def testRRN_Pretrain(sample,pre):
RRN_net = MyRRN()
RRN_net.to(cfg.DEVICE)
RRN_net.load_state_dict(torch.load(pre))
sam = sample['image']
bbb = sample['bb']
tm = sample['tm']
gt = sample['gt']
bz = bbb.size(0)
num = bbb.size(1)
bbb = bbb.view(-1, 5)
ind = torch.arange(bz,requires_grad=False).view(-1,1)
ind = ind.repeat(1,num).view(-1,1)
bbb[:,0] = ind[:,0]
labels_output = resizeThermal(tm, bbb)
# labels_output = labels_output.to(cfg.DEVICE)
sam,bbb,tm = sam.to(cfg.DEVICE), bbb.to(cfg.DEVICE), tm.to(cfg.DEVICE)
out_RRN = RRN_net(sam,bbb)
# out_RRN = out_RRN.cpu().detach().numpy()
bbb = bbb.cpu().detach().numpy()
ther = convertTensor2Img(tm,norm=False)
imgg = visualizeRP(ther, bbb)
# print(imgg.dtype)
cv2.imshow('aa',imgg)
for ind,labels in enumerate(out_RRN[:4]):
# print('output')
p = convertTensor2Img(labels,norm=False)
cv2.imshow('rrn{}'.format(ind), p)
for ind,labels in enumerate(labels_output[:4]):
p = convertTensor2Img(labels,norm=False)
cv2.imshow('bbs{}'.format(ind), p)
cv2.waitKey(0)
cv2.destroyAllWindows()
def testResizeThermal(sample):
sam = sample['image']
bbb = sample['bb']
tm = sample['tm']
print(sample['img_info'])
gt = sample['gt']
bbb = bbb.cpu()
bz = bbb.size(0)
num = bbb.size(1)
bbb = bbb.view(-1,5)
gt = gt.cpu()
gt = gt.view(-1,5)
gt = gt.detach().numpy()
ind = torch.arange(bz).view(-1,1)
ind = ind.repeat(1,num).view(-1,1)
bbb[:,0] = ind[:,0]
labels_output = resizeThermal(tm, bbb)
print(labels_output.shape)
# labels_output = labels_output.type('torch.ByteTensor')
exit()
ther = convertTensor2Img(tm,norm=False)
bbb = bbb.detach().numpy()
imgg = visualizeRP(ther, bbb[:9])
cv2.imshow('winname', imgg)
for ind,labels in enumerate(labels_output[:9]):
p = convertTensor2Img(labels,norm=False)
cv2.imshow('bbs{}'.format(ind), p)
cv2.waitKey(0)
cv2.destroyAllWindows()
def train():
THERMAL_PATH = '/storageStudents/K2015/duyld/dungnm/dataset/KAIST/train/images_train_tm/'
ROOT_DIR = '/storageStudents/K2015/duyld/dungnm/dataset/KAIST/train/images_train'
IMGS_CSV = 'mydata/imgs_train.csv'
ROIS_CSV = 'mydata/rois_trainKaist_thr70_1.csv'
params = {'batch_size': 6, 'shuffle': True, 'num_workers': 12}
print(params)
max_epoch = 10
print('max_epoch', max_epoch)
LR = 1e-6 #learning rate
print('learning_rate', LR)
MT = 0.9 #momentum
device = torch.device("cuda:0")
# cudnn.benchmark = True
# transform = ToTensor()
full_transform = transforms.Compose(
[RandomHorizontalFlip(),
ToTensor(), my_normalize()])
# Normalize(rgb_mean,rgb_std)])
my_dataset = MyDataset(imgs_csv=IMGS_CSV,
rois_csv=ROIS_CSV,
root_dir=ROOT_DIR,
ther_path=THERMAL_PATH,
transform=full_transform)
dataloader = DataLoader(my_dataset, **params)
RRN_net = MyRRN()
RRN_net.to(device)
RRN_net.load_state_dict(
torch.load(
'models/model24/model24_lr_1e-6_bz_6_NBS_128_norm_epoch_9.pth'))
criterion = nn.MSELoss()
optimizer = optim.SGD(filter(lambda p: p.requires_grad,
RRN_net.parameters()),
lr=LR,
momentum=MT)
for epoch in range(max_epoch): # Lặp qua bộ dữ liệu huấn luyện nhiều lần
running_loss = 0.0
st = time.time()
for i, data in enumerate(dataloader):
# Lấy dữ liệu
sample = data
sam = sample['image']
bbb = sample['bb']
num = bbb.size(1)
bbb = bbb.view(-1, 5)
#reset id
# bbb[:, 0] = bbb[:, 0] - bbb[0, 0]
ind = torch.arange(params['batch_size'],
requires_grad=False).view(-1, 1)
ind = ind.repeat(1, num).view(-1, 1)
bbb[:, 0] = ind[:, 0]
tm = sample['tm']
# print(bbb.shape)
# print(tm.shape)
sam, bbb, tm = sam.to(device), bbb.to(device), tm.to(device)
# roi_pool = ROIPool((50, 50), 1/1)
# labels_output = roi_pool(tm,bbb)
labels_output = resizeThermal(tm, bbb)
labels_output = labels_output.to(device)
# Xoá giá trị đạo hàm
optimizer.zero_grad()
# Tính giá trị tiên đoán, đạo hàm, và dùng bộ tối ưu hoá để cập nhật trọng số.
out_RRN = RRN_net(sam, bbb)
loss = criterion(out_RRN, labels_output)
# print('loss at {}: '.format(i),loss.item())
loss.backward()
optimizer.step()
# In ra số liệu trong quá trình huấn luyện
running_loss += loss.item()
if i % 10 == 9: # In mỗi 2000 mini-batches.
text = '[{}, {}] loss: {:.3f} time: {:.3f}'.format(
epoch + 1, i + 1, running_loss / 10,
time.time() - st)
print(text)
with open('models/model26/log26.txt', 'a') as f:
f.write(text + '\n')
running_loss = 0.0
st = time.time()
torch.save(
RRN_net.state_dict(),
'models/model26/model26_lr_1e-6_bz_6_NBS_128_norm_epoch_{}.pth'.
format(epoch))
print('Huấn luyện xong')
def test():
TEST_THERMAL_PATH = '/storageStudents/K2015/duyld/dungnm/dataset/KAIST/test/images_test_tm/'
TEST_ROOT_DIR = '/storageStudents/K2015/duyld/dungnm/dataset/KAIST/test/images_test'
IMGS_CSV = 'mydata/imgs_test.csv'
ROIS_CSV = 'mydata/rois_testKaist_thr70_0.csv'
params = {'batch_size': 2, 'shuffle': True, 'num_workers': 24}
device = torch.device("cuda:0")
full_transform = transforms.Compose([
RandomHorizontalFlip(),
ToTensor(),
])
# Normalize(rgb_mean,rgb_std)])
my_dataset = MyDataset(imgs_csv=IMGS_CSV,
rois_csv=ROIS_CSV,
root_dir=TEST_ROOT_DIR,
ther_path=TEST_THERMAL_PATH,
transform=full_transform)
print(my_dataset.__len__())
dataloader = DataLoader(my_dataset, **params)
RRN_net = MyRRN()
RRN_net.to(device)
RRN_net.load_state_dict(
torch.load(
'models/model21/model21_lr_1e-7_bz_6_NBS_128_data_True_epoch_7.ptx'
))
st = time.time()
running_loss = 0.0
total_loss = []
criterion = nn.MSELoss()
for i, sample in enumerate(dataloader):
# Lấy dữ liệu
sam = sample['image']
bbb = sample['bb']
num = bbb.size(1)
bbb = bbb.view(-1, 5)
#reset id
# bbb[:, 0] = bbb[:, 0] - bbb[0, 0]
ind = torch.arange(params['batch_size'],
requires_grad=False).view(-1, 1)
ind = ind.repeat(1, num).view(-1, 1)
bbb[:, 0] = ind[:, 0]
tm = sample['tm']
# print(bbb.shape)
# print(tm.shape)
sam, bbb, tm = sam.to(device), bbb.to(device), tm.to(device)
# roi_pool = ROIPool((50, 50), 1/1)
# labels_output = roi_pool(tm,bbb)
labels_output = resizeThermal(tm, bbb)
labels_output = labels_output.to(device)
# print('label shape',labels_output.shape)
# Tính giá trị tiên đoán, đạo hàm, và dùng bộ tối ưu hoá để cập nhật trọng số.
out_RRN = RRN_net(sam, bbb)
loss = criterion(out_RRN, labels_output)
# In ra số liệu trong quá trình huấn luyện
running_loss += loss.item()
if i % 10 == 9: # In mỗi 9 mini-batches.
text = '[{}, {}] loss: {:.3f} time: {:.3f}'.format(
0 + 1, i + 1, running_loss / 10,
time.time() - st)
print(text)
with open('test2_model21_epoch7.txt', 'a') as f:
f.write(text + '\n')
total_loss.append(running_loss)
running_loss = 0.0
st = time.time()
print("TOTAL: ", sum(total_loss) / len(total_loss))
def train():
params = {
'batch_size': cfg.TRAIN.BATCH_SIZE,
'shuffle': cfg.TRAIN.SHUFFLE,
'num_workers': cfg.TRAIN.NUM_WORKERS
}
print(params)
max_epoch = cfg.TRAIN.MAX_EPOCH
print('max_epoch', max_epoch)
LR = cfg.TRAIN.LEARNING_RATE #learning rate
print('learning_rate', LR)
MT = cfg.TRAIN.MOMENTUM
# cudnn.benchmark = True
# transform = ToTensor()
full_transform = transforms.Compose([
RandomHorizontalFlip(),
ToTensor(),
Normalize(cfg.BGR_MEAN, cfg.BGR_STD)
])
my_dataset = MyDataset(imgs_csv=cfg.TRAIN.IMGS_CSV,
rois_csv=cfg.TRAIN.ROIS_CSV,
root_dir=cfg.TRAIN.ROOT_DIR,
ther_path=cfg.TRAIN.THERMAL_PATH,
transform=full_transform)
dataloader = DataLoader(my_dataset, **params)
RRN_net = MyRRN()
RRN_net.to(cfg.DEVICE)
RRN_net.load_state_dict(
torch.load(
'models/RRN/model24/model24_lr_1e-6_bz_6_NBS_128_norm_epoch_9.pth')
)
criterion = nn.MSELoss()
optimizer = optim.SGD(filter(lambda p: p.requires_grad,
RRN_net.parameters()),
lr=LR,
momentum=MT)
f = open('models/RRN/model27/log27.txt', 'a')
for epoch in range(max_epoch): # Lặp qua bộ dữ liệu huấn luyện nhiều lần
running_loss = 0.0
st = time.time()
for i, data in enumerate(dataloader):
# Lấy dữ liệu
sample = data
sam = sample['image']
bbb = sample['bb']
num = bbb.size(1)
bbb = bbb.view(-1, 5)
#reset id
# bbb[:, 0] = bbb[:, 0] - bbb[0, 0]
ind = torch.arange(params['batch_size'],
requires_grad=False).view(-1, 1)
ind = ind.repeat(1, num).view(-1, 1)
bbb[:, 0] = ind[:, 0]
tm = sample['tm']
sam, bbb, tm = sam.to(cfg.DEVICE), bbb.to(cfg.DEVICE), tm.to(
cfg.DEVICE)
# labels_output = roi_pool(tm,bbb)
labels_output = resizeThermal(tm, bbb)
labels_output = labels_output.to(cfg.DEVICE)
# Xoá giá trị đạo hàm
optimizer.zero_grad()
# Tính giá trị tiên đoán, đạo hàm, và dùng bộ tối ưu hoá để cập nhật trọng số.
out_RRN = RRN_net(sam, bbb)
loss = criterion(out_RRN, labels_output)
# print('loss at {}: '.format(i),loss.item())
loss.backward()
optimizer.step()
# In ra số liệu trong quá trình huấn luyện
running_loss += loss.item()
if i % 10 == 9: # In mỗi 2000 mini-batches.
text = '[{}, {}] loss: {:.3f} time: {:.3f}'.format(
epoch + 1, i + 1, running_loss / 10,
time.time() - st)
print(text)
f.write(text + '\n')
running_loss = 0.0
st = time.time()
torch.save(
RRN_net.state_dict(),
'models/RRN/model27/model27_lr_1e-6_bz_6_NBS_128_norm_epoch_{}.pth'
.format(epoch))
f.close()
print('Huấn luyện xong')