-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathfuncs.py
196 lines (172 loc) · 6.88 KB
/
funcs.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
import numpy as np
import torch
from tqdm import tqdm
from utils import *
import torch.distributed as dist
import random
import os
from spikingjelly.activation_based import functional
from torch.utils.tensorboard import SummaryWriter
def seed_all(seed=42):
print(seed)
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
def amp_train_ann(train_dataloader, test_dataloader, model,
epochs, device, loss_fn,lr=0.1,lr_min=1e-5,wd=5e-4 , save=None, parallel=False,
rank=0):
use_amp=True
if rank==0:
with open('./runs/'+save+'_log.txt','a') as log:
log.write('lr={},epochs={},wd={}\n'.format(lr,epochs,wd))
model.cuda(device)
optimizer = torch.optim.SGD(model.parameters(),
lr=lr, weight_decay=wd, momentum=0.9)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer,eta_min=lr_min, T_max=epochs)
scaler = torch.cuda.amp.GradScaler(enabled=use_amp)
best_acc=0.
for epoch in range(epochs):
model.train()
if parallel:
train_dataloader.sampler.set_epoch(epoch)
epoch_loss = 0
length = 0
model.train()
for img, label in tqdm(train_dataloader):
img = img.to(device)
label = label.to(device)
with torch.autocast(device_type='cuda', dtype=torch.float16, enabled=use_amp):
out = model(img)
loss = loss_fn(out, label)
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
epoch_loss += loss.item()
length += len(label)
tmp_acc, val_loss = eval_ann(test_dataloader, model, loss_fn, device, rank)
if parallel:
dist.all_reduce(tmp_acc)
tmp_acc/=dist.get_world_size()
if rank == 0 and save != None and tmp_acc >= best_acc:
checkpoint = {"model": model.state_dict(),
"optimizer": optimizer.state_dict(),
"scaler": scaler.state_dict()}
torch.save(checkpoint, './saved_models/' + save + '.pth')
if rank == 0:
info='Epoch:{},Train_loss:{},Val_loss:{},Acc:{}'.format(epoch, epoch_loss/length,val_loss, tmp_acc.item())
with open('./runs/'+save+'_log.txt','a') as log:
log.write(info+'\n')
if epoch % 10 == 0:
print(model)
best_acc = max(tmp_acc, best_acc)
scheduler.step()
return best_acc, model
def train_ann(train_dataloader, test_dataloader, model,
epochs, device, loss_fn,lr=0.1,lr_min=1e-6,wd=5e-4 , save=None, parallel=False,
rank=0):
# model.cuda(device)
# writer = SummaryWriter('./runs/'+save)
# mt=monitor.InputMonitor(model,SteppedReLU)
# qcfs_vth={}
# cnt=1
# for name in mt.monitored_layers:
# qcfs=get_module_by_name(model,name)[1]
# #assert isinstance(qcfs,QCFS)
# qcfs_vth[str(cnt)+'+'+name]=qcfs.v_threshold
# #qcfs_p0[str(cnt)+'+'+name]=qcfs.p0
# cnt=cnt+1
# mt.clear_recorded_data()
# mt.remove_hooks()
if parallel:
wd=1e-4
if rank==0:
with open('./runs/'+save+'_log.txt','a') as log:
log.write('lr={},epochs={},wd={}\n'.format(lr,epochs,wd))
optimizer = torch.optim.SGD(model.parameters(),
lr=lr, weight_decay=wd, momentum=0.9)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer,eta_min=lr_min, T_max=epochs)
best_acc=eval_ann(test_dataloader, model, loss_fn, device, rank)[0]
if parallel:
dist.all_reduce(best_acc)
best_acc/=dist.get_world_size()
if rank==0:
print(best_acc)
for epoch in tqdm(range(epochs)):
model.train()
if parallel:
train_dataloader.sampler.set_epoch(epoch)
epoch_loss = 0
length = 0
model.train()
for img, label in tqdm(train_dataloader):
img = img.to(device)
label = label.to(device)
optimizer.zero_grad()
out = model(img)
loss = loss_fn(out, label)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
length += len(label)
tmp_acc, val_loss = eval_ann(test_dataloader, model, loss_fn, device, rank)
if parallel:
dist.all_reduce(tmp_acc)
tmp_acc/=dist.get_world_size()
if rank == 0 and save != None and tmp_acc >= best_acc:
torch.save(model.state_dict(), './saved_models/' + save + '.pth')
if rank == 0:
info='Epoch:{},Train_loss:{},Val_loss:{},Acc:{},lr:{}'.format(epoch, epoch_loss/length,val_loss, tmp_acc.item(),scheduler.get_last_lr()[0])
with open('./runs/'+save+'_log.txt','a') as log:
log.write(info+'\n')
best_acc = max(tmp_acc, best_acc)
# print('Epoch:{},Train_loss:{},Val_loss:{},Acc:{}'.format(epoch, epoch_loss/length,val_loss, tmp_acc), flush=True)
# print(f'lr={scheduler.get_last_lr()[0]}')
# print('best_acc: ', best_acc)
# writer.add_scalars('Acc',{'val_acc':tmp_acc,'best_acc':best_acc},epoch)
# writer.add_scalars('Loss',{'train_loss':epoch_loss/length,'val_loss':val_loss},epoch)
# writer.add_scalar('lr',scheduler.get_last_lr()[0],epoch)
# writer.add_scalars('vth',qcfs_vth,epoch)
scheduler.step()
#print(module)
# writer.close()
return best_acc, model
def eval_snn(test_dataloader, model,loss_fn, device, sim_len=8, rank=0):
tot = torch.zeros(sim_len).cuda()
length = 0
model = model.cuda()
model.eval()
with torch.no_grad():
for idx, (img, label) in enumerate(tqdm((test_dataloader))):
spikes = 0
length += len(label)
img = img.cuda()
label = label.cuda()
for t in range(sim_len):
out = model(img)
spikes += out
tot[t] += (label==spikes.max(1)[1]).sum()
spikes/=sim_len
loss = loss_fn(spikes, label)
functional.reset_net(model)
return (tot/length),loss.item()/length
def eval_ann(test_dataloader, model, loss_fn, device, rank=0):
epoch_loss = 0
tot = torch.tensor(0.).cuda(device)
model.eval()
model.cuda(device)
length = 0
with torch.no_grad():
for img, label in tqdm(test_dataloader):
img = img.cuda(device)
label = label.cuda(device)
out = model(img)
loss = loss_fn(out, label)
epoch_loss += loss.item()
length += len(label)
tot += (label==out.max(1)[1]).sum().data
return (tot/length), epoch_loss/length