-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
388 lines (318 loc) · 13.9 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
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
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
from __future__ import division
from model import *
from dataset import *
import os
import sys
import argparse
from torch import nn
from torch.utils import data
import lr_scheduler
from model import CrossEntropyLoss2d, PB_FCN, pruneModelNew, PB_FCN_2, DiceLoss
from dataset import SSYUVDataset
from transform import Scale, ToLabel, HorizontalFlip, VerticalFlip, ToYUV, maskLabel
from torchvision.transforms import Compose, Normalize, ToTensor, ColorJitter
import torch.optim as optim
torch.set_printoptions(precision=2,sci_mode=False)
import progressbar
def l1reg(model):
regularization_loss = 0
for param in model.parameters():
regularization_loss += torch.sum(torch.abs(param))
return regularization_loss
def train(epoch,epochs,bestLoss,indices = None):
#############
####TRAIN####
#############
lossreg = 0
losstotal = 0
running_acc = 0
imgCnt = 0
model.train()
bar = progressbar.ProgressBar(0, len(trainloader), redirect_stdout=False)
for batch_i, (imgs, targets) in enumerate(trainloader):
imgs = imgs.type(Tensor)
targets = targets.type(LongTensor).long()
targets = maskLabel(targets, nb, nr, ng, nl)
optimizer.zero_grad()
pred = model(imgs)
loss = criterion(pred,targets)
reg = Tensor([0.0])
if indices is None:
reg = decay * l1reg(model)
loss += reg
loss.backward()
if indices is not None:
pIdx = 0
for param in model.parameters():
if param.dim() > 1:
if param.grad is not None:
param.grad[indices[pIdx]] = 0
pIdx += 1
optimizer.step()
bar.update(batch_i)
_, predClass = torch.max(pred, 1)
running_acc += torch.sum(predClass == targets).item() * outSize * 100
lossreg += reg.item()
losstotal += loss.item()
imgCnt += imgs.shape[0]
bar.finish()
prune = count_zero_weights(model)
print(
"[Epoch Train %d/%d lr: %.4f][Losses: reg %f, pruned %f, total %f][Pixel Acc: %f]"
% (
epoch + 1,
epochs,
scheduler.get_lr()[-1]/learning_rate,
lossreg / float(len(trainloader)),
prune,
losstotal / float(len(trainloader)),
running_acc/(imgCnt),
)
)
if indices is None:
scheduler.step()
return bestLoss
def valid(epoch,epochs,bestLoss,pruned):
#############
####VALID####
#############
model.eval()
lossreg = 0
losstotal = 0
running_acc = 0
imgCnt = 0
conf = torch.zeros(numClass,numClass)
IoU = torch.zeros(numClass)
labCnts = torch.zeros(numClass)
bar = progressbar.ProgressBar(0, len(valloader), redirect_stdout=False)
for batch_i, (imgs, targets) in enumerate(valloader):
imgs = imgs.type(Tensor)
targets = targets.type(LongTensor)
targets = maskLabel(targets, nb, nr, ng, nl)
pred = model(imgs)
loss = criterion(pred, targets)
reg = Tensor([0.0])
if indices is None:
reg = decay * l1reg(model)
loss += reg
bar.update(batch_i)
_, predClass = torch.max(pred, 1)
running_acc += torch.sum(predClass == targets).item() * outSize * 100
lossreg += reg.item()
losstotal += loss.item()
bSize = imgs.shape[0]
imgCnt += bSize
maskPred = torch.zeros(numClass, bSize, int(labSize[0]), int(labSize[1])).long()
maskTarget = torch.zeros(numClass, bSize, int(labSize[0]), int(labSize[1])).long()
for currClass in range(numClass):
maskPred[currClass] = predClass == currClass
maskTarget[currClass] = targets == currClass
for imgInd in range(bSize):
for labIdx in range(numClass):
labCnts[labIdx] += torch.sum(maskTarget[labIdx, imgInd]).item()
for predIdx in range(numClass):
inter = torch.sum(maskPred[predIdx, imgInd] & maskTarget[labIdx, imgInd]).item()
conf[(predIdx, labIdx)] += inter
if labIdx == predIdx:
union = torch.sum(maskPred[predIdx, imgInd] | maskTarget[labIdx, imgInd]).item()
if union == 0:
IoU[labIdx] += 1
else:
IoU[labIdx] += inter / union
bar.finish()
prune = count_zero_weights(model)
for labIdx in range(numClass):
for predIdx in range(numClass):
conf[(predIdx, labIdx)] /= (labCnts[labIdx] / 100.0)
meanClassAcc = 0.0
meanIoU = torch.sum(IoU / imgCnt).item() / numClass * 100
for j in range(numClass):
meanClassAcc += conf[(j, j)] / numClass
currLoss = (meanClassAcc+meanIoU)/2
print(
"[Epoch Val %d/%d lr: %.4f][Losses: reg %f, pruned %f, total %f][Pixel Acc: %f, Mean Class Acc: %f, Mean IoU: %f]"
% (
epoch + 1,
epochs,
scheduler.get_lr()[-1] / learning_rate,
lossreg / float(len(valloader)),
prune,
losstotal / float(len(valloader)),
running_acc / (imgCnt),
meanClassAcc,
meanIoU,
)
)
name = "bestFinetune" if finetune else "best"
name += v2Str
name += scaleStr
name += unetStr
name += nbStr
name += ngStr
name += nrStr
name += nlStr
name += cameraSaveStr
if transfer != 0:
name += "T%d" % transfer
if pruned:
pruneP = round(prune * 100)
comp = round(sum(model.get_computations(True))/1000000)
name = name + ("%d_%d" %(pruneP,comp))
if bestLoss < (currLoss):
print("Saving best model")
print(conf)
bestLoss = (currLoss)
torch.save(model.state_dict(), "checkpoints/%s.weights" % name)
return bestLoss
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("--finetune", help="Finetuning", action="store_true", default=False)
parser.add_argument("--v2", help="Use v2 architecture", action="store_true", default=False)
parser.add_argument("--noScale", help="Use VGA resolution", action="store_true", default=False)
parser.add_argument("--UNet", help="Use Vanilla U-Net", action="store_true", default=False)
parser.add_argument("--useDice", help="Use Dice Loss", action="store_true", default=False)
parser.add_argument("--noBall", help="Treat Ball as Background", action="store_true")
parser.add_argument("--noGoal", help="Treat Goal as Background", action="store_true")
parser.add_argument("--noRobot", help="Treat Robot as Background", action="store_true")
parser.add_argument("--noLine", help="Treat Lines as Background", action="store_true")
parser.add_argument("--topCam", help="Use Top Camera images only", action="store_true")
parser.add_argument("--bottomCam", help="Use Bottom Camera images only", action="store_true")
parser.add_argument("--lr", help="Learning rate", type=float, default=1e-3)
parser.add_argument("--decay", help="Weight decay", type=float, default=1e-5)
parser.add_argument("--transfer", help="Layers to truly train", action="store_true")
opt = parser.parse_args()
finetune = opt.finetune
learning_rate = opt.lr if opt.transfer else opt.lr
dec = opt.decay if finetune and not opt.transfer else opt.decay / 10
transfers = [1, 2, 3, 4] if opt.transfer else [0]
decays = [10*dec, 5*dec, 2*dec, dec] if (finetune and not opt.transfer) else [dec]
if opt.v2:
decays = [d*1 for d in decays]
noScale = opt.noScale
unet = opt.UNet
v2 = opt.v2
nb = opt.noBall
ng = opt.noGoal
nr = opt.noRobot
nl = opt.noLine
tc = opt.topCam
bc = opt.bottomCam
fineTuneStr = "Finetuned" if finetune else ""
scaleStr = "VGA" if noScale else ""
v2Str = "v2" if v2 else ""
unetStr = "UNet" if unet else ""
nbStr = "NoBall" if nb else ""
ngStr = "NoGoal" if ng else ""
nrStr = "NoRobot" if nr else ""
nlStr = "NoLine" if nl else ""
cameraString = "" if tc == bc else ("top" if tc else "bottom")
cameraSaveStr = cameraString if finetune else ""
if tc == bc:
cameraString = "both"
scale = 2 if noScale else 4
labSize = (480//scale, 640//scale)
weights_path = "checkpoints/best%s%s%s%s%s%s%s%s.weights" % (v2Str,scaleStr,unetStr,nbStr,ngStr,nrStr,nlStr,cameraSaveStr)
if nb and ng and nr and nl:
print("You need to have at least one non-background class!")
exit(-1)
if cameraString != "both" and not finetune:
print("You can only select camera images for the finetune dataset. Using both cameras by default")
cameraString = "both"
n_cpu = 8
channels = 3
epochs = 100 if noScale or not finetune else 200
momentum = 0.5
if finetune:
#learning_rate *= 0.1
momentum = 0.5
epochs = 200 if noScale else 200
outSize = 1.0/(labSize[0] * labSize[1])
os.makedirs("output", exist_ok=True)
os.makedirs("checkpoints", exist_ok=True)
cuda = torch.cuda.is_available()
Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
LongTensor = torch.cuda.LongTensor if cuda else torch.LongTensor
seed = 12345678
np.random.seed(seed)
random.seed(seed)
torch.manual_seed(seed)
if cuda:
torch.cuda.manual_seed(seed)
batchSize = 16 if finetune else (32 if noScale else 64)
root = "../../Data/RoboCup" if sys.platform != 'win32' else "D:/Datasets/RoboCup"
trainloader = data.DataLoader(SSYUVDataset(root,img_size=labSize,train=True,finetune=finetune,camera=cameraString),
batch_size=batchSize, shuffle=True, num_workers=8)
valloader = data.DataLoader(SSYUVDataset(root,img_size=labSize,train=False,finetune=finetune,camera=cameraString),
batch_size=batchSize, shuffle=True, num_workers=8)
numClass = 5 - nb - ng - nr - nl
numPlanes = 8 if v2 else 8
levels = 3 if unet else (1 if v2 else 2)
depth = 4 if unet else 4
bellySize = 0 if unet else (9 if v2 else 5)
classSize = 3 if v2 else 1
bellyPlanes = numPlanes*pow(2,depth-1) if v2 else numPlanes*pow(2,depth)
weights = Tensor([1, 2, 6, 3, 2]) if opt.useDice else Tensor([1, 10, 30, 10, 2])
if finetune:
weights = Tensor([1, 6, 2, 10, 4])
classIndices = torch.LongTensor([1, (not nb), (not nr), (not ng), (not nl)])
weights = weights[classIndices == 1]
criterion = DiceLoss(weights) if opt.useDice else CrossEntropyLoss2d(weights)
indices = None
mapLoc = None if cuda else {'cuda:0': 'cpu'}
for transfer in transfers:
if len(transfers) > 1:
print("######################################################")
print("############# Finetune with transfer: %d #############" % transfer)
print("######################################################")
for decay in decays:
if len(decays) > 1:
print("######################################################")
print("############ Finetune with decay: %.1E ############" % decay)
print("######################################################")
torch.random.manual_seed(12345678)
if cuda:
torch.cuda.manual_seed(12345678)
# Initiate model
model = ROBO_UNet(noScale,planes=numPlanes,depth=depth,levels=levels,bellySize=bellySize,bellyPlanes=bellyPlanes,pool=unet,v2=v2,classSize=classSize)
comp = model.get_computations()
print(comp)
print(sum(comp))
if finetune:
model.load_state_dict(torch.load(weights_path))
if cuda:
model = model.cuda()
bestLoss = 0
'''optimizer = torch.optim.SGD([
{'params': model.downPart[0:transfer].parameters(), 'lr': learning_rate*10},
{'params': model.downPart[transfer:].parameters()},
{'params': model.PB.parameters()},
{'params': model.upPart.parameters()},
{'params': model.segmenter.parameters()}
],lr=learning_rate,momentum=momentum)'''
optimizer = torch.optim.Adam([
{'params': model.downPart[0:transfer].parameters(), 'lr': learning_rate*10},
{'params': model.downPart[transfer:].parameters()},
{'params': model.PB.parameters()},
{'params': model.upPart.parameters()},
{'params': model.segmenter.parameters()}
],lr=learning_rate)
#optimizer = torch.optim.SGD(model.parameters(),lr=learning_rate,momentum=momentum)
eta_min = learning_rate/25 if opt.transfer else learning_rate/10
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer,epochs,eta_min=eta_min)
for epoch in range(epochs):
#if finetune:
train(epoch,epochs,100)
bestLoss = valid(epoch,epochs,bestLoss,False)
#else:
#bestLoss = train(epoch,epochs,bestLoss)
if finetune and (transfer == 0):
model.load_state_dict(torch.load("checkpoints/bestFinetune%s%s%s%s%s%s%s%s.weights" % (v2Str,scaleStr,unetStr,nbStr,ngStr,nrStr,nlStr,cameraSaveStr)))
with torch.no_grad():
indices = pruneModelNew(model.parameters())
#optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate/20)
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate/20)
print("Finetuning")
bestLoss = 0
for epoch in range(25):
train(epoch, 25, 100, indices=indices)
bestLoss = valid(epoch,25,bestLoss,True)