data_module=data_module, args=args))
return
# training
try:
if trainer.global_rank == 0:
shutil.copytree("src", os.path.join(trainer.logger.log_dir, "src_backup")) # backup src directory
except AttributeError:
pass
trainer.fit(model=model, datamodule=data_module, ckpt_path=args.resume_ckpt_path)
# test after training
if args.do_test:
print(do_test(model=model, trainer=trainer, ckpt_path="best", data_module=data_module, args=args))
if __name__ == "__main__":
main()