def test(opt):
""" model configuration """
if 'CTC' in opt.Prediction:
converter = CTCLabelConverter(opt.character)
else:
converter = AttnLabelConverter(opt.character)
opt.num_class = len(converter.character)
if opt.rgb:
opt.input_channel = 3
model = Model(opt)
print('model input parameters', opt.imgH, opt.imgW, opt.num_fiducial,
opt.input_channel, opt.output_channel, opt.hidden_size,
opt.num_class, opt.batch_max_length, opt.Transformation,
opt.FeatureExtraction, opt.SequenceModeling, opt.Prediction)
model = torch.nn.DataParallel(model).to(device)
# load model
print('loading pretrained model from %s' % opt.saved_model)
model.load_state_dict(torch.load(opt.saved_model, map_location=device))
opt.exp_name = '_'.join(opt.saved_model.split('/')[1:])
# print(model)
""" keep evaluation model and result logs """
os.makedirs(f'./result/{opt.exp_name}', exist_ok=True)
os.system(f'cp {opt.saved_model} ./result/{opt.exp_name}/')
""" setup loss """
if 'CTC' in opt.Prediction:
criterion = torch.nn.CTCLoss(zero_infinity=True).to(device)
else:
criterion = torch.nn.CrossEntropyLoss(ignore_index=0).to(device)
""" evaluation """
model.eval()
with torch.no_grad():
if opt.benchmark_all_eval:
benchmark_all_eval(model, criterion, converter, opt)
else:
log = open(f'./result/{opt.exp_name}/log_evaluation.txt', 'a')
AlignCollate_evaluation = AlignCollate(imgH=opt.imgH,
imgW=opt.imgW,
keep_ratio_with_pad=opt.PAD)
eval_data, eval_data_log = hierarchical_dataset(root=opt.eval_data,
opt=opt)
evaluation_loader = torch.utils.data.DataLoader(
eval_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=int(opt.workers),
collate_fn=AlignCollate_evaluation,
pin_memory=True)
_, accuracy_by_best_model, _, _, _, _, _, _ = validation(
model, criterion, evaluation_loader, converter, opt)
log.write(eval_data_log)
print(f'{accuracy_by_best_model:0.3f}')
log.write(f'{accuracy_by_best_model:0.3f}\n')
log.close()
def train(opt):
""" dataset preparation """
opt.select_data = opt.select_data.split('-')
opt.batch_ratio = opt.batch_ratio.split('-')
train_dataset = Batch_Balanced_Dataset(opt)
AlignCollate_valid = AlignCollate(imgH=opt.imgH, imgW=opt.imgW, keep_ratio_with_pad=opt.PAD)
valid_dataset = hierarchical_dataset(root=opt.valid_data, opt=opt)
valid_loader = torch.utils.data.DataLoader(
valid_dataset, batch_size=opt.batch_size,
shuffle=True, # 'True' to check training progress with validation function.
num_workers=int(opt.workers),
collate_fn=AlignCollate_valid, pin_memory=True)
print('-' * 80)
""" model configuration """
if 'CTC' in opt.Prediction:
converter = CTCLabelConverter(opt.character)
else:
converter = AttnLabelConverter(opt.character)
opt.num_class = len(converter.character)
if opt.rgb:
opt.input_channel = 3
model = Model(opt)
print('model input parameters', opt.imgH, opt.imgW, opt.num_fiducial, opt.input_channel, opt.output_channel,
opt.hidden_size, opt.num_class, opt.batch_max_length, opt.Transformation, opt.FeatureExtraction,
opt.SequenceModeling, opt.Prediction)
# weight initialization
for name, param in model.named_parameters():
if 'localization_fc2' in name:
print(f'Skip {name} as it is already initialized')
continue
try:
if 'bias' in name:
init.constant_(param, 0.0)
elif 'weight' in name:
init.kaiming_normal_(param)
except Exception as e: # for batchnorm.
if 'weight' in name:
param.data.fill_(1)
continue
# data parallel for multi-GPU
model = torch.nn.DataParallel(model).cuda()
model.train()
if opt.continue_model != '':
if opt.without_prediction:
load_model_without_prediction(opt.continue_model, model)
print(f'loading pretrained model from {opt.continue_model}, without prediction layer')
else:
print(f'loading pretrained model from {opt.continue_model}')
model.load_state_dict(torch.load(opt.continue_model))
print("Model:")
print(model)
""" setup loss """
if 'CTC' in opt.Prediction:
criterion = torch.nn.CTCLoss(zero_infinity=True).cuda()
else:
criterion = torch.nn.CrossEntropyLoss(ignore_index=0).cuda() # ignore [GO] token = ignore index 0
# loss averager
loss_avg = Averager()
# filter that only require gradient decent
filtered_parameters = []
params_num = []
for p in filter(lambda p: p.requires_grad, model.parameters()):
filtered_parameters.append(p)
params_num.append(np.prod(p.size()))
print('Trainable params num : ', sum(params_num))
# [print(name, p.numel()) for name, p in filter(lambda p: p[1].requires_grad, model.named_parameters())]
# setup optimizer
if opt.adam:
optimizer = optim.Adam(filtered_parameters, lr=opt.lr, betas=(opt.beta1, 0.999))
else:
optimizer = optim.Adadelta(filtered_parameters, lr=opt.lr, rho=opt.rho, eps=opt.eps)
print("Optimizer:")
print(optimizer)
""" final options """
# print(opt)
with open(f'./saved_models/{opt.experiment_name}/opt.txt', 'a') as opt_file:
opt_log = '------------ Options -------------\n'
args = vars(opt)
for k, v in args.items():
opt_log += f'{str(k)}: {str(v)}\n'
opt_log += '---------------------------------------\n'
print(opt_log)
opt_file.write(opt_log)
""" start training """
start_iter = 0
if opt.continue_model != '':
print(f'continue to train, start_iter: {start_iter}')
start_time = time.time()
best_accuracy = -1
best_norm_ED = 1e+6
i = start_iter
while True:
# train part
for p in model.parameters():
p.requires_grad = True
image_tensors, labels = train_dataset.get_batch()
image = image_tensors.cuda()
text, length = converter.encode(labels)
batch_size = image.size(0)
if 'CTC' in opt.Prediction:
preds = model(image, text).log_softmax(2)
preds_size = torch.IntTensor([preds.size(1)] * batch_size)
preds = preds.permute(1, 0, 2) # to use CTCLoss format
cost = criterion(preds, text, preds_size, length)
else:
preds = model(image, text)
target = text[:, 1:] # without [GO] Symbol
cost = criterion(preds.view(-1, preds.shape[-1]), target.contiguous().view(-1))
model.zero_grad()
cost.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), opt.grad_clip) # gradient clipping with 5 (Default)
optimizer.step()
loss_avg.add(cost)
# validation part
if i % opt.valInterval == 0:
elapsed_time = time.time() - start_time
logging.info(f'[{i}/{opt.num_iter}] Loss: {loss_avg.val():0.5f} elapsed_time: {elapsed_time:0.5f}')
# for log
with open(f'./saved_models/{opt.experiment_name}/log_train.txt', 'a') as log:
log.write(f'[{i}/{opt.num_iter}] Loss: {loss_avg.val():0.5f} elapsed_time: {elapsed_time:0.5f}\n')
loss_avg.reset()
model.eval()
valid_loss, current_accuracy, current_norm_ED, preds, labels, infer_time, length_of_data = validation(
model, criterion, valid_loader, converter, opt)
model.train()
for pred, gt in zip(preds[:5], labels[:5]):
if 'Attn' in opt.Prediction:
pred = pred[:pred.find('[s]')]
gt = gt[:gt.find('[s]')]
print(f'{pred:20s}, gt: {gt:20s}, {str(pred == gt)}')
log.write(f'{pred:20s}, gt: {gt:20s}, {str(pred == gt)}\n')
valid_log = f'[{i}/{opt.num_iter}] valid loss: {valid_loss:0.5f}'
valid_log += f' accuracy: {current_accuracy:0.3f}, norm_ED: {current_norm_ED:0.2f}'
log.write(valid_log + '\n')
# keep best accuracy model
if current_accuracy > best_accuracy:
best_accuracy = current_accuracy
torch.save(model.state_dict(), f'./saved_models/{opt.experiment_name}/mtl_best_accuracy.pth')
if current_norm_ED < best_norm_ED:
best_norm_ED = current_norm_ED
torch.save(model.state_dict(), f'./saved_models/{opt.experiment_name}/best_norm_ED.pth')
best_model_log = f'best_accuracy: {best_accuracy:0.3f}, best_norm_ED: {best_norm_ED:0.2f}'
logging.info(best_model_log)
log.write(best_model_log + '\n')
# save model per 1e+5 iter.
if (i + 1) % 50000 == 0:
torch.save(
model.state_dict(), f'./saved_models/{opt.experiment_name}/iter_{i+1}.pth')
if i == opt.num_iter:
logging.info('end the training')
sys.exit()
i += 1
def train(opt):
""" Dataset Preparation """
if not opt.data_filtering_off:
print(
'Filtering the images containing characters which are not in opt.character'
)
print(
'Filtering the images whose label is longer than opt.batch_max_length'
)
opt.select_data = opt.select_data.split('-')
opt.batch_ratio = opt.batch_ratio.split('-')
train_dataset = Batch_Balanced_Dataset(opt)
log = open(f'./saved_models/{opt.exp_name}/log_dataset.txt', 'a')
AlignCollate_valid = AlignCollate(imgH=opt.imgH,
imgW=opt.imgW,
keep_ratio_with_pad=opt.PAD)
valid_dataset, valid_dataset_log = hierarchical_dataset(
root=opt.valid_data, opt=opt)
valid_loader = torch.utils.data.DataLoader(
valid_dataset,
batch_size=opt.batch_size,
shuffle=
True, # 'True' to check training progress with validation function.
num_workers=int(opt.workers),
collate_fn=AlignCollate_valid,
pin_memory=True)
log.write(valid_dataset_log)
print('-' * 80)
log.write('-' * 80 + '\n')
log.close()
""" Model Configuration """
if 'CTC' in opt.Prediction:
if opt.baiduCTC:
converter = CTCLabelConverterForBaiduWarpctc(opt.character)
else:
converter = CTCLabelConverter(opt.character)
else:
converter = AttnLabelConverter(opt.character)
opt.num_class = len(converter.character)
if opt.rgb:
opt.input_channel = 3
model = Model(opt)
print('model input parameters', opt.imgH, opt.imgW, opt.num_fiducial,
opt.input_channel, opt.output_channel, opt.hidden_size,
opt.num_class, opt.batch_max_length, opt.Transformation,
opt.FeatureExtraction, opt.SequenceModeling, opt.Prediction)
# weight initialization
for name, param in model.named_parameters():
if 'localization_fc2' in name:
print(f'Skip {name} as it is already initialized')
continue
try:
if 'bias' in name:
init.constant_(param, 0.0)
elif 'weight' in name:
init.kaiming_normal_(param)
except Exception as e: # for batchnorm.
if 'weight' in name:
param.data.fill_(1)
continue
# data parallel for multi-GPU
model = torch.nn.DataParallel(model).to(device)
model.train()
if opt.saved_model != '':
print(f'loading pretrained model from {opt.saved_model}')
if opt.FT:
model.load_state_dict(torch.load(opt.saved_model), strict=False)
else:
model.load_state_dict(torch.load(opt.saved_model))
print("Model:")
print(model)
""" Setup Loss """
if 'CTC' in opt.Prediction:
if opt.baiduCTC:
# need to install warpctc. see our guideline.
from warpctc_pytorch import CTCLoss
criterion = CTCLoss()
else:
criterion = torch.nn.CTCLoss(zero_infinity=True).to(device)
else:
criterion = torch.nn.CrossEntropyLoss(ignore_index=0).to(
device) # ignore [GO] token = ignore index 0
# loss averager
loss_avg = Averager()
# filter that only require gradient decent
filtered_parameters = []
params_num = []
for p in filter(lambda p: p.requires_grad, model.parameters()):
filtered_parameters.append(p)
params_num.append(np.prod(p.size()))
print('Trainable params num : ', sum(params_num))
# [print(name, p.numel()) for name, p in filter(lambda p: p[1].requires_grad, model.named_parameters())]
# setup optimizer
if opt.adam:
optimizer = optim.Adam(filtered_parameters,
lr=opt.lr,
betas=(opt.beta1, 0.999))
else:
optimizer = optim.Adadelta(filtered_parameters,
lr=opt.lr,
rho=opt.rho,
eps=opt.eps)
print("Optimizer:")
print(optimizer)
""" Final Options """
# print(opt)
with open(f'./saved_models/{opt.exp_name}/opt.txt', 'a') as opt_file:
opt_log = '------------ Options -------------\n'
args = vars(opt)
for k, v in args.items():
opt_log += f'{str(k)}: {str(v)}\n'
opt_log += '---------------------------------------\n'
print(opt_log)
opt_file.write(opt_log)
""" Start Training """
start_iter = 0
if opt.saved_model != '':
try:
start_iter = int(opt.saved_model.split('_')[-1].split('.')[0])
print(f'continue to train, start_iter: {start_iter}')
except:
pass
start_time = time.time()
best_accuracy = -1
best_norm_ED = -1
iteration = start_iter
while True:
# train part
image_tensors, labels = train_dataset.get_batch()
image = image_tensors.to(device)
text, length = converter.encode(labels,
batch_max_length=opt.batch_max_length)
batch_size = image.size(0)
if 'CTC' in opt.Prediction:
preds = model(image, text)
preds_size = torch.IntTensor([preds.size(1)] * batch_size)
if opt.baiduCTC:
preds = preds.permute(1, 0, 2) # to use CTCLoss format
cost = criterion(preds, text, preds_size, length) / batch_size
else:
preds = preds.log_softmax(2).permute(1, 0, 2)
cost = criterion(preds, text, preds_size, length)
else:
preds = model(image, text[:, :-1]) # align with Attention.forward
target = text[:, 1:] # without [GO] Symbol
cost = criterion(preds.view(-1, preds.shape[-1]),
target.contiguous().view(-1))
model.zero_grad()
cost.backward()
torch.nn.utils.clip_grad_norm_(
model.parameters(),
opt.grad_clip) # gradient clipping with 5 (Default)
optimizer.step()
loss_avg.add(cost)
# validation part
if (
iteration + 1
) % opt.valInterval == 0 or iteration == 0: # To see training progress, we also conduct validation when 'iteration == 0'
elapsed_time = time.time() - start_time
# for log
with open(f'./saved_models/{opt.exp_name}/log_train.txt',
'a') as log:
model.eval()
with torch.no_grad():
valid_loss, current_accuracy, current_norm_ED, preds, confidence_score, labels, infer_time, length_of_data = validation(
model, criterion, valid_loader, converter, opt)
model.train()
# training loss and validation loss
loss_log = f'[{iteration + 1}/{opt.num_iter}] Train loss: {loss_avg.val():0.5f}, Valid loss: {valid_loss:0.5f}, Elapsed_time: {elapsed_time:0.5f}'
loss_avg.reset()
current_model_log = f'{"Current_accuracy":17s}: {current_accuracy:0.3f}, {"Current_norm_ED":17s}: {current_norm_ED:0.2f}'
# keep best accuracy model (on valid dataset)
if current_accuracy > best_accuracy:
best_accuracy = current_accuracy
torch.save(
model.state_dict(),
f'./saved_models/{opt.exp_name}/best_accuracy.pth')
if current_norm_ED > best_norm_ED:
best_norm_ED = current_norm_ED
torch.save(
model.state_dict(),
f'./saved_models/{opt.exp_name}/best_norm_ED.pth')
best_model_log = f'{"Best_accuracy":17s}: {best_accuracy:0.3f}, {"Best_norm_ED":17s}: {best_norm_ED:0.2f}'
loss_model_log = f'{loss_log}\n{current_model_log}\n{best_model_log}'
print(loss_model_log)
log.write(loss_model_log + '\n')
# show some predicted results
dashed_line = '-' * 80
head = f'{"Ground Truth":25s} | {"Prediction":25s} | Confidence Score & T/F'
predicted_result_log = f'{dashed_line}\n{head}\n{dashed_line}\n'
for gt, pred, confidence in zip(labels[:5], preds[:5],
confidence_score[:5]):
if 'Attn' in opt.Prediction:
gt = gt[:gt.find('[s]')]
pred = pred[:pred.find('[s]')]
predicted_result_log += f'{gt:25s} | {pred:25s} | {confidence:0.4f}\t{str(pred == gt)}\n'
predicted_result_log += f'{dashed_line}'
print(predicted_result_log)
log.write(predicted_result_log + '\n')
# save model per 1e+5 iter.
if (iteration + 1) % 1e+5 == 0:
torch.save(
model.state_dict(),
f'./saved_models/{opt.exp_name}/iter_{iteration + 1}.pth')
if (iteration + 1) == opt.num_iter:
print('end the training')
sys.exit()
iteration += 1
def benchmark_all_eval(model,
criterion,
converter,
opt,
calculate_infer_time=False):
""" Evaluation with 10 benchmark evaluation datasets """
# The evaluation datasets, dataset order is same with Table 1 in our paper.
eval_data_list = [
'IIIT5k_3000', 'SVT', 'IC03_860', 'IC03_867', 'IC13_857', 'IC13_1015',
'IC15_1811', 'IC15_2077', 'SVTP', 'CUTE80'
]
# # To easily compute the total accuracy of our paper.
# eval_data_list = ['IIIT5k_3000', 'SVT', 'IC03_867',
# 'IC13_1015', 'IC15_2077', 'SVTP', 'CUTE80']
if calculate_infer_time:
evaluation_batch_size = 1 # batch_size should be 1 to calculate the GPU inference time per image.
else:
evaluation_batch_size = opt.batch_size
list_accuracy = []
total_forward_time = 0
total_evaluation_data_number = 0
total_correct_number = 0
log = open(f'./result/{opt.exp_name}/log_all_evaluation.txt', 'a')
dashed_line = '-' * 80
print(dashed_line)
log.write(dashed_line + '\n')
for eval_data in eval_data_list:
eval_data_path = os.path.join(opt.eval_data, eval_data)
AlignCollate_evaluation = AlignCollate(imgH=opt.imgH,
imgW=opt.imgW,
keep_ratio_with_pad=opt.PAD)
eval_data, eval_data_log = hierarchical_dataset(root=eval_data_path,
opt=opt)
evaluation_loader = torch.utils.data.DataLoader(
eval_data,
batch_size=evaluation_batch_size,
shuffle=False,
num_workers=int(opt.workers),
collate_fn=AlignCollate_evaluation,
pin_memory=True)
_, accuracy_by_best_model, norm_ED_by_best_model, _, _, _, infer_time, length_of_data = validation(
model, criterion, evaluation_loader, converter, opt)
list_accuracy.append(f'{accuracy_by_best_model:0.3f}')
total_forward_time += infer_time
total_evaluation_data_number += len(eval_data)
total_correct_number += accuracy_by_best_model * length_of_data
log.write(eval_data_log)
print(
f'Acc {accuracy_by_best_model:0.3f}\t normalized_ED {norm_ED_by_best_model:0.3f}'
)
log.write(
f'Acc {accuracy_by_best_model:0.3f}\t normalized_ED {norm_ED_by_best_model:0.3f}\n'
)
print(dashed_line)
log.write(dashed_line + '\n')
averaged_forward_time = total_forward_time / total_evaluation_data_number * 1000
total_accuracy = total_correct_number / total_evaluation_data_number
params_num = sum([np.prod(p.size()) for p in model.parameters()])
evaluation_log = 'accuracy: '
for name, accuracy in zip(eval_data_list, list_accuracy):
evaluation_log += f'{name}: {accuracy}\t'
evaluation_log += f'total_accuracy: {total_accuracy:0.3f}\t'
evaluation_log += f'averaged_infer_time: {averaged_forward_time:0.3f}\t# parameters: {params_num / 1e6:0.3f}'
print(evaluation_log)
log.write(evaluation_log + '\n')
log.close()
return None