-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
275 lines (235 loc) · 11.1 KB
/
train.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
import torch
from torch.utils.tensorboard import SummaryWriter
from torch.utils.data.distributed import DistributedSampler
from sacred import Experiment
import datetime
from tqdm import tqdm
from pathlib import Path
import os
import numpy as np
from warmup_scheduler import GradualWarmupScheduler
from lib.datasets import ds
from lib.datasets import HdF5Dataset
from lib.third_party.clevrtex_eval import CLEVRTEX, collate_fn, texds
from lib.model import net
from lib.model import EfficientMORL, SlotAttention
from lib.geco import GECO
from lib.visualization import visualize_output, visualize_slots
from sacred.observers import FileStorageObserver
ex = Experiment('TRAINING', ingredients=[ds,texds, net])
local_rank = os.environ['LOCAL_RANK']
if local_rank == '0':
# optionally set the env variable "SACRED_OBSERVATORY"
# to a path to store sacred run files
sacred_run_dir = os.environ['SACRED_OBSERVATORY']
if sacred_run_dir != '':
ex.observers.append(FileStorageObserver(sacred_run_dir))
@ex.config
def cfg():
training = {
'DDP_port': 29500, # torch.distributed config
'batch_size': 16, # training mini-batch size
'num_workers': 8, # pytorch dataloader workers
'mode': 'train', # dataset
'model': 'EfficientMORL', # model name
'iters': 500000, # gradient steps to take
'refinement_curriculum': [(-1,3), (100000,1), (200000,1)], # (step,I): Update refinement iters I at step
'lr': 3e-4, # Adam LR
'warmup': 10000, # LR warmup
'decay_rate': 0.5, # LR decay
'decay_steps': 100000, # LR decay steps
'kl_beta_init': 1, # kl_beta from beta-VAE
'use_scheduler': False, # LR scheduler
'tensorboard_freq': 100, # how often to write to TB
'checkpoint_freq': 25000, # save checkpoints every % steps
'load_from_checkpoint': False, # whether to load from a checkpoint or not
'checkpoint': '', # name of .pth file to load model state
'run_suffix': 'debug', # string to append to run name
'out_dir': 'experiments', # where output folders for run results go
'use_geco': True, # Use GECO (Rezende & Viola 2018)
'clip_grad_norm': True, # Grad norm clipping to 5.0
'geco_reconstruction_target': -23000, # GECO C
'geco_step_size_acceleration': 1, # multiplies beta once the target is reached
'geco_ema_alpha': 0.99, # GECO EMA step parameter
'geco_beta_stepsize': 1e-6, # GECO Lagrange parameter beta
'tqdm': False # Show training progress in CLI
}
def save_checkpoint(step, kl_beta, model, model_opt, filepath):
state = {
'step': step,
'model': model.state_dict(),
'model_opt': model_opt.state_dict(),
'kl_beta': kl_beta
}
torch.save(state, filepath)
@ex.automain
def run(training, seed, _run):
global local_rank
run_dir = Path(training['out_dir'], 'runs')
checkpoint_dir = Path(training['out_dir'], 'weights')
tb_dir = Path(training['out_dir'], 'tb')
# Avoid issues with torch distributed and just create directory structure
# beforehand
# training['out_dir']/runs
# training['out_dir']/weights
# training['out_dir']/tb
for dir_ in [run_dir, checkpoint_dir, tb_dir]:
if not dir_.exists():
print(f'Create {dir_} before running!')
exit(1)
tb_dbg = tb_dir / (training['run_suffix'] + '_' + \
datetime.datetime.now().strftime("%Y-%m-%d--%H-%M-%S"))
#local_rank = 'cuda:{}'.format(_run.info['local_rank'])
local_rank = f'cuda:{local_rank}'
if local_rank == 'cuda:0':
print(f'Creating SummaryWriter! ({local_rank})')
writer = SummaryWriter(tb_dbg)
# Fix random seed
print(f'Local rank: {local_rank}. Setting random seed to {seed}')
# Auto-set by sacred
# torch.manual_seed(seed)
torch.backends.cudnn.deterministic=True
torch.backends.cudnn.benchmark=False
# Auto-set by sacred
#np.random.seed(seed)
os.environ['MASTER_ADDR'] = '127.0.0.1'
os.environ['MASTER_PORT'] = str(training['DDP_port'])
torch.distributed.init_process_group(backend='nccl')
torch.cuda.set_device(local_rank)
if training['model'] == 'EfficientMORL':
model = EfficientMORL(batch_size=training['batch_size'],
use_geco=training['use_geco'])
elif training['model'] == 'SlotAttention':
model = SlotAttention(batch_size=training['batch_size'])
else:
raise RuntimeError('Model {} unknown'.format(training['model']))
model_geco = None
if training['use_geco']:
C, H, W = model.input_size
recon_target = training['geco_reconstruction_target'] * (C * H * W)
model_geco = GECO(recon_target,
training['geco_ema_alpha'],
training['geco_step_size_acceleration'])
model = model.to(local_rank)
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[local_rank],
output_device=local_rank,
find_unused_parameters=True)
model.train()
if local_rank == 'cuda:0':
print(model)
# Optimization
model_opt = torch.optim.Adam(model.parameters(), lr=training['lr'])
if training['use_scheduler']:
scheduler = torch.optim.lr_scheduler.LambdaLR(
model_opt,
lr_lambda=lambda epoch: 0.5 ** (epoch / 100000)
)
scheduler_warmup = GradualWarmupScheduler(model_opt, multiplier=1,
total_epoch=training['warmup'],
after_scheduler=scheduler)
else:
scheduler_warmup = None
if not training['load_from_checkpoint']:
step = 0
kl_beta = training['kl_beta_init']
else:
checkpoint = checkpoint_dir / training['checkpoint']
map_location = {'cuda:0': local_rank}
state = torch.load(checkpoint, map_location=map_location)
model.load_state_dict(state['model'])
model_opt.load_state_dict(state['model_opt'])
kl_beta = state['kl_beta']
step = state['step']
if "clevrtex" in training['run_suffix']:
tr_dataset = CLEVRTEX(dataset_variant='full',
split='train',
crop=True,
resize=(128,128),
return_metadata=False)
else:
tr_dataset = HdF5Dataset(d_set=training['mode'])
batch_size = training['batch_size']
tr_sampler = DistributedSampler(dataset=tr_dataset)
def worker_init_fn(worker_id):
np.random.seed(np.random.get_state()[1][0] + worker_id)
tr_dataloader = torch.utils.data.DataLoader(tr_dataset,
batch_size=batch_size, sampler=tr_sampler,
num_workers=training['num_workers'],
worker_init_fn=worker_init_fn,
drop_last=True)
max_iters = training['iters']
print('Num parameters: {}'.format(sum(p.numel() for p in model.parameters())))
epoch_idx = 0
while step <= max_iters:
# Re-shuffle every epoch
tr_sampler.set_epoch(epoch_idx)
if training['tqdm'] and local_rank == 'cuda:0':
data_iter = tqdm(tr_dataloader)
else:
data_iter = tr_dataloader
for batch in data_iter:
# Update refinement iterations by curriculum
for rf in range(len(training['refinement_curriculum'])-1,-1,-1):
if step >= training['refinement_curriculum'][rf][0]:
model.module.refinement_iters = training['refinement_curriculum'][rf][1]
break
if 'clevrtex' in training['run_suffix']:
_, img_batch, mask_batch = batch
img_batch = img_batch.to(local_rank)
img_batch = (img_batch * 2) - 1
else:
img_batch = batch['imgs'].to(local_rank)
model_opt.zero_grad()
# Forward
if training['model'] == 'SlotAttention':
out_dict = model(img_batch)
else:
out_dict = model(img_batch, model_geco, step, kl_beta)
# Backward
total_loss = out_dict['total_loss']
total_loss.backward()
if training['use_scheduler']:
scheduler_warmup.step(step)
if training['clip_grad_norm']:
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=5.)
model_opt.step()
if training['use_geco']:
if step == model.module.geco_warm_start:
model.module.geco_C_ema = model_geco.init_ema(model.module.geco_C_ema, out_dict['nll'])
elif step > model.module.geco_warm_start:
model.module.geco_C_ema = model_geco.update_ema(model.module.geco_C_ema, out_dict['nll'])
model.module.geco_beta = model_geco.step_beta(model.module.geco_C_ema,
model.module.geco_beta, training['geco_beta_stepsize'])
# logging
if step % training['tensorboard_freq'] == 0 and local_rank == 'cuda:0':
if training['model'] == 'SlotAttention':
writer.add_scalar('train/MSE', total_loss, step)
visualize_slots(writer, (img_batch+1)/2., out_dict, step)
else:
writer.add_scalar('train/total_loss', total_loss, step)
writer.add_scalar('train/KL', out_dict['kl'], step)
writer.add_scalar('train/KL_beta', kl_beta, step)
writer.add_scalar('train/NLL', out_dict['nll'], step)
visualize_output(writer, (img_batch+1)/2., out_dict,
model.module.stochastic_layers, model.module.refinement_iters,
step)
if training['use_geco']:
writer.add_scalar('train/geco_beta', model.module.geco_beta, step)
writer.add_scalar('train/geco_C_ema', model.module.geco_C_ema, step)
if 'deltas' in out_dict:
for refine_iter in range(out_dict['deltas'].shape[0]):
writer.add_scalar(f'train/norm_delta_lamda_{refine_iter}',
out_dict['deltas'][refine_iter], step)
if (step > 0 and step % training['checkpoint_freq'] == 0 and
local_rank == 'cuda:0'):
# Save the model
prefix = training['run_suffix']
save_checkpoint(step, kl_beta, model, model_opt,
checkpoint_dir / f'{prefix}-state-{step}.pth')
if step >= max_iters:
step += 1
break
step += 1
epoch_idx += 1
if local_rank == 'cuda:0':
writer.close()