def test(args, logger):
# 返回Net.train()或Net.eval()
lprnet = build_lprnet(lpr_max_len=args.lpr_max_len, phase=args.phase_train, class_num=len(CHARS), dropout_rate=args.dropout_rate)
device = torch.device("cuda:0" if args.cuda else "cpu")
lprnet.to(device) # 实例化后使用.to方法将网络移动到GPU或CPU
print("Successful to build network!") # 到此位置模型搭建完成
# load pretrained model
if args.pretrained_model:
lprnet.load_state_dict(torch.load(args.pretrained_model, map_location=device))
print("load pretrained model successful!\n")
else:
print("[Error] Can't found pretrained mode, please check!")
return False
test_img_dirs = os.path.expanduser(args.test_img_dirs) # 把path中包含的"~"和"~user"转换成用户目录
test_dataset = LPRDataLoader(test_img_dirs.split(','), args.img_size, args.lpr_max_len) # lpr_max_len为车牌最大字符数
# epsilons = np.arange(0, 0.1, 0.005).tolist()
epsilons = np.arange(0, 0.1, 0.01).tolist()
for epsilon in epsilons:
args.epsilon = epsilon
acc = Greedy_Decode_Eval(lprnet, test_dataset, args)
logger.info("Epsilon: {}\tAccuracy: {:.4f}".format(args.epsilon, acc))
print("Epsilon: {}\tAccuracy: {:.4f}".format(args.epsilon, acc))
def test(args):
test_img_dirs = './data/my_test' # 这是裁剪出车牌后的路径
# 返回Net.train()或Net.eval()
lprnet = build_lprnet(lpr_max_len=args.lpr_max_len,
phase=args.phase_train,
class_num=len(CHARS),
dropout_rate=args.dropout_rate)
device = torch.device("cuda:0" if args.cuda else "cpu")
lprnet.to(device) # 实例化后使用.to方法将网络移动到GPU或CPU
print("Successful to build network!") # 到此位置模型搭建完成
# load pretrained model
if args.pretrained_model:
lprnet.load_state_dict(
torch.load(args.pretrained_model, map_location=device))
print("load pretrained model successful!\n")
else:
print("[Error] Can't found pretrained mode, please check!")
return False
test_img_dirs = os.path.expanduser(
test_img_dirs) # 把path中包含的"~"和"~user"转换成用户目录
test_dataset = LPRDataLoader(test_img_dirs.split(','), args.img_size,
args.lpr_max_len) # lpr_max_len为车牌最大字符数
try:
Greedy_Decode_Eval(lprnet, test_dataset, args)
finally:
cv2.destroyAllWindows()
def test():
args = get_parser()
lprnet = build_lprnet(lpr_max_len=args.lpr_max_len,
phase=args.phase_train,
class_num=len(CHARS),
dropout_rate=args.dropout_rate)
device = torch.device("cuda:0" if args.cuda else "cpu")
lprnet.to(device)
print("Successful to build network!")
# load pretrained model
if args.pretrained_model:
lprnet.load_state_dict(torch.load(args.pretrained_model))
print("load pretrained model successful!")
else:
print("[Error] Can't found pretrained mode, please check!")
return False
test_img_dirs = os.path.expanduser(args.test_img_dirs)
test_dataset = LPRDataLoader(test_img_dirs.split(','), args.img_size,
args.lpr_max_len)
try:
Greedy_Decode_Eval(lprnet, test_dataset, args)
finally:
cv2.destroyAllWindows()
def test():
args = get_parser()
# epsilons = [0, 0.005, 0.01, 0.015, 0.02, 0.025, 0.03, 0.035, 0.04] # 0.07时能降到10%左右
epsilons = np.arange(0, 0.1, 0.005).tolist()
accuracies = []
# 返回Net.train()或Net.eval()
lprnet = build_lprnet(lpr_max_len=args.lpr_max_len, phase=args.phase_train, class_num=len(CHARS), dropout_rate=args.dropout_rate)
device = torch.device("cuda:0" if args.cuda else "cpu")
lprnet.to(device) # 实例化后使用.to方法将网络移动到GPU或CPU
print("Successful to build network!") # 到此位置模型搭建完成
# load pretrained model
if args.pretrained_model:
lprnet.load_state_dict(torch.load(args.pretrained_model, map_location=device))
print("load pretrained model successful!\n")
else:
print("[Error] Can't found pretrained mode, please check!")
return False
test_img_dirs = os.path.expanduser(args.test_img_dirs) # 把path中包含的"~"和"~user"转换成用户目录
test_dataset = LPRDataLoader(test_img_dirs.split(','), args.img_size, args.lpr_max_len) # lpr_max_len为车牌最大字符数
epoch_size = len(test_dataset) // args.test_batch_size # 整除，多余的末尾就不会包括进来了
test_dataset = DataLoader(test_dataset, args.test_batch_size, shuffle=True, num_workers=args.num_workers, collate_fn=collate_fn)
# collate_fn：如何取样本的，我们可以定义自己的函数来准确地实现想要的功能
# shuffle：设置为True的时候，每个世代都会打乱数据集
examples = []
for epsilon in epsilons:
batch_iterator = iter(test_dataset)
perturbed_image = Greedy_Decode_Eval(lprnet, batch_iterator, args, epsilon, epoch_size)
b, g, r = cv2.split(perturbed_image)
perturbed_image = cv2.merge([r, g, b])
examples.append(perturbed_image)
draw_all_images(epsilons, examples) # 画图
def visualize_stn():
with torch.no_grad():
# Get a batch of training data
dataset = LPRDataLoader([args.img_dirs], args.img_size)
dataloader = DataLoader(dataset,
batch_size=1,
shuffle=False,
num_workers=2,
collate_fn=collate_fn)
#imgs, labels, lengths = next(iter(dataloader))
for imgs, labels, lengths in dataloader:
input_tensor = imgs.cpu()
transformed_input_tensor = STN(imgs.to(device)).cpu()
in_grid = convert_image(torchvision.utils.make_grid(input_tensor))
out_grid = convert_image(
torchvision.utils.make_grid(transformed_input_tensor))
# Plot the results side-by-side
f, axarr = plt.subplots(1, 2)
axarr[0].imshow(in_grid)
axarr[0].set_title('Dataset Images')
axarr[1].imshow(out_grid)
axarr[1].set_title('Transformed Images')
plt.show()
def test(args):
test_img_dirs = './data/my_test' # 这是裁剪出车牌后的路径
epsilons = [0, 0.005, 0.01, 0.015, 0.02, 0.025, 0.03, 0.035,
0.04] # 0.07时能降到10%左右
accuracies = []
# 返回Net.train()或Net.eval()
lprnet = build_lprnet(lpr_max_len=args.lpr_max_len,
phase=args.phase_train,
class_num=len(CHARS),
dropout_rate=args.dropout_rate)
device = torch.device("cuda:0" if args.cuda else "cpu")
lprnet.to(device) # 实例化后使用.to方法将网络移动到GPU或CPU
print("Successful to build network!") # 到此位置模型搭建完成
# load pretrained model
if args.pretrained_model:
lprnet.load_state_dict(
torch.load(args.pretrained_model, map_location=device))
print("load pretrained model successful!\n")
else:
print("[Error] Can't found pretrained mode, please check!")
return False
test_img_dirs = os.path.expanduser(
test_img_dirs) # 把path中包含的"~"和"~user"转换成用户目录
test_dataset = LPRDataLoader(test_img_dirs.split(','), args.img_size,
args.lpr_max_len) # lpr_max_len为车牌最大字符数
epoch_size = len(test_dataset) // args.test_batch_size # 整除，多余的末尾就不会包括进来了
test_dataset = DataLoader(test_dataset,
args.test_batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
# collate_fn：如何取样本的，我们可以定义自己的函数来准确地实现想要的功能
# shuffle：设置为True的时候，每个世代都会打乱数据集
try:
for epsilon in epsilons:
batch_iterator = iter(test_dataset)
accuracies.append(
Greedy_Decode_Eval(lprnet, batch_iterator, args, epsilon,
epoch_size))
finally:
cv2.destroyAllWindows()
# 画出epsilon变化趋势图
plt.figure(figsize=(5, 5))
plt.plot(epsilons, accuracies, "*-")
plt.yticks(np.arange(0, 1.1, step=0.1))
plt.xticks(np.arange(0, 0.045, step=0.005))
plt.title("Accuracy vs Epsilon")
plt.xlabel("Epsilon")
plt.ylabel("Accuracy")
plt.show()
def test():
args = get_parser()
lprnet = build_lprnet(lpr_max_len=args.lpr_max_len,
phase=args.phase_train,
class_num=len(CHARS),
dropout_rate=args.dropout_rate)
device = torch.device("cuda:0" if args.cuda else "cpu")
lprnet.to(device)
print("Successful to build network!")
## 搭建空间变换网络
STN = STNet()
STN.to(device)
STN.load_state_dict(
torch.load('STN/weights/STN_Model_LJK_CA_XZH.pth',
map_location=lambda storage, loc: storage))
STN.eval()
print("空间变换网络搭建完成")
# load pretrained model
if args.pretrained_model:
lprnet.load_state_dict(torch.load(args.pretrained_model))
print("load pretrained model successful!")
else:
print("[Error] Can't found pretrained mode, please check!")
return False
test_img_dirs = os.path.expanduser(args.test_img_dirs)
test_dataset = LPRDataLoader(test_img_dirs.split(','), args.img_size,
args.lpr_max_len)
try:
Greedy_Decode_Eval(lprnet, test_dataset, args, STN, device)
finally:
cv2.destroyAllWindows()
def train(args, logger, epsilon=0.04, alpha=0.5):
T_length = 18 # args.lpr_max_len
epoch = 0 + args.resume_epoch
loss_val = 0
cnt = 0
# lr = args.learning_rate # 学习率
if not os.path.exists(args.save_folder):
os.mkdir(args.save_folder)
lprnet = build_lprnet(lpr_max_len=args.lpr_max_len, phase=args.phase_train, class_num=len(CHARS), dropout_rate=args.dropout_rate)
device = torch.device("cuda:0" if args.cuda else "cpu")
lprnet.to(device)
print("Successful to build network!")
# load pretrained model
if args.pretrained_model:
lprnet.load_state_dict(torch.load(args.pretrained_model, map_location=device))
print("load pretrained model successful!") # 从模型net.train()和net.eval()得出的结果完全不一样
# test_img_dirs = os.path.expanduser(args.test_img_dirs) # 测试集
# test_dataset = LPRDataLoader(test_img_dirs.split(','), args.img_size, args.lpr_max_len)
# Greedy_Decode_Eval(lprnet, test_dataset, args)
else:
def xavier(param):
nn.init.xavier_uniform(param)
def weights_init(m):
for key in m.state_dict():
if key.split('.')[-1] == 'weight':
if 'conv' in key:
nn.init.kaiming_normal_(m.state_dict()[key], mode='fan_out')
if 'bn' in key:
m.state_dict()[key][...] = xavier(1)
elif key.split('.')[-1] == 'bias':
m.state_dict()[key][...] = 0.01
lprnet.backbone.apply(weights_init)
lprnet.container.apply(weights_init)
print("initial net weights successful!")
# define optimizer
# optimizer = optim.SGD(lprnet.parameters(), lr=args.learning_rate,
# momentum=args.momentum, weight_decay=args.weight_decay)
optimizer = optim.RMSprop(lprnet.parameters(), lr=args.learning_rate, alpha = 0.9, eps=1e-08,
momentum=args.momentum, weight_decay=args.weight_decay)
# os.path.expanduser把path中包含的"~"和"~user"转换成用户目录
train_img_dirs = os.path.expanduser(args.train_img_dirs) # 训练集
test_img_dirs = os.path.expanduser(args.test_img_dirs) # 测试集
train_dataset = LPRDataLoader(train_img_dirs.split(','), args.img_size, args.lpr_max_len)
test_dataset = LPRDataLoader(test_img_dirs.split(','), args.img_size, args.lpr_max_len)
epoch_size = len(train_dataset) // args.train_batch_size # 求得批数
max_iter = args.max_epoch * epoch_size # 共需要循环的次数
ctc_loss = nn.CTCLoss(blank=len(CHARS)-1, reduction='mean') # reduction: 'none' | 'mean' | 'sum'
if args.resume_epoch > 0:
start_iter = args.resume_epoch * epoch_size
else:
start_iter = 0
for iteration in range(start_iter, max_iter):
if iteration % epoch_size == 0: # 说明新的一个周期开始
# create batch iterator
batch_iterator = iter(DataLoader(train_dataset, args.train_batch_size, shuffle=True, num_workers=args.num_workers, collate_fn=collate_fn))
loss_val = 0
cnt = 0 # 用于统计加了几次loss
epoch += 1
if iteration !=0 and iteration % args.save_interval == 0:
torch.save(lprnet.state_dict(), args.save_folder + 'LPRNet_' + '_iteration_' + repr(iteration) + '.pth') # 经过一定的间隔后就保存状态
if (iteration + 1) % args.test_interval == 0: # 经过一定间隔就评估模型
Greedy_Decode_Eval(lprnet, test_dataset, args, logger)
# lprnet.train() # should be switch to train mode
start_time = time.time()
# load train data
images, labels, lengths = next(batch_iterator)
# labels = np.array([el.numpy() for el in labels]).T
# print(labels)
# update lr
lr = adjust_learning_rate(optimizer, epoch, args.learning_rate, args.lr_schedule)
if args.cuda:
images = Variable(images.cuda(), requires_grad=True)
labels = Variable(labels, requires_grad=False).cuda()
else:
images = Variable(images, requires_grad=True)
labels = Variable(labels, requires_grad=False)
lprnet.eval() # 先开测试模式
# forward
logits = lprnet(images)
log_probs = logits.permute(2, 0, 1) # for ctc loss: T x N x C
# print(labels.shape)
log_probs = log_probs.log_softmax(2).requires_grad_() # requires_grad_()相当于把requires_grad属性置为1
# log_probs = log_probs.detach().requires_grad_()
# print(log_probs.shape)
# get ctc parameters
input_lengths, target_lengths = sparse_tuple_for_ctc(T_length, lengths)
# backprop
loss1 = ctc_loss(log_probs, labels, input_lengths=input_lengths, target_lengths=target_lengths)
lprnet.zero_grad()
loss1.backward(retain_graph=True)
data_grad = images.grad.data
perturbed_images = fgsm_attack(images, epsilon, data_grad) # FGSM攻击
logits = lprnet(perturbed_images)
log_probs = logits.permute(2, 0, 1) # for ctc loss: T x N x C
log_probs = log_probs.log_softmax(2).requires_grad_() # requires_grad_()相当于把requires_grad属性置为1
loss2 = ctc_loss(log_probs, labels, input_lengths=input_lengths, target_lengths=target_lengths)
loss = alpha * loss1 + (1-alpha) * loss2 # 新的loss值
lprnet.train()
lprnet.zero_grad()
optimizer.zero_grad() # 梯度置0
loss.backward()
optimizer.step()
loss_val += loss.item() # 在输出的时候可以loss_val取平均
cnt += 1 # loss_val的次数
end_time = time.time()
if (iteration + 1) % 20 == 0:
msg = 'Epoch:' + repr(epoch) + ' || epochiter: ' + repr(iteration % epoch_size) + '/' + repr(epoch_size) \
+ '|| Totel iter ' + repr(iteration) + ' || Loss: %.4f || ' % (loss_val / cnt) + \
'Batch time: %.4f sec. ||' % (end_time - start_time) + 'LR: %.8f' % (lr)
print(msg)
logger.info(msg) # 存入日志
# print('Epoch:' + repr(epoch) + ' || epochiter: ' + repr(iteration % epoch_size) + '/' + repr(epoch_size)
# + '|| Totel iter ' + repr(iteration) + ' || Loss: %.4f||' % (loss_val / cnt) +
# 'Batch time: %.4f sec. ||' % (end_time - start_time) + 'LR: %.8f' % (lr))
# final test
print("Final test Accuracy:")
logger.info("Final test Accuracy:")
Greedy_Decode_Eval(lprnet, test_dataset, args, logger)
# save final parameters
torch.save(lprnet.state_dict(), args.save_folder + 'Ans_LPRNet_model.pth')
def train():
args = get_parser()
T_length = 18 # args.lpr_max_len
epoch = 0 + args.resume_epoch
loss_val = 0
if not os.path.exists(args.save_folder):
os.mkdir(args.save_folder)
lprnet = build_lprnet(lpr_max_len=args.lpr_max_len, phase=args.phase_train, class_num=len(CHARS), dropout_rate=args.dropout_rate)
device = torch.device("cuda:0" if args.cuda else "cpu")
lprnet.to(device)
print("Successful to build network!")
# load pretrained model
if args.pretrained_model:
lprnet.load_state_dict(torch.load(args.pretrained_model))
print("load pretrained model successful!")
else:
def xavier(param):
nn.init.xavier_uniform(param)
def weights_init(m):
for key in m.state_dict():
if key.split('.')[-1] == 'weight':
if 'conv' in key:
nn.init.kaiming_normal_(m.state_dict()[key], mode='fan_out')
if 'bn' in key:
m.state_dict()[key][...] = xavier(1)
elif key.split('.')[-1] == 'bias':
m.state_dict()[key][...] = 0
lprnet.backbone.apply(weights_init)
lprnet.container.apply(weights_init)
print("initial net weights successful!")
# define optimizer
# optimizer = optim.SGD(lprnet.parameters(), lr=args.learning_rate,
# momentum=args.momentum, weight_decay=args.weight_decay)
optimizer = optim.RMSprop(lprnet.parameters(), lr=args.learning_rate, alpha = 0.9, eps=1e-08,
momentum=args.momentum, weight_decay=args.weight_decay)
train_img_dirs = os.path.expanduser(args.train_img_dirs)
test_img_dirs = os.path.expanduser(args.test_img_dirs)
train_dataset = LPRDataLoader(train_img_dirs.split(','), args.img_size, args.lpr_max_len)
test_dataset = LPRDataLoader(test_img_dirs.split(','), args.img_size, args.lpr_max_len)
epoch_size = len(train_dataset) // args.train_batch_size
max_iter = args.max_epoch * epoch_size
ctc_loss = nn.CTCLoss(blank=len(CHARS)-1, reduction='mean') # reduction: 'none' | 'mean' | 'sum'
if args.resume_epoch > 0:
start_iter = args.resume_epoch * epoch_size
else:
start_iter = 0
for iteration in range(start_iter, max_iter):
if iteration % epoch_size == 0:
# create batch iterator
batch_iterator = iter(DataLoader(train_dataset, args.train_batch_size, shuffle=True, num_workers=args.num_workers, collate_fn=collate_fn))
loss_val = 0
epoch += 1
if iteration !=0 and iteration % args.save_interval == 0:
torch.save(lprnet.state_dict(), args.save_folder + 'LPRNet_' + '_iteration_' + repr(iteration) + '.pth')
if (iteration + 1) % args.test_interval == 0:
Greedy_Decode_Eval(lprnet, test_dataset, args)
lprnet.train() # should be switch to train mode
start_time = time.time()
# load train data
images, labels, lengths = next(batch_iterator)
# labels = np.array([el.numpy() for el in labels]).T
# print(labels)
# get ctc parameters
input_lengths, target_lengths = sparse_tuple_for_ctc(T_length, lengths)
# update lr
lr = adjust_learning_rate(optimizer, epoch, args.learning_rate, args.lr_schedule)
if args.cuda:
images = Variable(images.cuda())
labels = Variable(labels.cuda(), requires_grad=False)
else:
images = Variable(images)
labels = Variable(labels, requires_grad=False)
# forward
logits = lprnet(images)
log_probs = logits.permute(2, 0, 1) # for ctc loss: T x N x C
# print(labels.shape)
log_probs = log_probs.log_softmax(2).requires_grad_()
# log_probs = log_probs.detach().requires_grad_()
# print(log_probs.shape)
# backprop
optimizer.zero_grad()
loss = ctc_loss(log_probs, labels, input_lengths=input_lengths, target_lengths=target_lengths)
if loss.item() == np.inf:
continue
loss.backward()
optimizer.step()
loss_val += loss.item()
end_time = time.time()
if iteration % 20 == 0:
print('Epoch:' + repr(epoch) + ' || epochiter: ' + repr(iteration % epoch_size) + '/' + repr(epoch_size)
+ '|| Totel iter ' + repr(iteration) + ' || Loss: %.4f||' % (loss.item()) +
'Batch time: %.4f sec. ||' % (end_time - start_time) + 'LR: %.8f' % (lr))
# final test
print("Final test Accuracy:")
Greedy_Decode_Eval(lprnet, test_dataset, args)
# save final parameters
torch.save(lprnet.state_dict(), args.save_folder + 'Final_LPRNet_model.pth')
lprnet = LPRNet(class_num=len(CHARS), dropout_rate=args.dropout_rate)
lprnet.to(device)
lprnet.load_state_dict(torch.load('weights/Final_LPRNet_model.pth', map_location=lambda storage, loc: storage))
# checkpoint = torch.load('saving_ckpt/lprnet_Iter_023400_model.ckpt')
# lprnet.load_state_dict(checkpoint['net_state_dict'])
lprnet.eval()
print("LPRNet loaded")
# torch.save(lprnet.state_dict(), 'weights/Final_LPRNet_model.pth')
STN = STNet()
STN.to(device)
STN.load_state_dict(torch.load('weights/Final_STN_model.pth', map_location=lambda storage, loc: storage))
# checkpoint = torch.load('saving_ckpt/stn_Iter_023400_model.ckpt')
# STN.load_state_dict(checkpoint['net_state_dict'])
STN.eval()
print("STN loaded")
# torch.save(STN.state_dict(), 'weights/Final_STN_model.pth')
dataset = LPRDataLoader([args.img_dirs], args.img_size)
dataloader = DataLoader(dataset, batch_size=args.batch_size, shuffle=False, num_workers=2, collate_fn=collate_fn)
print('dataset loaded with length : {}'.format(len(dataset)))
ACC = eval(lprnet, STN, dataloader, dataset, device)
print('the accuracy is {:.2f} %'.format(ACC*100))
visualize_stn()
#lprnet.load_state_dict(torch.load('weights/Final_LPRNet_model.pth', map_location=lambda storage, loc: storage))
print("LPRNet loaded")
print('LPRNet params: ', sum(p.numel() for p in lprnet.parameters()))
STN = STNet(r=r, resw=resw, resh=resh)
STN.to(device)
#summary(STN, (3,24,94))
summary(STN, (3, resh, resw))
#STN_load=torch.load('weights/Final_LPRNet_model.pth', map_location=lambda storage, loc: storage.cuda)
#STN.load_state_dict(torch.load('weights/Final_STN_model.pth', map_location=lambda storage, loc: storage))
print("STN loaded")
print('STN params: ', sum(p.numel() for p in STN.parameters()))
# raise NameError('stop')
dataset = {
'train':
LPRDataLoader([args.img_dirs_train], args.img_size,
aug_transform=True),
'val':
LPRDataLoader([args.img_dirs_val], args.img_size, aug_transform=False)
} ###shuffle
dataloader = {
'train':
DataLoader(dataset['train'],
batch_size=args.batch_size,
shuffle=False,
num_workers=4,
collate_fn=collate_fn),
'val':
DataLoader(dataset['val'],
batch_size=args.batch_size,
shuffle=False,
num_workers=4,
elif key.split('.')[-1] == 'bias':
m.state_dict()[key][...] = 0.01
lprnet.backbone.apply(weights_init)
lprnet.container.apply(weights_init)
print("initial net weights successful!")
STN = STNet()
STN.to(device)
STN.load_state_dict(
torch.load('weights/STN_model_Init.pth',
map_location=lambda storage, loc: storage))
print("STN loaded")
dataset = {
'train': LPRDataLoader([args.img_dirs_train], args.img_size),
'val': LPRDataLoader([args.img_dirs_val], args.img_size)
}
dataloader = {
'train':
DataLoader(dataset['train'],
batch_size=args.batch_size,
shuffle=False,
num_workers=4,
collate_fn=collate_fn),
'val':
DataLoader(dataset['val'],
batch_size=args.batch_size,
shuffle=False,
num_workers=4,
collate_fn=collate_fn)