-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathmain.py
204 lines (174 loc) · 7.64 KB
/
main.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
import argparse
import importlib
import os
from copy import deepcopy
import wandb
import torch
from accelerate import Accelerator
from accelerate.utils import set_seed
from cleanfid import fid
import global_v as glv
from network_parser import parse
from utils import get_git_info, add_hook
from envs import Trainer
import snn_layers
from fad.autoencoder_fid import compute_autoencoder_frechet_distance
def train(trainer: Trainer, start_epoch: int, args: argparse.Namespace,
accelerator: Accelerator):
for epoch in range(start_epoch, start_epoch + glv.network_config['epochs']):
train_metrics = trainer.train(epoch)
test_metrics = trainer.test(epoch)
if accelerator.is_main_process:
trainer.save_model(epoch,
'./debug' if args.debug else wandb.run.dir)
metrics = {**train_metrics, **test_metrics}
if not args.debug:
wandb.log(metrics, step=epoch, commit=True)
print(f'epoch: {epoch}, ' + ', '.join(
k + ': ' + f'{v:.4f}'
for k, v in metrics.items() if isinstance(v, (int, float))))
if glv.network_config.get('calc_metrics_at_last', True):
metrics = trainer.calc_final_metrics()
if accelerator.is_main_process:
if not args.debug:
if metrics:
wandb.log(metrics)
def calc_metrics(trainer: Trainer, args: argparse.Namespace, checkpoint: dict,
accelerator: Accelerator):
if args.generated_image_dir:
output_dir = args.generated_image_dir
else:
epoch = checkpoint['epoch']
output_dir = os.path.join('output', args.name, 'img', str(epoch))
if accelerator.is_main_process:
os.makedirs(output_dir, exist_ok=args.debug)
metrics_config = glv.network_config['metrics']
metrics = {}
if metrics_config['sample']:
assert checkpoint is not None
trainer.sample(output_dir, metrics_config.get('num_images', None))
if not accelerator.is_main_process:
return
if metrics_config['fid']:
name = glv.network_config['dataset'].lower()
if 'image_size' in glv.network_config:
name += '-' + str(glv.network_config['image_size'])
if 'center_crop' in glv.network_config:
name += '-crop' + str(glv.network_config['center_crop'])
fid_score = fid.compute_fid(fdir1=output_dir,
batch_size=metrics_config.get(
'batch_size', 128),
device=trainer.device,
dataset_name=name,
dataset_split='custom',
num_workers=0,
use_dataparallel=False)
metrics['fid'] = fid_score
if metrics_config['fad']:
name = glv.network_config['dataset'].lower()
fad_score = compute_autoencoder_frechet_distance(
name,
fdir=output_dir,
device=trainer.device,
num_gen=metrics_config.get('num_images', 10000),
batch_size=metrics_config.get('batch_size', 128))
metrics['fad'] = fad_score
if metrics_config['calc_mul_add']:
count_mul_add, hook_handles = add_hook(trainer.model)
trainer.model.eval()
dummy_input = torch.randn(1,
glv.network_config['in_channels'],
glv.network_config['image_size'],
glv.network_config['image_size'],
glv.network_config['n_steps'],
device=trainer.device)
trainer.model(dummy_input, torch.tensor([0], device=trainer.device))
count_mul_add.clear()
sample_num = metrics_config.get('calc_mul_add_sample', 100)
trainer.diffusion_sample(sample_num)
mul_sum = count_mul_add.mul_sum / sample_num
add_sum = count_mul_add.add_sum / sample_num
mac_sum = count_mul_add.mac_sum / sample_num
ac_sum = count_mul_add.ac_sum / sample_num
for handle in hook_handles:
handle.remove()
metrics['mul_add_num/mul'] = mul_sum
metrics['mul_add_num/add'] = add_sum
metrics['mul_add_num/mac'] = mac_sum
metrics['mul_add_num/ac'] = ac_sum
print('mul:', mul_sum)
print('add:', add_sum)
print('mac:', mac_sum)
print('ac:', ac_sum)
if not args.debug:
wandb.log(metrics, commit=True)
print(', '.join(k + ': ' + f'{v:.4f}' for k, v in metrics.items()
if isinstance(v, (int, float))))
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-c', '--config', action='store')
parser.add_argument('-d', '--gpu-id', type=int, default=0)
parser.add_argument('-s', '--seed', type=int, default=0)
parser.add_argument('-n', '--name', type=str, default='test')
parser.add_argument('-m', '--model-checkpoint', type=str, default='')
parser.add_argument('-i', '--resume-id', type=str)
parser.add_argument('-o', '--generated-image-dir', type=str, default='')
parser.add_argument('--debug', action='store_true')
args = parser.parse_args()
accelerator = Accelerator()
set_seed(args.seed + accelerator.process_index)
params = parse(args.config)
network_config = deepcopy(params['network'])
if accelerator.is_main_process:
print(network_config)
glv.init(network_config, accelerator.num_processes)
snn_layers.init_layer_config(glv.layer_config)
models = {}
for model_name, model_config in network_config['models'].items():
module, cls = model_config['class'].rsplit(".", 1)
Model = getattr(importlib.import_module(module), cls)
model = Model(**model_config.get('args', {}))
models[model_name] = model
start_epoch = 0
if args.model_checkpoint:
checkpoint = torch.load(args.model_checkpoint,
map_location=torch.device('cpu'))
if args.resume_id is not None:
start_epoch = checkpoint['epoch'] + 1
else:
checkpoint = None
if args.resume_id is not None:
raise ValueError(
'--resume-id is specified but --model-checkpoint is not specified'
)
if params['experiment_type'] == 'diffusion':
trainer = Trainer(models, network_config, checkpoint)
else:
ValueError('invalid experiment type', params['experiment_type'])
if accelerator.is_main_process and not args.debug:
if args.resume_id is not None:
kwargs = dict(resume='must', id=args.resume_id)
else:
kwargs = {}
wandb.init(dir=network_config.get('wandb_dir', None),
project='fsddim',
name=args.name,
config={
'original': params['network'],
'adjusted': network_config
},
**kwargs)
commit, diff = get_git_info()
table = wandb.Table(columns=['title', 'contents'],
data=[['commit', commit], ['diff', diff]])
wandb.log({'git_info': table}, step=0)
for model in models.values():
wandb.watch(model)
else:
os.makedirs('./debug', exist_ok=True)
train(trainer, start_epoch, args, accelerator)
checkpoint = torch.load(os.path.join(
'./debug' if args.debug else wandb.run.dir,
'best_model_checkpoint.pth'),
map_location=torch.device('cpu'))
calc_metrics(trainer, args, checkpoint, accelerator)