def main():
if not os.path.exists(opt.output):
os.makedirs(opt.output)
converter = utils.strLabelConverter(opt.alphabet)
collate = dataset.AlignCollate()
train_dataset = dataset.TextLineDataset(text_file=opt.train_list, transform=dataset.ResizeNormalize(100, 32), converter=converter)
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=opt.batchsize, shuffle=True,
num_workers=opt.num_workers, collate_fn=collate)
test_dataset = dataset.TextLineDataset(text_file=opt.train_list, transform=dataset.ResizeNormalize(100, 32), converter=converter)
test_loader = torch.utils.data.DataLoader(test_dataset, shuffle=False, batch_size=opt.batchsize,
num_workers=opt.num_workers, collate_fn=collate)
criterion = nn.CTCLoss()
import models.crnn as crnn
crnn = crnn.CRNN(opt.imgH, opt.nc, opt.num_classes, opt.nh)
crnn.apply(utils.weights_init)
if opt.pretrained != '':
print('loading pretrained model from %s' % opt.pretrained)
crnn.load_state_dict(torch.load(opt.pretrained), strict=False)
print(crnn)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
crnn = crnn.to(device)
criterion = criterion.to(device)
# setup optimizer
optimizer = optim.Adam(crnn.parameters(), lr=opt.lr)
for epoch in range(opt.num_epochs):
loss_avg = 0.0
i = 0
while i < len(train_loader):
time0 = time.time()
# 训练
train_iter = iter(train_loader)
cost = trainBatch(crnn, train_iter, criterion, optimizer, device) # 一个批次，一个批次训练
loss_avg += cost
i += 1
if i % opt.interval == 0:
print('[%d/%d][%d/%d] Loss: %f Time: %f s' %
(epoch, opt.num_epochs, i, len(train_loader), loss_avg,
time.time() - time0))
loss_avg = 0.0
if (epoch + 1) % opt.valinterval == 0:
val(crnn, test_loader, criterion, converter=converter, device=device, max_iter=100)
criterion = CTCLoss()
writer = SummaryWriter()
# custom weights initialization called on crnn
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
m.weight.data.normal_(0.0, 0.02)
elif classname.find('BatchNorm') != -1:
m.weight.data.normal_(1.0, 0.02)
m.bias.data.fill_(0)
crnn = crnn.CRNN(opt.imgH, num_channels, nclass, opt.hidden_size)
crnn.apply(weights_init)
if opt.pretrained != '':
print('loading pretrained model from %s' % opt.pretrained)
crnn.load_state_dict(torch.load(opt.pretrained))
print(crnn)
image = torch.FloatTensor(opt.batchSize, 3, opt.imgH, opt.imgH)
text = torch.IntTensor(opt.batchSize * 5)
length = torch.IntTensor(opt.batchSize)
if torch.cuda.is_available():
crnn = crnn.cuda(opt.gpu)
# crnn = torch.nn.DataParallel(crnn, device_ids=range(opt.ngpu))
image = image.cuda(opt.gpu)
criterion = criterion.cuda(opt.gpu)
converter = utils.strLabelConverter(params.alphabet)
#criterion = CTCLoss()
criterion = torch.nn.CTCLoss()
# cnn and rnn
image = torch.FloatTensor(params.batchSize, 3, params.imgH, params.imgH)
text = torch.IntTensor(params.batchSize * 5)
length = torch.IntTensor(params.batchSize)
crnn = crnn.CRNN(params.imgH, nc, nclass, params.nh)
if opt.cuda:
crnn.cuda()
image = image.cuda()
criterion = criterion.cuda()
crnn.apply(weights_init) #参数初始化
if params.crnn != '':
print('loading pretrained model from %s' % params.crnn)
crnn.load_state_dict(torch.load(params.crnn))
image = Variable(image)
text = Variable(text)
length = Variable(length)
# loss averager
loss_avg = utils.averager()
# setup optimizer
if params.adam:
optimizer = optim.Adam(crnn.parameters(),
lr=params.lr,
def main(arg):
print(arg)
train_dataset = dataset.lmdbDataset(
path=arg.train_root,
# transform=dataset.resizeNormalize((imgW,imgH)),
)
test_dataset = dataset.lmdbDataset(
path=arg.test_root,
# transform=dataset.resizeNormalize((arg.imgW,arg.imgH)),
)
d = test_dataset.__getitem__(0)
l = test_dataset.__len__()
train_loader = DataLoader(train_dataset,
num_workers=arg.num_workers,
batch_size=arg.batch_size,
collate_fn=dataset.alignCollate(
imgH=arg.imgH,
imgW=arg.imgW,
keep_ratio=arg.keep_ratio),
shuffle=True,
drop_last=True)
criterion = CTCLoss()
converter = utils.Converter(arg.num_class)
crnn = CRNN(imgH=arg.imgH, nc=3, nclass=arg.num_class + 1, nh=256)
# custom weights initialization called on crnn
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
m.weight.data.normal_(0.0, 0.02)
elif classname.find('BatchNorm') != -1:
m.weight.data.normal_(1.0, 0.02)
m.bias.data.fill_(0)
crnn.apply(weights_init)
print(crnn)
image = torch.FloatTensor(arg.batch_size, 3, arg.imgH, arg.imgW)
text = torch.IntTensor(arg.batch_size * 5)
length = torch.IntTensor(arg.batch_size)
image = Variable(image)
text = Variable(text)
length = Variable(length)
# loss averager
loss_avg = utils.averager()
# setup optimizer
if arg.opt == 'adam':
optimizer = optim.Adam(crnn.parameters(), 0.01, betas=(0.5, 0.999))
elif arg.opt == 'adadelta':
optimizer = optim.Adadelta(crnn.parameters())
else:
optimizer = optim.RMSprop(crnn.parameters(), 0.01)
for epoch in range(arg.n_epoch):
train_iter = iter(train_loader)
i = 0
while i < len(train_loader):
for p in crnn.parameters():
p.requires_grad = True
crnn.train()
data = train_iter.next()
cpu_images, cpu_texts = data
batch_size = cpu_images.size(0)
utils.loadData(image, cpu_images)
text_labels, l = converter.encode(cpu_texts)
utils.loadData(text, text_labels)
utils.loadData(length, l)
preds = crnn(image)
preds_size = Variable(torch.IntTensor([preds.size(0)] *
batch_size))
cost = criterion(preds, text, preds_size, length) / batch_size
crnn.zero_grad()
cost.backward()
optimizer.step()
loss_avg.add(cost)
i += 1
if i % arg.displayInterval == 0:
print(
'[%d/%d][%d/%d] Loss: %f' %
(epoch, arg.n_epoch, i, len(train_loader), loss_avg.val()))
loss_avg.reset()
if i % arg.testInterval == 0:
test(arg, crnn, test_dataset, criterion, image, text, length)
# do checkpointing
if i % arg.saveInterval == 0:
name = '{0}/netCRNN_{1}_{2}_{3}_{4}.pth'.format(
arg.model_dir, arg.num_class, arg.type, epoch, i)
torch.save(crnn.state_dict(), name)
print('model saved at ', name)
torch.save(
crnn.state_dict(),
'{0}/netCRNN_{1}_{2}.pth'.format(arg.model_dir, arg.num_class,
arg.type))
converter = utils.strLabelConverter(alphabet)
criterion = CTCLoss()
# custom weights initialization called on crnn
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
m.weight.data.normal_(0.0, 0.02)
elif classname.find('BatchNorm') != -1:
m.weight.data.normal_(1.0, 0.02)
m.bias.data.fill_(0)
crnn = crnn.CRNN(opt.imgH, nc, nclass, nh, ngpu)
crnn.apply(weights_init)
if opt.crnn != '':
print('loading pretrained model from %s' % opt.crnn)
crnn.load_state_dict(torch.load(opt.crnn))
print(crnn)
image = torch.FloatTensor(opt.batchSize, 3, opt.imgH, opt.imgH)
text = torch.IntTensor(opt.batchSize * 5)
length = torch.IntTensor(opt.batchSize)
if opt.cuda:
crnn.cuda()
image = image.cuda()
criterion = criterion.cuda()
image = Variable(image)
text = torch.IntTensor(params.batchSize *
5) # text length no less than batchsize * 5
length = torch.IntTensor(params.batchSize) # define each word length
crnn = crnn.CRNN(params.imgH, nc, nclass, params.nh)
#if opt.cuda:
# crnn.cuda()
# image = image.cuda()
# criterion = criterion.cuda()
crnn = crnn.to(device)
image = image.to(device)
criterion = criterion.to(device)
text = text.to(device)
#length = length.to(device)
crnn.apply(weights_init) # self-define weight initialize function
#print("crnn =",crnn)
#print("crnn parameters =",crnn.cnn)
if params.crnn != '':
print('loading pretrained model from %s' % params.crnn)
crnn.load_state_dict(torch.load(params.crnn))
#for para in crnn.parameters():
#print("parameters =", para)#para.requires_grad=False
#crnn.state_dict()#get parameters list-------->cnn.conv0.weight cnn.conv0.bias cnn.conv1.weight cnn.conv1.bias cnn.conv2.weight cnn.conv2.bias cnn.batchnorm2.weight cnn.batchnorm2.bias
#crnn.state_dict().items()#get parameters name and its value ------------>('rnn.1.embedding.bias', tensor([-0.0602, -0.3962, -0.3687, -0.3052, -0.2965, -0.3442, -0.4302, -0.3631,\
#-0.3303, -0.2937, -0.2485, -0.4897, -0.2815, -0.3473, -0.3228, -0.2575,\
#-0.3200, -0.3391, -0.4191, -0.2042, -0.5009, -0.4935, -0.3103, -0.2821, \
#-0.3521, -0.2895, -0.3934, -0.2745, -0.3072, -0.2851, -0.2467, -0.3485,\
#-0.2747, -0.2944, -0.3731, -0.4065, -0.3084, -0.3154, -0.4246],\
#device='cuda:0'))