def val(net, dataset, criterion, max_iter=100):
print('Start val')
for p in crnn.parameters():
p.requires_grad = False
net.eval()
data_loader = torch.utils.data.DataLoader(
dataset, shuffle=True, batch_size=opt.batchSize, num_workers=int(opt.workers))
val_iter = iter(data_loader)
i = 0
n_correct = 0
loss_avg = utils.averager()
for i in range(min(max_iter, len(data_loader))):
data = val_iter.next()
i += 1
cpu_images, cpu_texts = data
batch_size = cpu_images.size(0)
utils.loadData(image, cpu_images)
t, l = converter.encode(cpu_texts)
utils.loadData(text, t)
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
loss_avg.add(cost)
_, preds = preds.max(2)
preds = preds.squeeze(2)
preds = preds.transpose(1, 0).contiguous().view(-1)
sim_preds = converter.decode(preds.data, preds_size.data, raw=False)
for pred, target in zip(sim_preds, cpu_texts):
if pred == target.lower():
n_correct += 1
raw_preds = converter.decode(preds.data, preds_size.data, raw=True)
for raw_pred, pred, gt in zip(raw_preds, sim_preds, cpu_texts):
print('%-20s => %-20s, gt: %-20s' % (raw_pred, pred, gt))
accuracy = n_correct / float(max_iter * opt.batchSize)
print('Test loss: %f, accuray: %f' % (loss_avg.val(), accuracy))
def val(net, dataset, criterion, max_iter=100):
print('Start val')
for p in crnn.parameters():
p.requires_grad = False
net.eval()
data_loader = torch.utils.data.DataLoader(dataset,
shuffle=True,
batch_size=opt.batchSize,
num_workers=int(opt.workers))
val_iter = iter(data_loader)
i = 0
n_correct = 0
loss_avg = utils.averager()
max_iter = min(max_iter, len(data_loader))
for i in range(max_iter):
try:
data = val_iter.next()
except:
continue
i += 1
try:
cpu_images, cpu_texts = data
except ValueError:
cpu_images, cpu_texts, _ = data
batch_size = cpu_images.size(0)
utils.loadData(image, cpu_images)
t, l = converter.encode(cpu_texts)
global text
utils.loadData(text, t)
utils.loadData(length, l)
text = text.view((batch_size, -1)).cuda()
preds = crnn(image)
preds_size = Variable(torch.IntTensor([preds.size(0)] * batch_size))
cost = criterion(F.log_softmax(preds, dim=-1), text, preds_size,
length).sum() / batch_size
loss_avg.add(cost)
_, preds = preds.max(2)
#preds = preds.squeeze(2)
preds = preds.transpose(1, 0).contiguous().view(-1)
sim_preds = converter.decode(preds.data, preds_size.data, raw=False)
for pred, target in zip(sim_preds, cpu_texts):
if pred == target:
n_correct += 1
elif val_wrong is not None:
val_wrong.append(target + '_' + pred)
writer.add_scalars(
'Test', {
'loss': cost.item(),
'acc': (np.array(sim_preds) == np.array(cpu_texts)).mean()
},
global_step=global_step * max_iter + i)
raw_preds = converter.decode(preds.data, preds_size.data,
raw=True)[:opt.n_test_disp]
for raw_pred, pred, gt in zip(raw_preds, sim_preds, cpu_texts):
print('%-20s => %-20s, gt: %-20s' % (raw_pred, pred, gt))
accuracy = n_correct / float(max_iter * opt.batchSize)
print('Test loss: %f, accuray: %f' % (loss_avg.val(), accuracy))
crnn = crnn.CRNN(params.imgH, nc, nclass, params.nh)
if args.cuda:
crnn.cuda()
image = image.cuda()
criterion = criterion.cuda()
crnn.apply(weights_init)
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,
betas=(params.beta1, 0.999))
elif params.adadelta:
optimizer = optim.Adadelta(crnn.parameters(), lr=params.lr)
else:
optimizer = optim.RMSprop(crnn.parameters(), lr=params.lr)
training(crnn,train_loader,criterion,optimizer)
np.save('loss.npy', training_Loss)
np.save('Accuracy.npy', testing_Rate)
def val(crnn, val_data_list_param, criterion, max_iter_num=100):
# print('开始校验准确性')
for p in crnn.parameters():
p.requires_grad = False
crnn.eval()
i = 0
all_Count = 0
correct_Count = 0
while i < len(val_data_list_param):
val_data = val_data_list_param[i]
# datasetOne = datasetList[i]
# data_loader = torch.utils.data.DataLoader(
# datasetOne, shuffle=True, batch_size=opt.batchSize, num_workers=int(opt.workers))
data_set = val_data['dataset']
data_loader = torch.utils.data.DataLoader(data_set,
shuffle=True,
batch_size=opt.batchSize,
num_workers=1)
i += 1
# print("验证进度:{}/{},当前Flag:{}".format(i, len(val_data_list_param), val_data['dir']))
val_iter = iter(data_loader)
one_index = 0
one_correct = 0
loss_avg = utils.averager()
# 检测所用的图片数量
max_iter = min(max_iter_num, len(data_loader))
# 检测的总数量增加
all_Count += max_iter * opt.batchSize
for one_index in range(max_iter):
data = val_iter.next()
one_index += 1
cpu_images, cpu_texts = data
batch_size = cpu_images.size(0)
utils.loadData(image, cpu_images)
t, l = converter.encode(cpu_texts)
utils.loadData(text, t)
utils.loadData(length, l)
preds = crnn(image)
preds_size = Variable(torch.IntTensor([preds.size(0)] *
batch_size))
# cost = criterion(preds, text, preds_size, length)
cost = criterion(preds, text, preds_size, length) / batch_size
loss_avg.add(cost)
_, preds = preds.max(2, keepdim=True)
preds = preds.squeeze(2)
preds = preds.transpose(1, 0).contiguous().view(-1)
sim_preds = converter.decode(preds.data,
preds_size.data,
raw=False)
for pred, target in zip(sim_preds, cpu_texts):
if pred == target.lower():
# 两个成功数量都加1
one_correct += 1
correct_Count += 1
del val_iter
raw_preds = converter.decode(preds.data, preds_size.data,
raw=True)[:opt.n_test_disp]
accuracy = one_correct / float(max_iter * opt.batchSize)
if accuracy < 0.95:
for raw_pred, pred, gt in zip(raw_preds, sim_preds, cpu_texts):
print_msg('%-20s => %-20s, gt: %-20s' % (raw_pred, pred, gt))
print_msg('验证 %-3d/%d,Loss: %f,Flag: [%-15s] 的成功率: %f' %
(i, len(val_data_list_param), loss_avg.val(),
val_data['dir'], accuracy))
accuracy = correct_Count / float(all_Count)
print_msg('总的成功率: %f ,总验证文件数: %d ' % (accuracy, all_Count))
return accuracy
def val(net, test_dataset, criterion, max_iter=100):
print('Start val')
for p in crnn.parameters():
p.requires_grad = False
# layer_dict = net.state_dict()
# print(layer_dict['cnn.conv1.weight'])
net.eval()
data_loader = torch.utils.data.DataLoader(test_dataset,
shuffle=True,
batch_size=opt.batchSize,
num_workers=int(opt.workers),
collate_fn=dataset.alignCollate(
imgH=32,
imgW=100,
keep_ratio=True))
val_iter = iter(data_loader)
i = 0
n = 0
n_correct = 0
n_text = 0
loss_avg = util.averager()
max_iter = len(data_loader)
for i in range(max_iter):
data = val_iter.next()
i += 1
cpu_images, cpu_texts = data
batch_size = cpu_images.size(0)
util.loadData(image, cpu_images)
t, l = converter.encode(cpu_texts)
util.loadData(text, t)
util.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
loss_avg.add(cost)
_, preds = preds.max(2)
# preds = preds.squeeze(2)
preds = preds.transpose(1, 0).contiguous().view(-1)
# print (preds)
sim_preds = converter.decode(preds.data, preds_size.data, raw=False)
for pred, target in zip(sim_preds, cpu_texts):
if isinstance(target, unicode) is False:
target = target.decode('utf-8')
pred_encode, _ = converter.encode(pred)
target_encode, _ = converter.encode(target)
t = editdistance.eval(pred_encode, target_encode)
l = len(target_encode)
n_correct += t
n_text += l
n += 1
raw_preds = converter.decode(preds.data, preds_size.data,
raw=True)[:opt.n_test_disp]
for raw_pred, sim_pred, gt in zip(raw_preds, sim_preds, cpu_texts):
print('%-20s => %-20s, gt: %-20s' % (raw_pred, sim_pred, gt))
len_edit = n_correct / float(n)
len_text = n_text / float(n)
norm = 1 - len_edit / len_text
print('aver editd: %f, norm acc: %f' % (len_edit, norm))
if opt.cuda:
crnn.cuda()
image = image.cuda()
criterion = criterion.cuda()
image = Variable(image)
text = Variable(text)
length = Variable(length)
# loss averager
loss_avg = utils.averager()
# setup optimizer
if opt.adam:
optimizer = optim.Adam(
crnn.parameters(), lr=opt.lr, betas=(opt.beta1, 0.999))
elif opt.adadelta:
optimizer = optim.Adadelta(crnn.parameters(), lr=opt.lr)
else:
optimizer = optim.RMSprop(crnn.parameters(), lr=opt.lr)
def val(net, dataset, criterion, max_iter=2):
print('Start val')
for p in crnn.parameters():
p.requires_grad = False
net.eval()
data_loader = torch.utils.data.DataLoader(
dataset,
def val(net, dataset, criterion, max_iter=20):
print('Start val')
for p in crnn.parameters():
p.requires_grad = False
net.eval()
data_loader = torch.utils.data.DataLoader(dataset,
shuffle=True,
batch_size=opt.batchSize,
num_workers=int(opt.workers))
val_iter = iter(data_loader)
i = 0
n_correct = 0
sum_NED = 0
n_dataset = 0
loss_avg = utils.averager()
max_iter = min(max_iter, len(data_loader))
for i in range(max_iter):
data = val_iter.next()
i += 1
cpu_images, cpu_texts = data
batch_size = cpu_images.size(0)
utils.loadData(image, cpu_images)
decode_texts = [text[2:-1] for text in cpu_texts]
t, l = converter.encode(decode_texts)
utils.loadData(text, t)
utils.loadData(length, l)
preds = crnn(image)
preds_chunks = preds.chunk(len(gpu_ids), dim=0)
preds = torch.cat(
preds_chunks, dim=1
) # [num_gpu * time_step, batch_size / num_gpu, nOut] -> [time_step, batch_size, nOut]
preds_size = Variable(torch.IntTensor([preds.size(0)] * batch_size))
cost = criterion(preds, text, preds_size, length) / batch_size
loss_avg.add(cost)
# [max_values,...,], [max_index,...,] = tensor.max(dim=k)
# preds = [time_step, batch_size, nOut]
# _, preds = preds.max(2)
# preds = [time_step, batch_size]
_, preds = preds.max(2)
# preds = preds.squeeze(2)
# after transpose, preds = [batch_size, time_step]
preds = preds.transpose(1, 0).contiguous().view(-1)
# preds = [batch_size * time_step]
sim_preds = converter.decode(preds.data, preds_size.data, raw=False)
# len(sim_preds) = batch_size, len(data_loader) = sizeof(datasets) / batch_size
for pred, target in zip(sim_preds, cpu_texts):
n_dataset += 1
gt = target[2:-1].lower()
if pred == gt:
n_correct += 1
else:
# sum of normalized edit distance
sum_NED += (EditDistance(pred, gt) / len(gt))
raw_preds = converter.decode(preds.data, preds_size.data,
raw=True)[:opt.n_test_disp]
for raw_pred, pred, gt in zip(raw_preds, sim_preds, cpu_texts):
# print('%-20s => %-20s, gt: %-20s' % (raw_pred, pred, gt))
print('%-20s => %-20s' % (raw_pred, pred))
gt = gt[2:-1].lower()
prompt = ""
ned = 0
if pred != gt:
maxLen = max(len(pred), len(gt))
minLen = min(len(pred), len(gt))
for i in range(maxLen):
if i < minLen:
if pred[i] != gt[i]:
prompt += "^"
else:
prompt += " "
else:
prompt += "^"
ned = EditDistance(pred, gt)
print("pred:%-20s\ngt :%-20s\nerr :%-20s" % (pred, gt, prompt))
print("Edit Distance:\t%f" % (ned))
# accuracy = n_correct / float(max_iter * opt.batchSize)
accuracy = n_correct / float(n_dataset)
print('Test loss: %f, accuray: %f, NED: %f' %
(loss_avg.val(), accuracy, sum_NED))
return accuracy, sum_NED
# 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(parameters.max_image_height, 3, parameters.number_of_classes,
256)
crnn.apply(weights_init)
criterion = CTCLoss(zero_infinity=True)
optimizer = optim.Adam(crnn.parameters(), lr=0.001, betas=(0.5, 0.9))
loss_avg = averager()
image = torch.FloatTensor(dataloader_params['batch_size'], 3,
parameters.max_image_width,
parameters.max_image_height)
text = torch.IntTensor(dataloader_params['batch_size'] * 5)
length = torch.IntTensor(dataloader_params['batch_size'])
if torch.cuda.is_available():
crnn.cuda()
crnn = torch.nn.DataParallel(crnn, device_ids=range(1))
image = image.cuda()
text = text.cuda()
length = length.cuda()
criterion = criterion.cuda()
def val(net, dataset, max_iter=100):
print('Start val')
for p in crnn.parameters():
p.requires_grad = False
net.eval()
data_loader = torch.utils.data.DataLoader(dataset,
shuffle=True,
batch_size=opt.batchSize,
num_workers=int(opt.workers))
val_iter = iter(data_loader)
i = 0
n_correct = 0
# loss averager
avg_h_val = utils.averager()
avg_cost_val = utils.averager()
avg_h_cost_val = utils.averager()
if opt.eval_all:
max_iter = len(data_loader)
else:
max_iter = min(max_iter, len(data_loader))
for i in range(max_iter):
data = val_iter.next()
# print('data: ', data)
# print(data)
i += 1
cpu_images, cpu_texts = data
batch_size = cpu_images.size(0)
utils.loadData(image, cpu_images)
t, l = converter.encode(cpu_texts)
utils.loadData(text, t)
utils.loadData(length, l)
# print('len(cpu_images): ', len(cpu_images))
# print('cpu_texts ', cpu_texts)
# print('len(cpu_texts): ', len(cpu_texts))
# print('l ', l)
# print(len(l))
# print('length ', length)
preds = crnn(image) # size = 26, 64, 96
# print(preds.size())
preds_size = Variable(torch.IntTensor([preds.size(0)] * batch_size))
H, cost = seg_ctc_ent_cost(preds,
text,
preds_size,
length,
uni_rate=opt.uni_rate)
h_cost = (1 - opt.h_rate) * cost - opt.h_rate * H
avg_h_val.add(H / batch_size)
avg_cost_val.add(cost / batch_size)
avg_h_cost_val.add(h_cost / batch_size)
_, preds = preds.max(2) # size = 26, 64
# print(preds.size())
# preds = preds.squeeze(2)
preds = preds.transpose(1, 0).contiguous().view(-1)
sim_preds = converter.decode(preds.data, preds_size.data, raw=False)
for idx, (pred, target) in enumerate(zip(sim_preds, cpu_texts)):
if pred == target.lower():
n_correct += 1
raw_preds = converter.decode(preds.data, preds_size.data,
raw=True)[:opt.n_test_disp]
for raw_pred, pred, gt in zip(raw_preds, sim_preds, cpu_texts):
print('%-30s => %-30s, gt: %-30s' % (raw_pred, pred, gt))
accuracy = n_correct / float(max_iter * opt.batchSize)
print(
'Test H: %f, Cost: %f, H Cost: %f, accuray: %f' %
(avg_h_val.val(), avg_cost_val.val(), avg_h_cost_val.val(), accuracy))
def val(net, dataset, criterion, max_iter=100):
print('Start val')
for p in crnn.parameters():
p.requires_grad = False
net.eval()
data_loader = torch.utils.data.DataLoader(dataset,
shuffle=True,
batch_size=opt.batchSize,
num_workers=int(opt.workers))
val_iter = iter(data_loader)
i = 0
n_correct = 0
loss_avg = utils.averager()
max_iter = min(max_iter, len(data_loader))
print("max_iter", max_iter, "len(data_loader)", len(data_loader))
for i in range(max_iter):
data = val_iter.next()
# print(data)
i += 1
cpu_images, cpu_texts = data
batch_size = cpu_images.size(0)
utils.loadData(image, cpu_images)
t, l = converter.encode(cpu_texts)
utils.loadData(text, t)
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
loss_avg.add(cost)
_, preds = preds.max(2)
# preds = preds.squeeze(2)
preds = preds.transpose(1, 0).contiguous().view(-1)
list_cpu_texts = []
for i in cpu_texts:
list_cpu_texts.append(i.decode('utf-8', 'strict'))
sim_preds = converter.decode(preds.data, preds_size.data, raw=False)
if (i == 1):
print(sim_preds)
print(cpu_texts)
# cpu_texts = byte_to_zh(cpu_texts)
# print("sim_preds",sim_preds)
for pred, target in zip(sim_preds, list_cpu_texts):
if (pred == target.lower()) | (pred == target):
n_correct += 1
raw_preds = converter.decode(preds.data, preds_size.data,
raw=True)[:opt.n_test_disp]
for raw_pred, pred, gt in zip(raw_preds, sim_preds, list_cpu_texts):
print('%-20s => %-20s, gt: %-20s' % (raw_pred, pred, gt))
accuracy = n_correct / float(max_iter * opt.batchSize)
print('Test loss: %f, accuray: %f' % (loss_avg.val(), accuracy))
if opt.cuda:
crnn.cuda()
image = image.cuda()
criterion = criterion.cuda()
image = Variable(image)
text = Variable(text)
length = Variable(length)
# loss averager
loss_avg = utils.averager()
# setup optimizer
if opt.adam:
optimizer = optim.Adam(crnn.parameters(), lr=opt.lr,
betas=(opt.beta1, 0.999))
elif opt.adadelta:
optimizer = optim.Adadelta(crnn.parameters(), lr=opt.lr)
else:
optimizer = optim.RMSprop(crnn.parameters(), lr=opt.lr)
def val(net, dataset, criterion, max_iter=100):
print('Start val')
for p in crnn.parameters():
p.requires_grad = False
net.eval()
data_loader = torch.utils.data.DataLoader(
def val(net, dataset, criterion, max_iter=1000, test_aug=False, n_aug=1):
print('Start validation set')
for p in crnn.parameters():
p.requires_grad = False
net.eval()
i = 0
n_correct = 0
loss_avg = utils.averager()
image_count = 0
# Character and word error rate lists
char_error = []
w_error = []
pred_dict = {}
gt_dict = {}
for epoch in range(n_aug):
max_iter = len(dataset) if test_aug else min(max_iter, len(dataset))
val_iter = iter(dataset)
for i in range(max_iter):
data = val_iter.next()
i += 1
cpu_images, cpu_texts, cpu_files = data
batch_size = cpu_images.size(0)
image_count = image_count + batch_size
utils.loadData(image, cpu_images)
t, l = converter.encode(cpu_texts)
utils.loadData(text, t)
utils.loadData(length, l)
preds = crnn(image)
#print(preds.size())
preds_size = Variable(torch.IntTensor([preds.size(0)] * batch_size))
cost = criterion(preds, text, preds_size, length) / batch_size
loss_avg.add(cost)
# RA: While I am not sure yet, it looks like a greedy decoder and not beam search is being used here
# Case is ignored in the accuracy, which is not ideal for an actual working system
_, preds = preds.max(2)
if torch.__version__ < '0.2':
preds = preds.squeeze(2) # https://github.com/meijieru/crnn.pytorch/issues/31
preds = preds.transpose(1, 0).contiguous().view(-1)
sim_preds = converter.decode(preds.data, preds_size.data, raw=False)
for pred, target, f in zip(sim_preds, cpu_texts, cpu_files):
if f not in gt_dict:
gt_dict[f] = target
pred_dict[f] = []
pred_dict[f].append(pred)
if pred == target:
n_correct += 1
# Case-sensitive character and word error rates
for f, target in gt_dict.items():
# Finds the most commonly predicted string for all the augmented images
best_pred = Counter(pred_dict[f]).most_common(1)[0][0]
char_error.append(cer(best_pred, target))
w_error.append(wer(best_pred, target))
raw_preds = converter.decode(preds.data, preds_size.data, raw=True)[:opt.n_test_disp]
for raw_pred, pred, gt in zip(raw_preds, sim_preds, cpu_texts):
print('%-20s => %-20s, gt: %-20s' % (raw_pred, pred, gt))
print("Total number of images in validation set: %8d" % image_count)
accuracy = n_correct / float(max_iter * opt.batchSize)
print('Test loss: %f, accuracy: %f' % (loss_avg.val(), accuracy))
char_arr = np.array(char_error)
w_arr = np.array(w_error)
char_mean_error = np.mean(char_arr)
word_mean_error = np.mean(w_arr)
print("Character error rate mean: %4.4f; Character error rate sd: %4.4f" % (
char_mean_error, np.std(char_arr, ddof=1)))
print("Word error rate mean: %4.4f; Word error rate sd: %4.4f" % (word_mean_error, np.std(w_arr, ddof=1)))
return char_mean_error, word_mean_error, accuracy
if opt.cuda:
crnn.cuda()
traject = traject.cuda()
criterion = criterion.cuda()
traject = Variable(traject)
text = Variable(text)
length = Variable(length)
# loss averager
loss_avg = utils.averager()
# setup optimizer
if opt.adam:
optimizer = optim.Adam(crnn.parameters(), lr=opt.lr,
betas=(opt.beta1, 0.999))
elif opt.adadelta:
optimizer = optim.Adadelta(crnn.parameters(), lr=opt.lr)
elif opt.rmsprop:
optimizer = optim.RMSprop(crnn.parameters(), lr=opt.lr)
else:
optimizer = torch.optim.SGD(crnn.parameters(), lr=opt.lr,
momentum=0.9,
weight_decay=1e-4)
def padding2tensor(trajects):
nTrajs = len(trajects)
maxLen = max([len(trajects[i].split(',')) for i in range(len(trajects))])
new_trajects = torch.LongTensor(nTrajs, maxLen)
for i in range(nTrajs):
def val(net, dataset, criterion, max_iter=100):
print('Start val')
for p in crnn.parameters():
p.requires_grad = False
net.eval()
data_loader = torch.utils.data.DataLoader(dataset,
shuffle=True,
batch_size=params.batchSize,
num_workers=int(params.workers))
val_iter = iter(data_loader)
i = 0
n_correct = 0
loss_avg = utils.averager()
max_iter = min(max_iter, len(data_loader))
for i in range(max_iter):
data = val_iter.next()
i += 1
cpu_images, cpu_texts = data
batch_size = cpu_images.size(0)
utils.loadData(image, cpu_images)
t, l = converter.encode(cpu_texts)
utils.loadData(text, t)
utils.loadData(length, l)
preds = crnn(image)
# print('-----preds-----')
# print(preds)
preds_size = Variable(torch.IntTensor([preds.size(0)] * batch_size))
# print('-----preds_size-----')
# print(preds_size)
cost = criterion(preds, text, preds_size, length) / batch_size
loss_avg.add(cost)
_, preds = preds.max(2)
# print('-----preds.max(2)-----')
# print(preds)
preds = preds.transpose(1, 0).contiguous().view(-1)
# print('-----preds.transpose(1, 0)-----')
# print(preds)
sim_preds = converter.decode(preds.data, preds_size.data, raw=False)
list_1 = []
for m in cpu_texts:
list_1.append(m.decode('utf-8', 'strict'))
# if (i - 1) % 10 == 0:
# print('-----sim_preds-----list_1-----')
# print(sim_preds, list_1)
for pred, target in zip(sim_preds, list_1):
if pred == target:
n_correct += 1
# else:
# print('%-20s, gt: %-20s' % (pred, target))
raw_preds = converter.decode(preds.data, preds_size.data,
raw=True)[:params.n_test_disp]
for raw_pred, pred, gt in zip(raw_preds, sim_preds, list_1):
print('%-20s => %-20s, gt: %-20s' % (raw_pred, pred, gt))
print(n_correct)
print(max_iter * params.batchSize)
accuracy = n_correct / float(max_iter * params.batchSize)
print('Test loss: %f, accuray: %f' % (loss_avg.val(), accuracy))
if opt.cuda:
crnn.cuda()
crnn = torch.nn.DataParallel(crnn, device_ids=range(opt.ngpu))
image = image.cuda()
criterion = criterion.cuda()
image = Variable(image)
text = Variable(text)
length = Variable(length)
# loss averager
loss_avg = utils.averager()
# setup optimizer
if opt.adam:
optimizer = optim.Adam(crnn.parameters(),
lr=opt.lr,
betas=(opt.beta1, 0.999))
elif opt.adadelta:
optimizer = optim.Adadelta(crnn.parameters(), lr=opt.lr)
else:
optimizer = optim.RMSprop(crnn.parameters(), lr=opt.lr)
def val(net, dataset, criterion, max_iter=100):
print('Start val')
for p in crnn.parameters():
p.requires_grad = False
net.eval()
nclass = len(opt.alphabet) + 1
nc = 3
converter = utils.strLabelConverter(opt.alphabet)
criterion = CTCLoss()
crnn = crnn.CRNN(opt.imgH, nc, nclass, opt.nh)
print('loading pretrained model from %s' % opt.crnn)
old = torch.load(opt.crnn)
new_state_dict = {}
for k, v in old.items():
new_state_dict[k[7:]] = v
crnn.load_state_dict(new_state_dict)
for p in crnn.parameters():
p.requires_grad = False
if opt.cuda:
crnn.cuda()
crnn = torch.nn.DataParallel(crnn, device_ids=range(opt.ngpu))
def bytes2cv2(b):
nparr = np.fromstring(b, np.uint8)
img_np = cv2.imdecode(nparr, cv2.IMREAD_COLOR)
return img_np
def run(tensor):
tensor = tensor.view(-1, *tensor.shape)
preds = crnn(tensor)
if opt.cuda:
crnn.cuda()
image = image.cuda()
criterion = criterion.cuda()
summary(crnn.cnn, (3, opt.imgH, opt.imgW))
image = Variable(image)
text = Variable(text)
length = Variable(length)
train_loss_avg = utils.averager()
train_cer_avg = utils.averager()
# setup optimizer
optimizer = optim.Adam(crnn.parameters(), lr=opt.lr)
def val(net, data_loader, criterion, max_iter=1000):
print('Start val')
for p in crnn.parameters():
p.requires_grad = False
net.eval()
val_iter = iter(data_loader)
val_loss_avg = utils.averager()
val_cer_avg = utils.averager()
max_iter = min(max_iter, len(data_loader))
def val(net, dataset, criterion, max_iter=2):
print('Start val')
for p in crnn.parameters():
p.requires_grad = False
net.eval()
data_loader = torch.utils.data.DataLoader(dataset,
shuffle=False,
batch_size=opt.batchSize,
num_workers=int(opt.workers))
val_iter = iter(data_loader)
i = 0
n_correct = 0
loss_avg = utils.averager()
max_iter = min(max_iter, len(data_loader))
for i in range(max_iter):
data = val_iter.next()
i += 1
cpu_images, cpu_texts = data
batch_size = cpu_images.size(0)
utils.loadData(image, cpu_images)
if ifUnicode:
cpu_texts = [clean_txt(tx.decode('utf-8')) for tx in cpu_texts]
# print(cpu_texts)
t, l = converter.encode(cpu_texts)
# print(t)
# print(l)
utils.loadData(text, t)
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
loss_avg.add(cost)
# print(preds)
# print(preds.shape)
_, preds = preds.max(2)
# print(preds)
# print(preds.shape)
# preds = preds.squeeze(2)
# print(preds)
# print(preds.shape)
preds = preds.transpose(1, 0).contiguous().view(-1)
# print(preds)
# print(preds.shape)
sim_preds = converter.decode(preds.data, preds_size.data, raw=False)
print(sim_preds)
print(cpu_texts)
for pred, target in zip(sim_preds, cpu_texts):
if pred.strip() == target.strip():
n_correct += 1
# raw_preds = converter.decode(preds.data, preds_size.data, raw=True)[:opt.n_test_disp]
# for raw_pred, pred, gt in zip(raw_preds, sim_preds, cpu_texts):
# print((pred, gt))
# print
accuracy = n_correct / float(max_iter * opt.batchSize)
testLoss = loss_avg.val()
print('Test loss: %f, accuray: %f' % (testLoss, accuracy))
return testLoss, accuracy