-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_DIR_cls_ft_NUCLA.py
215 lines (189 loc) · 11.1 KB
/
train_DIR_cls_ft_NUCLA.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
# DIR-4: fine-tune CVAR cls with CL feature, N-UCLA dataset
import time
import argparse
from matplotlib import pyplot as plt
import torch
from torch.optim import lr_scheduler
from torch.utils.data import DataLoader
from dataset.crossView_UCLA_ske import os, np, random, NUCLA_CrossView
from modelZoo.BinaryCoding import nn, gridRing,contrastiveNet
from utils import load_fineTune_model
from test_cls_CV_DIR import testing, getPlots
random.seed(0)
np.random.seed(0)
torch.manual_seed(0)
def get_parser():
def str2bool(v):
if v.lower() in ('yes', 'true', 't', 'y', '1'): return True
elif v.lower() in ('no', 'false', 'f', 'n', '0'): return False
else: raise argparse.ArgumentTypeError('Unsupported value encountered.')
parser = argparse.ArgumentParser(description='CVARDIF')
parser.add_argument('--modelRoot', default='/data/Dan/202111_CVAR/NUCLA/CV_DIR_cls_ft',
help='the work folder for storing experiment results')
parser.add_argument('--path_list', default='', help='')
parser.add_argument('--pretrain', default='', help='')
parser.add_argument('--cus_n', default='', help='customized name')
parser.add_argument('--dataset', default='NUCLA', help='') # dataset = 'NUCLA'
parser.add_argument('--setup', default='setup1', help='') # setup = 'setup1' # v1,v2 train, v3 test;
parser.add_argument('--num_class', default=10, type=int, help='') # num_class = 10
parser.add_argument('--dataType', default='2D', help='') # dataType = '2D'
parser.add_argument('--sampling', default='Single', help='') # sampling = 'Multi' #sampling strategy
parser.add_argument('--nClip', default=1, type=int, help='') # sampling=='multi' or sampling!='Single'
parser.add_argument('--bs', default=8, type=int, help='')
parser.add_argument('--nw', default=8, type=int, help='')
parser.add_argument('--mode', default='dy+bi+cl', help='dy+bi+cl | dy+cl | rgb+dy') # mode = 'dy+bi+cl'
parser.add_argument('--RHdyan', default='1', type=str2bool, help='') # RHdyan = True
parser.add_argument('--withMask', default='0', type=str2bool, help='') # withMask = False
parser.add_argument('--maskType', default='None', help='') # maskType = 'score'
parser.add_argument('--contrastive', default='1', type=str2bool, help='') # constrastive = True
parser.add_argument('--finetune', default='1', type=str2bool, help='')
parser.add_argument('--fusion', default='0', type=str2bool, help='') # fusion = False
parser.add_argument('--groupLasso', default='0', type=str2bool, help='')
parser.add_argument('--T', default=36, type=int, help='') # T = 36 # input clip length
parser.add_argument('--N', default=80*2, type=int, help='') # N = 80*2
parser.add_argument('--lam_f', default=0.1, type=float) # fistaLam = 0.1
parser.add_argument('--gumbel_thresh', default=0.505, type=float) # gumbel_thresh = 0.505
parser.add_argument('--gpu_id', default=1, type=int, help='') # gpu_id = 5
parser.add_argument('--Epoch', default=50, type=int, help='') # Epoch = 100
parser.add_argument('--lr', default=1e-4, type=float, help='sparse coding') # lr = 1e-3 # classifier
parser.add_argument('--lr_2', default=5e-4, type=float, help='classifier') # lr_2 = 1e-4 # sparse codeing
parser.add_argument('--Alpha', default=0.1, type=float, help='bi loss') # Alpha = 0.1
parser.add_argument('--lam1', default=1, type=float, help='cls loss') # lam1 = 2
parser.add_argument('--lam2', default=0.5, type=float, help='mse loss') # lam2 = 1
return parser
def main(args):
args.saveModel = os.path.join(args.modelRoot,
f'NUCLA_CV_{args.setup}_{args.sampling}/DIR_cls_ft_{args.mode}/')
if not os.path.exists(args.saveModel): os.makedirs(args.saveModel)
print('mode:',args.mode, 'model path:', args.saveModel, 'gpu:', args.gpu_id)
'============================================= Main Body of script================================================='
P,Pall = gridRing(args.N)
Drr = abs(P)
Drr = torch.from_numpy(Drr).float()
Dtheta = np.angle(P)
Dtheta = torch.from_numpy(Dtheta).float()
# Dataset
assert args.path_list!='', '!!! NO Dataset Sample LIST !!!'
path_list = args.path_list + f"/data/CV/{args.setup}/"
# root_skeleton = '/data/Dan/N-UCLA_MA_3D/openpose_est'
trainSet = NUCLA_CrossView(root_list=path_list, phase='train',
setup=args.setup, dataType=args.dataType,
sampling=args.sampling, nClip=args.nClip,
T=args.T, maskType=args.maskType)
trainloader = DataLoader(trainSet, shuffle=True,
batch_size=args.bs, num_workers=args.nw)
testSet = NUCLA_CrossView(root_list=path_list, phase='test',
setup=args.setup, dataType=args.dataType,
sampling=args.sampling, nClip=args.nClip,
T=args.T, maskType=args.maskType)
testloader = DataLoader(testSet, shuffle=True,
batch_size=args.bs, num_workers=args.nw)
net = contrastiveNet(dim_embed=128, Npole=args.N+1,
Drr=Drr, Dtheta=Dtheta, fistaLam=args.lam_f,
mode=args.mode, Inference=True,
dataType=args.dataType, dim=2, nClip=args.nClip,
fineTune=args.finetune, useCL=args.contrastive,
gpu_id=args.gpu_id).cuda(args.gpu_id)
# 'load pre-trained contrastive model'
#pre_train = modelRoot + sampling + '/' + mode + '/T36_contrastive_fineTune_all/' + '40.pth'
assert args.pretrain!='', '!!! NO Pretrained Dictionary !!!'
print('pretrain:', args.pretrain)
state_dict = torch.load(args.pretrain, map_location=args.map_loc)
net = load_fineTune_model(state_dict, net)
net.train()
# frozen all layers except last classification layer
for p in net.backbone.sparseCoding.parameters(): p.requires_grad = False
for p in net.backbone.Classifier.parameters(): p.requires_grad = True
# for p in net.backbone.Classifier.parameters(): p.requires_grad = False
# for p in net.backbone.Classifier.cls[-1].parameters(): p.requires_grad = True
optimizer = torch.optim.SGD(
[{'params': filter(lambda x: x.requires_grad,net.backbone.sparseCoding.parameters()),
'lr':args.lr},
{'params': filter(lambda x: x.requires_grad, net.backbone.Classifier.parameters()),
'lr': args.lr_2}], weight_decay=1e-3, momentum=0.9)
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[50, 70], gamma=0.1)
Criterion = torch.nn.CrossEntropyLoss()
mseLoss = torch.nn.MSELoss()
L1loss = torch.nn.SmoothL1Loss()
LOSS = []
ACC = []
LOSS_CLS = []
LOSS_MSE = []
LOSS_BI = []
print('Experiment config | setup:',args.setup,'sampling:', args.sampling, 'gumbel_thresh:', args.gumbel_thresh,
'\n\tAlpha(bi):',args.Alpha,'lam1(cls):',args.lam1,'lam2(mse):',args.lam2,
'lr(mse):',args.lr,'lr_2(cls):',args.lr_2)
print('RHdyan:',args.RHdyan, 'useCL:', args.contrastive, 'fineTune:', args.finetune )
for epoch in range(1, args.Epoch+1):
print('start training epoch:', epoch)
lossVal = []
lossCls = []
lossBi = []
lossMSE = []
start_time = time.time()
for i, sample in enumerate(trainloader):
optimizer.zero_grad()
skeletons = sample['input_skeletons']['normSkeleton'].float().cuda(args.gpu_id)
# skeletons = sample['input_skeletons']['unNormSkeleton'].float().cuda(gpu_id)
# skeletons = sample['input_skeletons']['affineSkeletons'].float().cuda(gpu_id)
visibility = sample['input_skeletons']['visibility'].float().cuda(args.gpu_id)
gt_label = sample['action'].cuda(args.gpu_id)
if args.sampling == 'Single':
t = skeletons.shape[1]
input_skeletons = skeletons.reshape(skeletons.shape[0], t, -1) #bz, T, 25, 2
input_mask = visibility.reshape(visibility.shape[0], t, -1)
nClip = 1
else:
t = skeletons.shape[2]
input_skeletons = skeletons.reshape(skeletons.shape[0],skeletons.shape[1], t, -1) #bz,clip, T, 25, 2 --> bz*clip, T, 50
# input_mask = visibility.reshape(visibility.shape[0]*visibility.shape[1], t, -1)
nClip = skeletons.shape[1]
actPred, lastFeat, binaryCode, output_skeletons = net(input_skeletons, args.gumbel_thresh)
bi_gt = torch.zeros_like(binaryCode).cuda(args.gpu_id)
actPred = actPred.reshape(skeletons.shape[0], nClip, args.num_class)
actPred = torch.mean(actPred, 1)
# target_skeletons = skeletons.reshape(skeletons.shape[0]*skeletons.shape[1],t,-1)
# loss = args.lam1 * Criterion(actPred, gt_label) \
# + args.lam2 * mseLoss(output_skeletons, target_skeletons.squeeze(-1)) \
# + args.Alpha * L1loss(binaryCode, bi_gt)
# lossMSE.append(mseLoss(output_skeletons, target_skeletons.squeeze(-1)).data.item())
loss = args.lam1 * Criterion(actPred, gt_label) \
+ args.lam2 * mseLoss(output_skeletons, input_skeletons) \
+ args.Alpha * L1loss(binaryCode, bi_gt)
lossBi.append(L1loss(binaryCode, bi_gt).data.item())
lossMSE.append(mseLoss(output_skeletons, input_skeletons.squeeze(-1)).data.item())
loss.backward()
# print('rr.grad:', net.sparseCoding.rr.grad, 'mse:', lossMSE[-1])
optimizer.step()
lossVal.append(loss.data.item())
lossCls.append(Criterion(actPred, gt_label).data.item())
loss_val = np.mean(np.array(lossVal))
LOSS.append(loss_val)
LOSS_CLS.append(np.mean(np.array((lossCls))))
LOSS_MSE.append(np.mean(np.array(lossMSE)))
LOSS_BI.append(np.mean(np.array(lossBi)))
print('epoch:', epoch, '|loss:', loss_val, '|cls:', np.mean(np.array(lossCls)),
'|mse:', np.mean(np.array(lossMSE)), '|bi:', np.mean(np.array(lossBi)))
end_time = time.time()
print('training time(h):', (end_time - start_time) / 3600)
scheduler.step()
if epoch % 5 == 0:
torch.save({'state_dict': net.state_dict(),
'optimizer': optimizer.state_dict()},
args.saveModel +'dir_cls_ft'+ str(epoch) + '.pth')
Acc = testing(testloader, net, args.gpu_id,
args.sampling, args.mode,
args.withMask, args.gumbel_thresh, None)
print('testing epoch:', epoch, f'Acc: {Acc*100:.4f}%')
ACC.append(Acc)
if __name__ == "__main__":
parser = get_parser()
args=parser.parse_args()
args.map_loc = "cuda:"+str(args.gpu_id) # map_loc = "cuda:"+str(gpu_id)
if args.sampling == 'Single': args.nClip = 1
else: args.nClip = 6
main(args)
# 'plotting results:'
# getPlots(LOSS,LOSS_CLS, LOSS_MSE, LOSS_BI, ACC,fig_name='DY_CL.pdf')
torch.cuda.empty_cache()
print('done')