-
Notifications
You must be signed in to change notification settings - Fork 4
/
finetune.py
309 lines (241 loc) · 13.2 KB
/
finetune.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
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
import sys
import os
sys.path.append(['.','./../'])
os.environ['OMP_NUM_THREADS'] = '16'
import json
import time
import argparse
import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
import matplotlib.pyplot as plt
import operator
from functools import reduce
from functools import partial
from timeit import default_timer
from torch.optim.lr_scheduler import OneCycleLR, StepLR, LambdaLR, CosineAnnealingWarmRestarts, CyclicLR
from torch.utils.tensorboard import SummaryWriter
from utils.optimizer import Adam, Lamb
from utils.utilities import count_parameters, get_grid, load_model_from_checkpoint, load_components_from_pretrained
from utils.criterion import SimpleLpLoss
from utils.griddataset import MixedTemporalDataset
from utils.make_master_file import DATASET_DICT
from models.fno import FNO2d
from models.dpot import DPOTNet
import pickle
# torch.manual_seed(0)
# np.random.seed(0)
################################################################
# configs
################################################################
parser = argparse.ArgumentParser(description='Training or pretraining for the same data type')
### currently no influence
parser.add_argument('--model', type=str, default='DPOT')
parser.add_argument('--dataset',type=str, default='ns2d')
parser.add_argument('--train_paths',nargs='+', type=str, default=['swe_pdb'])
parser.add_argument('--test_paths',nargs='+',type=str, default=['swe_pdb'])
parser.add_argument('--resume_path',type=str, default='')
parser.add_argument('--ntrain_list', nargs='+', type=int, default=[900])
parser.add_argument('--data_weights',nargs='+',type=int, default=[1])
parser.add_argument('--use_writer', action='store_true',default=False)
parser.add_argument('--res', type=int, default=64)
parser.add_argument('--noise_scale',type=float, default=0.0)
# parser.add_argument('--n_channels',type=int,default=-1)
### shared params
parser.add_argument('--width', type=int, default=20)
parser.add_argument('--n_layers',type=int, default=4)
parser.add_argument('--act',type=str, default='gelu')
### GNOT params
parser.add_argument('--max_nodes',type=int, default=-1)
### FNO params
parser.add_argument('--modes', type=int, default=12)
parser.add_argument('--use_ln',type=int, default=0)
parser.add_argument('--normalize',type=int, default=0)
### AFNO
parser.add_argument('--patch_size',type=int, default=1)
parser.add_argument('--n_blocks',type=int, default=8)
parser.add_argument('--mlp_ratio',type=int, default=1)
parser.add_argument('--out_layer_dim', type=int, default=32)
parser.add_argument('--batch_size', type=int, default=20)
parser.add_argument('--epochs', type=int, default=500)
parser.add_argument('--lr', type=float, default=0.001)
parser.add_argument('--opt',type=str, default='adam', choices=['adam','lamb'])
parser.add_argument('--beta1',type=float,default=0.9)
parser.add_argument('--beta2',type=float,default=0.9)
parser.add_argument('--lr_method',type=str, default='step')
parser.add_argument('--grad_clip',type=float, default=10000.0)
parser.add_argument('--step_size', type=int, default=100)
parser.add_argument('--step_gamma', type=float, default=0.5)
parser.add_argument('--warmup_epochs',type=int, default=50)
parser.add_argument('--sub', type=int, default=1)
parser.add_argument('--T_in', type=int, default=10)
parser.add_argument('--T_ar', type=int, default=1)
parser.add_argument('--T_bundle', type=int, default=1)
parser.add_argument('--gpu', type=str, default="3")
parser.add_argument('--comment',type=str, default="")
parser.add_argument('--log_path',type=str,default='')
### finetuning parameters
parser.add_argument('--n_channels',type=int, default=4)
parser.add_argument('--n_class',type=int,default=12)
parser.add_argument('--load_components',nargs='+', type=str, default=['blocks','pos','time_agg'])
args = parser.parse_args()
device = torch.device("cuda:{}".format(args.gpu))
print(f"Current working directory: {os.getcwd()}")
################################################################
# load data and dataloader
################################################################
train_paths = args.train_paths
test_paths = args.test_paths
args.data_weights = [1] * len(args.train_paths) if len(args.data_weights) == 1 else args.data_weights
print('args',args)
train_dataset = MixedTemporalDataset(args.train_paths, args.ntrain_list, res=args.res, t_in = args.T_in, t_ar = args.T_ar, normalize=False,train=True, data_weights=args.data_weights, n_channels=args.n_channels)
test_datasets = [MixedTemporalDataset(test_path, res=args.res, n_channels = train_dataset.n_channels,t_in = args.T_in, t_ar=-1, normalize=False, train=False) for i, test_path in enumerate(test_paths)]
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=8)
test_loaders = [torch.utils.data.DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False,num_workers=8) for test_dataset in test_datasets]
ntrain, ntests = len(train_dataset), [len(test_dataset) for test_dataset in test_datasets]
print('Train num {} test num {}'.format(train_dataset.n_sizes, ntests))
################################################################
# load model
################################################################
if args.model == "FNO":
model = FNO2d(args.modes, args.modes, args.width, img_size = args.res, patch_size=args.patch_size, in_timesteps = args.T_in, out_timesteps=1,normalize=args.normalize,n_layers = args.n_layers,use_ln = args.use_ln, n_channels=train_dataset.n_channels, n_cls=len(args.train_paths)).to(device)
elif args.model == 'DPOT':
model = DPOTNet(img_size=args.res, patch_size=args.patch_size, in_channels=train_dataset.n_channels, in_timesteps = args.T_in, out_timesteps = args.T_bundle, out_channels=train_dataset.n_channels, normalize=args.normalize, embed_dim=args.width, modes=args.modes, depth=args.n_layers, n_blocks = args.n_blocks, mlp_ratio=args.mlp_ratio, out_layer_dim=args.out_layer_dim, act=args.act, n_cls=args.n_class).to(device)
else:
raise NotImplementedError
if args.resume_path:
print('Loading models and fine tune from {}'.format(args.resume_path))
load_components_from_pretrained(model, torch.load(args.resume_path,map_location='cuda:{}'.format(args.gpu))['model'], components=args.load_components)
#### set optimizer
if args.opt == 'lamb':
optimizer = Lamb(model.parameters(), lr=args.lr, betas = (args.beta1, args.beta2), adam=True, debias=False,weight_decay=1e-4)
else:
optimizer = Adam(model.parameters(), lr=args.lr, betas=(args.beta1, args.beta2), weight_decay=1e-6)
if args.lr_method == 'cycle':
print('Using cycle learning rate schedule')
scheduler = OneCycleLR(optimizer, max_lr=args.lr, div_factor=1e4, pct_start=(args.warmup_epochs / args.epochs), final_div_factor=1e4, steps_per_epoch=len(train_loader), epochs=args.epochs)
elif args.lr_method == 'step':
print('Using step learning rate schedule')
scheduler = StepLR(optimizer, step_size=args.step_size * len(train_loader), gamma=args.step_gamma)
elif args.lr_method == 'warmup':
print('Using warmup learning rate schedule')
scheduler = LambdaLR(optimizer, lambda steps: min((steps + 1) / (args.warmup_epochs * len(train_loader)), np.power(args.warmup_epochs * len(train_loader) / float(steps + 1), 0.5)))
elif args.lr_method == 'linear':
print('Using warmup learning rate schedule')
scheduler = LambdaLR(optimizer, lambda steps: (1 - steps / (args.epochs * len(train_loader))))
elif args.lr_method == 'restart':
print('Using cos anneal restart')
scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=len(train_loader) * args.lr_step_size, eta_min=0.)
elif args.lr_method == 'cyclic':
scheduler = CyclicLR(optimizer, base_lr=1e-5, max_lr=1e-3, step_size_up=args.lr_step_size * len(train_loader),mode='triangular2', cycle_momentum=False)
else:
raise NotImplementedError
comment = args.comment + '_{}_{}'.format(len(train_paths), ntrain)
log_path = './logs/' + time.strftime('%m%d_%H_%M_%S') + comment if len(args.log_path)==0 else os.path.join('./logs',args.log_path + comment)
model_path = log_path + '/model.pth'
if args.use_writer:
writer = SummaryWriter(log_dir=log_path)
fp = open(log_path + '/logs.txt', 'w+',buffering=1)
json.dump(vars(args), open(log_path + '/params.json', 'w'),indent=4)
sys.stdout = fp
else:
writer = None
print(model)
count_parameters(model)
################################################################
# Main function for pretraining
################################################################
myloss = SimpleLpLoss(size_average=False)
clsloss = torch.nn.CrossEntropyLoss(reduction='sum')
iter = 0
for ep in range(args.epochs):
model.train()
t1 = t_1 = default_timer()
t_load, t_train = 0., 0.
train_l2_step = 0
train_l2_full = 0
cls_total, cls_correct, cls_acc = 0, 0, 0.
loss_previous = np.inf
for xx, yy, msk, cls in train_loader:
t_load += default_timer() - t_1
t_1 = default_timer()
loss, cls_loss = 0. , 0.
xx = xx.to(device) ## B, n, n, T_in, C
yy = yy.to(device) ## B, n, n, T_ar, C
msk = msk.to(device)
cls = cls.to(device)
## auto-regressive training loop, support 1. noise injection, 2. long rollout backward, 3. temporal bundling prediction
for t in range(0, yy.shape[-2], args.T_bundle):
y = yy[..., t:t + args.T_bundle, :]
### auto-regressive training
xx = xx + args.noise_scale *torch.sum(xx**2, dim=(1,2,3),keepdim=True)**0.5 * torch.randn_like(xx)
im, cls_pred = model(xx)
loss += myloss(im, y, mask=msk)
### classification
pred_labels = torch.argmax(cls_pred,dim=1)
cls_loss += clsloss(cls_pred, cls.squeeze())
cls_correct += (pred_labels == cls.squeeze()).sum().item()
cls_total += cls.shape[0]
if t == 0:
pred = im
else:
pred = torch.cat((pred, im), dim=-2)
xx = torch.cat((xx[..., args.T_bundle:, :], im), dim=-2)
train_l2_step += loss.item()
l2_full = myloss(pred, yy, mask=msk)
train_l2_full += l2_full.item()
optimizer.zero_grad()
total_loss = loss # + 1.0 * cls_loss
total_loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
optimizer.step()
scheduler.step()
train_l2_step_avg, train_l2_full_avg = train_l2_step / ntrain / (yy.shape[-2] / args.T_bundle), train_l2_full / ntrain
cls_acc = cls_correct / cls_total
iter +=1
if args.use_writer:
writer.add_scalar("train_loss_step", loss.item()/(xx.shape[0] * yy.shape[-2] / args.T_bundle), iter)
writer.add_scalar("train_loss_full", l2_full / xx.shape[0], iter)
## reset model
if loss.item() > 10 * loss_previous : # or (ep > 50 and l2_full / xx.shape[0] > 0.9):
print('loss explodes, loading model from previous epoch')
checkpoint = torch.load(model_path,map_location='cuda:{}'.format(args.gpu))
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint["optimizer"])
loss_previous = loss.item()
t_train += default_timer() - t_1
t_1 = default_timer()
test_l2_fulls, test_l2_steps = [], []
with torch.no_grad():
model.eval()
for id, test_loader in enumerate(test_loaders):
test_l2_full, test_l2_step = 0, 0
for xx, yy, msk, _ in test_loader:
loss = 0
xx = xx.to(device)
yy = yy.to(device)
msk = msk.to(device)
for t in range(0, yy.shape[-2], args.T_bundle):
y = yy[..., t:t + args.T_bundle, :]
im, _ = model(xx)
loss += myloss(im, y, mask=msk)
if t == 0:
pred = im
else:
pred = torch.cat((pred, im), -2)
xx = torch.cat((xx[..., args.T_bundle:,:], im), dim=-2)
test_l2_step += loss.item()
test_l2_full += myloss(pred, yy, mask=msk)
test_l2_step_avg, test_l2_full_avg = test_l2_step / ntests[id] / (yy.shape[-2] / args.T_bundle), test_l2_full / ntests[id]
test_l2_steps.append(test_l2_step_avg)
test_l2_fulls.append(test_l2_full_avg)
if args.use_writer:
writer.add_scalar("test_loss_step_{}".format(test_paths[id]), test_l2_step_avg, ep)
writer.add_scalar("test_loss_full_{}".format(test_paths[id]), test_l2_full_avg, ep)
if args.use_writer:
torch.save({'args': args, 'model': model.state_dict(), 'optimizer': optimizer.state_dict()}, model_path)
t_test = default_timer() - t_1
t2 = t_1 = default_timer()
lr = optimizer.param_groups[0]['lr']
print('epoch {}, time {:.5f}, lr {:.2e}, train l2 step {:.5f} train l2 full {:.5f}, test l2 step {} test l2 full {}, cls acc {:.5f}, time train avg {:.5f} load avg {:.5f} test {:.5f}'.format(ep, t2 - t1, lr,train_l2_step_avg, train_l2_full_avg,', '.join(['{:.5f}'.format(val) for val in test_l2_steps]),', '.join(['{:.5f}'.format(val) for val in test_l2_fulls]), cls_acc, t_train / len(train_loader), t_load / len(train_loader), t_test))