src_path = opt.src_path
tgt_path = opt.tgt_path
swapper.swap_setup(src_path, tgt_path)
if opt.post_tune:
print('\n\t\t\tPersonalization: meta cycle finetune...')
swapper.post_personalize(opt.output_dir,
visualizer=None,
verbose=False)
# swapper.post_personalize(opt.output_dir, visualizer=visualizer, verbose=True, bidirection=opt.bidirection)
print('\n\t\t\tPersonalization: completed...')
# if a->b
print('\n\t\t\tSwapping: {} wear the clothe of {}...'.format(
src_path, tgt_path))
preds = swapper.swap(src_info=swapper.src_info,
tgt_info=swapper.tsf_info,
target_part=opt.swap_part,
visualizer=visualizer)
if opt.save_res:
pred_output_dir = mkdir(os.path.join(opt.output_dir, 'swappers'))
save_results(src_path, tgt_path, pred_output_dir, preds)
# # else b->a
# preds = swapper.swap(src_info=swapper.tgt_info, tgt_info=swapper.src_info,
# target_part=opt.swap_part, visualizer=visualizer)
def main(jsonPath):
# options
opt = option.parse(jsonPath, is_train=False)
util.mkdirs((path for key, path in opt["path"].items()
if not key == "pretrain_model_G"))
opt = option.dict_to_nonedict(opt)
util.setup_logger(None,
opt["path"]["log"],
"test.log",
level=logging.INFO,
screen=True)
logger = logging.getLogger("base")
logger.info(option.dict2str(opt))
# Create test dataset and dataloader
test_loaders = []
for phase, dataset_opt in sorted(opt["datasets"].items()):
test_set = create_dataset(dataset_opt)
test_loader = create_dataloader(test_set, dataset_opt)
logger.info("Number of test images in [{:s}]: {:d}".format(
dataset_opt["name"], len(test_set)))
test_loaders.append(test_loader)
# Create model
model = create_model(opt)
for test_loader in test_loaders:
test_set_name = test_loader.dataset.opt["name"]
logger.info("\nTesting [{:s}]...".format(test_set_name))
# test_start_time = time.time()
dataset_dir = os.path.join(opt["path"]["results_root"], test_set_name)
util.mkdir(dataset_dir)
test_results = OrderedDict()
test_results["psnr"] = []
test_results["ssim"] = []
test_results["psnr_y"] = []
test_results["ssim_y"] = []
for data in test_loader:
need_HR = False if test_loader.dataset.opt[
"dataroot_HR"] is None else True
model.feed_data(data, need_HR=need_HR)
img_path = data["LR_path"][0]
img_name = os.path.splitext(os.path.basename(img_path))[0]
model.test() # test
visuals = model.get_current_visuals(need_HR=need_HR)
sr_img = util.tensor2img(visuals["SR"]) # uint8
# save images
suffix = opt["suffix"]
if suffix:
save_img_path = os.path.join(dataset_dir,
img_name + suffix + ".png")
else:
save_img_path = os.path.join(dataset_dir, img_name + ".png")
util.save_img(sr_img, save_img_path)
# calculate PSNR and SSIM
if need_HR:
gt_img = util.tensor2img(visuals["HR"])
gt_img = gt_img / 255.0
sr_img = sr_img / 255.0
crop_border = test_loader.dataset.opt["scale"]
cropped_sr_img = sr_img[crop_border:-crop_border,
crop_border:-crop_border, :]
cropped_gt_img = gt_img[crop_border:-crop_border,
crop_border:-crop_border, :]
psnr = util.calculate_psnr(cropped_sr_img * 255,
cropped_gt_img * 255)
ssim = util.calculate_ssim(cropped_sr_img * 255,
cropped_gt_img * 255)
test_results["psnr"].append(psnr)
test_results["ssim"].append(ssim)
if gt_img.shape[2] == 3: # RGB image
sr_img_y = bgr2ycbcr(sr_img, only_y=True)
gt_img_y = bgr2ycbcr(gt_img, only_y=True)
cropped_sr_img_y = sr_img_y[crop_border:-crop_border,
crop_border:-crop_border]
cropped_gt_img_y = gt_img_y[crop_border:-crop_border,
crop_border:-crop_border]
psnr_y = util.calculate_psnr(cropped_sr_img_y * 255,
cropped_gt_img_y * 255)
ssim_y = util.calculate_ssim(cropped_sr_img_y * 255,
cropped_gt_img_y * 255)
test_results["psnr_y"].append(psnr_y)
test_results["ssim_y"].append(ssim_y)
logger.info(
"{:20s} - PSNR: {:.6f} dB; SSIM: {:.6f}; PSNR_Y: {:.6f} dB; SSIM_Y: {:.6f}."
.format(img_name, psnr, ssim, psnr_y, ssim_y))
else:
logger.info(
"{:20s} - PSNR: {:.6f} dB; SSIM: {:.6f}.".format(
img_name, psnr, ssim))
else:
logger.info(img_name)
if need_HR: # metrics
# Average PSNR/SSIM results
ave_psnr = sum(test_results["psnr"]) / len(test_results["psnr"])
ave_ssim = sum(test_results["ssim"]) / len(test_results["ssim"])
logger.info(
"----Average PSNR/SSIM results for {}----\n\tPSNR: {:.6f} dB; SSIM: {:.6f}\n"
.format(test_set_name, ave_psnr, ave_ssim))
if test_results["psnr_y"] and test_results["ssim_y"]:
ave_psnr_y = sum(test_results["psnr_y"]) / len(
test_results["psnr_y"])
ave_ssim_y = sum(test_results["ssim_y"]) / len(
test_results["ssim_y"])
logger.info(
"----Y channel, average PSNR/SSIM----\n\tPSNR_Y: {:.6f} dB; SSIM_Y: {:.6f}\n"
.format(ave_psnr_y, ave_ssim_y))
def main(config_path):
# Read a config file (.yml)
with open(config_path, "r") as f:
config = yaml.load(f)
corpus = config['corpus']
feature = config['feature']
param = config['param']
# TODO: Solve conflict (batch_norm & layer norm)
if corpus['label_type'] == 'phone61':
output_size = 61
elif corpus['label_type'] == 'phone48':
output_size = 48
elif corpus['label_type'] == 'phone39':
output_size = 39
elif corpus['label_type'] == 'character':
output_size = 30
# Model setting
CTCModel = load(model_type=config['model_name'])
network = CTCModel(batch_size=param['batch_size'],
input_size=feature['input_size'] * feature['num_stack'],
num_cell=param['num_cell'],
num_layer=param['num_layer'],
output_size=output_size,
clip_gradients=param['clip_grad'],
clip_activation=param['clip_activation'],
dropout_ratio_input=param['dropout_input'],
dropout_ratio_hidden=param['dropout_hidden'],
num_proj=param['num_proj'],
weight_decay=param['weight_decay'])
network.model_name = config['model_name'].upper()
network.model_name += '_' + str(param['num_cell'])
network.model_name += '_' + str(param['num_layer'])
network.model_name += '_' + param['optimizer']
network.model_name += '_lr' + str(param['learning_rate'])
if param['num_proj'] != 0:
network.model_name += '_proj' + str(param['num_proj'])
if feature['num_stack'] != 1:
network.model_name += '_stack' + str(feature['num_stack'])
if param['weight_decay'] != 0:
network.model_name += '_weightdecay' + str(param['weight_decay'])
# Set save path
network.model_dir = mkdir('/n/sd8/inaguma/result/timit/ctc/')
network.model_dir = mkdir_join(network.model_dir, corpus['label_type'])
network.model_dir = mkdir_join(network.model_dir, network.model_name)
# Reset model directory
if not isfile(join(network.model_dir, 'complete.txt')):
tf.gfile.DeleteRecursively(network.model_dir)
tf.gfile.MakeDirs(network.model_dir)
else:
raise ValueError('File exists.')
# Set process name
setproctitle('ctc_timit_' +
corpus['label_type'] + '_' + param['optimizer'])
# Save config file
shutil.copyfile(config_path, join(network.model_dir, 'config.yml'))
sys.stdout = open(join(network.model_dir, 'train.log'), 'w')
print(network.model_name)
do_train(network=network,
optimizer=param['optimizer'],
learning_rate=param['learning_rate'],
batch_size=param['batch_size'],
epoch_num=param['num_epoch'],
label_type=corpus['label_type'],
num_stack=feature['num_stack'],
num_skip=feature['num_skip'])
sys.stdout = sys.__stdout__
def main():
#### options
parser = argparse.ArgumentParser()
parser.add_argument('-opt', type=str, help='Path to option YAML file.')
parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none',
help='job launcher')
parser.add_argument('--local_rank', type=int, default=0)
args = parser.parse_args()
opt = option.parse(args.opt, is_train=True)
label_path = opt['datasets']['val']['dataroot_label_file']
#### distributed training settings
if args.launcher == 'none': # disabled distributed training
opt['dist'] = False
rank = -1
print('Disabled distributed training.')
else:
opt['dist'] = True
init_dist()
world_size = torch.distributed.get_world_size()
rank = torch.distributed.get_rank()
#### loading resume state if exists
if opt['path'].get('resume_state', None):
# distributed resuming: all load into default GPU
device_id = torch.cuda.current_device()
resume_state = torch.load(opt['path']['resume_state'],
map_location=lambda storage, loc: storage.cuda(device_id))
option.check_resume(opt, resume_state['iter']) # check resume options
else:
resume_state = None
#### mkdir and loggers
if rank <= 0: # normal training (rank -1) OR distributed training (rank 0)
if resume_state is None:
util.mkdir_and_rename(
opt['path']['experiments_root']) # rename experiment folder if exists
util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root'
and 'pretrain_model' not in key and 'resume' not in key))
# config loggers. Before it, the log will not work
util.setup_logger('base', opt['path']['log'], 'train_' + opt['name'], level=logging.INFO,
screen=True, tofile=True)
logger = logging.getLogger('base')
logger.info(option.dict2str(opt))
# tensorboard logger
if opt['use_tb_logger'] and 'debug' not in opt['name']:
version = float(torch.__version__[0:3])
if version >= 1.1: # PyTorch 1.1
from torch.utils.tensorboard import SummaryWriter
else:
logger.info(
'You are using PyTorch {}. Tensorboard will use [tensorboardX]'.format(version))
from tensorboardX import SummaryWriter
tb_logger = SummaryWriter(log_dir='../tb_logger/' + opt['name'])
else:
util.setup_logger('base', opt['path']['log'], 'train', level=logging.INFO, screen=True)
logger = logging.getLogger('base')
# convert to NoneDict, which returns None for missing keys
opt = option.dict_to_nonedict(opt)
#### random seed
seed = opt['train']['manual_seed']
if seed is None:
seed = random.randint(1, 10000)
if rank <= 0:
logger.info('Random seed: {}'.format(seed))
util.set_random_seed(seed)
torch.backends.cudnn.benchmark = True
# torch.backends.cudnn.deterministic = True
#### create train and val dataloader
dataset_ratio = 200 # enlarge the size of each epoch
for phase, dataset_opt in opt['datasets'].items():
if phase == 'train':
train_set = create_dataset(dataset_opt)
train_size = int(math.ceil(len(train_set) / dataset_opt['batch_size']))
total_iters = int(opt['train']['niter'])
total_epochs = int(math.ceil(total_iters / train_size))
if opt['dist']:
train_sampler = DistIterSampler(train_set, world_size, rank, dataset_ratio)
total_epochs = int(math.ceil(total_iters / (train_size * dataset_ratio)))
else:
train_sampler = None
train_loader = create_dataloader(train_set, dataset_opt, opt, train_sampler)
if rank <= 0:
logger.info('Number of train images: {:,d}, iters: {:,d}'.format(
len(train_set), train_size))
logger.info('Total epochs needed: {:d} for iters {:,d}'.format(
total_epochs, total_iters))
elif phase == 'val':
val_set = create_dataset(dataset_opt,is_train = False)
val_loader = create_dataloader(val_set, dataset_opt, opt, None)
if rank <= 0:
logger.info('Number of val images in [{:s}]: {:d}'.format(
dataset_opt['name'], len(val_set)))
else:
raise NotImplementedError('Phase [{:s}] is not recognized.'.format(phase))
assert train_loader is not None
#### create model
model = create_model(opt)
#### resume training
if resume_state:
logger.info('Resuming training from epoch: {}, iter: {}.'.format(
resume_state['epoch'], resume_state['iter']))
start_epoch = resume_state['epoch']
current_step = resume_state['iter']
model.resume_training(resume_state) # handle optimizers and schedulers
else:
current_step = 0
start_epoch = 0
#### training
logger.info('Start training from epoch: {:d}, iter: {:d}'.format(start_epoch, current_step))
for epoch in range(start_epoch, total_epochs + 1):
if opt['dist']:
train_sampler.set_epoch(epoch)
for _, train_data in enumerate(train_loader):
current_step += 1
if current_step > total_iters:
break
#### update learning rate
model.update_learning_rate(current_step, warmup_iter=opt['train']['warmup_iter'])
#### training
model.feed_data(train_data)
model.optimize_parameters(current_step)
#### log
if current_step % opt['logger']['print_freq'] == 0:
logs = model.get_current_log()
message = '[epoch:{:3d}, iter:{:8,d}, lr:('.format(epoch, current_step)
for v in model.get_current_learning_rate():
message += '{:.3e},'.format(v)
message += ')] '
for k, v in logs.items():
message += '{:s}: {:.4e} '.format(k, v)
# tensorboard logger
if opt['use_tb_logger'] and 'debug' not in opt['name']:
if rank <= 0:
tb_logger.add_scalar(k, v, current_step)
if rank <= 0:
logger.info(message)
#### validation
if opt['datasets'].get('val', None) and current_step % opt['train']['val_freq'] == 0:
if rank <= 0: #
# does not support multi-GPU validation
pbar = util.ProgressBar(len(val_loader))
idx = 0
for val_data in val_loader:
idx += 1
img_name = os.path.splitext(os.path.basename(val_data['img1_path'][0]))[0]
img_dir = os.path.join(opt['path']['val_images'], str(current_step))
util.mkdir(img_dir)
f = open(os.path.join(img_dir, 'predict_score.txt'), 'a')
model.feed_data(val_data)
model.test()
visuals = model.get_current_visuals()
predict_score1 = visuals['predict_score1'].numpy()
# Save predict scores
f.write('%s %f\n' % (img_name + '.png', predict_score1))
f.close()
pbar.update('Test {}'.format(img_name))
# calculate accuracy
aligned_pair_accuracy, accuracy_esrganbig, accuracy_srganbig = rank_pair_test(\
os.path.join(img_dir, 'predict_score.txt'), label_path)
# log
logger.info(
'# Validation # Accuracy: {:.4e}, Accuracy_pair1_class1: {:.4e}, Accuracy_pair1_class2: {:.4e} '.format(
aligned_pair_accuracy, accuracy_esrganbig, accuracy_srganbig))
logger_val = logging.getLogger('val') # validation logger
logger_val.info(
'<epoch:{:3d}, iter:{:8,d}> Accuracy: {:.4e}, Accuracy_pair1_class1: {:.4e}, Accuracy_pair1_class2: {:.4e} '.format(
epoch, current_step, aligned_pair_accuracy, accuracy_esrganbig, accuracy_srganbig))
# tensorboard logger
if opt['use_tb_logger'] and 'debug' not in opt['name']:
tb_logger.add_scalar('Accuracy', aligned_pair_accuracy, current_step)
tb_logger.add_scalar('Accuracy_pair1_class1', accuracy_esrganbig, current_step)
tb_logger.add_scalar('Accuracy_pair1_class2', accuracy_srganbig, current_step)
#### save models and training states
if current_step % opt['logger']['save_checkpoint_freq'] == 0:
if rank <= 0:
logger.info('Saving models and training states.')
model.save(current_step)
model.save_training_state(epoch, current_step)
if rank <= 0:
logger.info('Saving the final model.')
model.save('latest')
logger.info('End of training.')
tb_logger.close()
def summary_dir(self): return ut.mkdir(pj(self.resdir, 'summary'))
if test_opt.ip:
visualizer = VisdomVisualizer(env=test_opt.name,
ip=test_opt.ip,
port=test_opt.port)
else:
visualizer = None
# set imitator
imitator = Imitator(test_opt)
if test_opt.post_tune:
adaptive_personalize(test_opt, imitator, visualizer)
imitator.personalize(test_opt.src_path, visualizer=visualizer)
print('\n\t\t\tPersonalization: completed...')
if test_opt.save_res:
pred_output_dir = mkdir(os.path.join(test_opt.output_dir, 'imitators'))
pred_output_dir = clear_dir(pred_output_dir)
else:
pred_output_dir = None
print('\n\t\t\tImitating `{}`'.format(test_opt.tgt_path))
tgt_paths = scan_tgt_paths(test_opt.tgt_path, itv=1)
imitator.inference(tgt_paths,
tgt_smpls=None,
cam_strategy='smooth',
output_dir=pred_output_dir,
visualizer=visualizer,
verbose=True)
def fit(model, opt, dataloaders, steps_states, data_params, loggers):
# read data_params
batch_size = data_params['batch_size']
virtual_batch_size = data_params['virtual_batch_size']
total_iters = data_params['total_iters']
total_epochs = data_params['total_epochs']
# read steps_states
start_epoch = steps_states["start_epoch"]
current_step = steps_states["current_step"]
virtual_step = steps_states["virtual_step"]
# read loggers
logger = util.get_root_logger()
tb_logger = loggers["tb_logger"]
# training
logger.info('Start training from epoch: {:d}, iter: {:d}'.format(start_epoch, current_step))
try:
timer = metrics.Timer() # iteration timer
timerData = metrics.TickTock() # data timer
timerEpoch = metrics.TickTock() # epoch timer
# outer loop for different epochs
for epoch in range(start_epoch, (total_epochs * (virtual_batch_size // batch_size))+1):
timerData.tick()
timerEpoch.tick()
# inner iteration loop within one epoch
for n, train_data in enumerate(dataloaders['train'], start=1):
timerData.tock()
virtual_step += 1
take_step = False
if virtual_step > 0 and virtual_step * batch_size % virtual_batch_size == 0:
current_step += 1
take_step = True
if current_step > total_iters:
break
# training
model.feed_data(train_data) # unpack data from dataset and apply preprocessing
model.optimize_parameters(virtual_step) # calculate loss functions, get gradients, update network weights
# log
def eta(t_iter):
# calculate training ETA in hours
return (t_iter * (opt['train']['niter'] - current_step)) / 3600 if t_iter > 0 else 0
if current_step % opt['logger']['print_freq'] == 0 and take_step:
# iteration end time
avg_time = timer.get_average_and_reset()
avg_data_time = timerData.get_average_and_reset()
# print training losses and save logging information to disk
message = '<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}, t:{:.4f}s, td:{:.4f}s, eta:{:.4f}h> '.format(
epoch, current_step, model.get_current_learning_rate(current_step), avg_time,
avg_data_time, eta(avg_time))
# tensorboard training logger
if opt['use_tb_logger'] and 'debug' not in opt['name']:
if current_step % opt['logger'].get('tb_sample_rate', 1) == 0: # Reduce rate of tb logs
# tb_logger.add_scalar('loss/nll', nll, current_step)
tb_logger.add_scalar('lr/base', model.get_current_learning_rate(), current_step)
tb_logger.add_scalar('time/iteration', timer.get_last_iteration(), current_step)
tb_logger.add_scalar('time/data', timerData.get_last_iteration(), current_step)
logs = model.get_current_log()
for k, v in logs.items():
message += '{:s}: {:.4e} '.format(k, v)
# tensorboard loss logger
if opt['use_tb_logger'] and 'debug' not in opt['name']:
if current_step % opt['logger'].get('tb_sample_rate', 1) == 0: # Reduce rate of tb logs
tb_logger.add_scalar(k, v, current_step)
# tb_logger.flush()
logger.info(message)
# start time for next iteration #TODO:skip the validation time from calculation
timer.tick()
# update learning rate
if model.optGstep and model.optDstep and take_step:
model.update_learning_rate(current_step, warmup_iter=opt['train'].get('warmup_iter', -1))
# save latest models and training states every <save_checkpoint_freq> iterations
if current_step % opt['logger']['save_checkpoint_freq'] == 0 and take_step:
if model.swa:
model.save(current_step, opt['logger']['overwrite_chkp'], loader=dataloaders['train'])
else:
model.save(current_step, opt['logger']['overwrite_chkp'])
model.save_training_state(
epoch=epoch + (n >= len(dataloaders['train'])),
iter_step=current_step,
latest=opt['logger']['overwrite_chkp']
)
logger.info('Models and training states saved.')
# validation
if dataloaders.get('val', None) and current_step % opt['train']['val_freq'] == 0 and take_step:
val_metrics = metrics.MetricsDict(metrics=opt['train'].get('metrics', None))
nlls = []
for val_data in dataloaders['val']:
model.feed_data(val_data) # unpack data from data loader
model.test() # run inference
if hasattr(model, 'nll'):
nll = model.nll if model.nll else 0
nlls.append(nll)
"""
Get Visuals
"""
visuals = model.get_current_visuals() # get image results
img_name = os.path.splitext(os.path.basename(val_data['LR_path'][0]))[0]
img_dir = os.path.join(opt['path']['val_images'], img_name)
util.mkdir(img_dir)
# Save SR images for reference
sr_img = None
if hasattr(model, 'heats'): # SRFlow
opt['train']['val_comparison'] = False
for heat in model.heats:
for i in range(model.n_sample):
sr_img = tensor2np(visuals['SR', heat, i], denormalize=opt['datasets']['train']['znorm'])
if opt['train']['overwrite_val_imgs']:
save_img_path = os.path.join(img_dir,
'{:s}_h{:03d}_s{:d}.png'.format(img_name, int(heat * 100), i))
else:
save_img_path = os.path.join(img_dir,
'{:s}_{:09d}_h{:03d}_s{:d}.png'.format(img_name,
current_step,
int(heat * 100), i))
util.save_img(sr_img, save_img_path)
else: # regular SR
sr_img = tensor2np(visuals['SR'], denormalize=opt['datasets']['train']['znorm'])
if opt['train']['overwrite_val_imgs']:
save_img_path = os.path.join(img_dir, '{:s}.png'.format(img_name))
else:
save_img_path = os.path.join(img_dir, '{:s}_{:d}.png'.format(img_name, current_step))
if not opt['train']['val_comparison']:
util.save_img(sr_img, save_img_path)
assert sr_img is not None
# Save GT images for reference
gt_img = tensor2np(visuals['HR'], denormalize=opt['datasets']['train']['znorm'])
if opt['train']['save_gt']:
save_img_path_gt = os.path.join(img_dir,
'{:s}_GT.png'.format(img_name))
if not os.path.isfile(save_img_path_gt):
util.save_img(gt_img, save_img_path_gt)
# Save LQ images for reference
if opt['train']['save_lr']:
save_img_path_lq = os.path.join(img_dir,
'{:s}_LQ.png'.format(img_name))
if not os.path.isfile(save_img_path_lq):
lq_img = tensor2np(visuals['LR'], denormalize=opt['datasets']['train']['znorm'])
util.save_img(lq_img, save_img_path_lq, scale=opt['scale'])
# save single images or LQ / SR comparison
if opt['train']['val_comparison']:
lr_img = tensor2np(visuals['LR'], denormalize=opt['datasets']['train']['znorm'])
util.save_img_comp([lr_img, sr_img], save_img_path)
# else:
# util.save_img(sr_img, save_img_path)
"""
Get Metrics
# TODO: test using tensor based metrics (batch) instead of numpy.
"""
val_metrics.calculate_metrics(sr_img, gt_img, crop_size=opt['scale']) # , only_y=True)
avg_metrics = val_metrics.get_averages()
if nlls:
avg_nll = sum(nlls) / len(nlls)
del val_metrics
# log
logger_m = ''
for r in avg_metrics:
formatted_res = r['name'].upper() + ': {:.5g}, '.format(r['average'])
logger_m += formatted_res
if nlls:
logger_m += 'avg_nll: {:.4e} '.format(avg_nll)
logger.info('# Validation # ' + logger_m[:-2])
logger_val = logging.getLogger('val') # validation logger
logger_val.info('<epoch:{:3d}, iter:{:8,d}> '.format(epoch, current_step) + logger_m[:-2])
# memory_usage = torch.cuda.memory_allocated()/(1024.0 ** 3) # in GB
# tensorboard logger
if opt['use_tb_logger'] and 'debug' not in opt['name']:
for r in avg_metrics:
tb_logger.add_scalar(r['name'], r['average'], current_step)
if nlls:
tb_logger.add_scalar('average nll', avg_nll, current_step)
# tb_logger.flush()
# tb_logger_valid.add_scalar(r['name'], r['average'], current_step)
# tb_logger_valid.flush()
timerData.tick()
timerEpoch.tock()
logger.info('End of epoch {} / {} \t Time Taken: {:.4f} sec'.format(
epoch, total_epochs, timerEpoch.get_last_iteration()))
logger.info('Saving the final model.')
if model.swa:
model.save('latest', loader=dataloaders['train'])
else:
model.save('latest')
logger.info('End of training.')
except KeyboardInterrupt:
# catch a KeyboardInterrupt and save the model and state to resume later
if model.swa:
model.save(current_step, True, loader=dataloaders['train'])
else:
model.save(current_step, True)
model.save_training_state(epoch + (n >= len(dataloaders['train'])), current_step, True)
logger.info('Training interrupted. Latest models and training states saved.')
def weights_best(self):
path = "weights/{}/best".format(self.cnf['name'])
mkdir(path)
return os.path.join(path, '{epoch}_{timestamp}_{loss}.pkl')
def main():
############################################
#
# set options
#
############################################
parser = argparse.ArgumentParser()
parser.add_argument('--opt', type=str, help='Path to option YAML file.')
parser.add_argument('--launcher',
choices=['none', 'pytorch'],
default='none',
help='job launcher')
parser.add_argument('--local_rank', type=int, default=0)
args = parser.parse_args()
opt = option.parse(args.opt, is_train=True)
############################################
#
# distributed training settings
#
############################################
if args.launcher == 'none': # disabled distributed training
opt['dist'] = False
rank = -1
print('Disabled distributed training.')
else:
opt['dist'] = True
init_dist()
world_size = torch.distributed.get_world_size()
rank = torch.distributed.get_rank()
print("Rank:", rank)
print("------------------DIST-------------------------")
############################################
#
# loading resume state if exists
#
############################################
if opt['path'].get('resume_state', None):
# distributed resuming: all load into default GPU
device_id = torch.cuda.current_device()
resume_state = torch.load(
opt['path']['resume_state'],
map_location=lambda storage, loc: storage.cuda(device_id))
option.check_resume(opt, resume_state['iter']) # check resume options
else:
resume_state = None
############################################
#
# mkdir and loggers
#
############################################
if rank <= 0: # normal training (rank -1) OR distributed training (rank 0)
if resume_state is None:
util.mkdir_and_rename(
opt['path']
['experiments_root']) # rename experiment folder if exists
util.mkdirs(
(path for key, path in opt['path'].items()
if not key == 'experiments_root'
and 'pretrain_model' not in key and 'resume' not in key))
# config loggers. Before it, the log will not work
util.setup_logger('base',
opt['path']['log'],
'train_' + opt['name'],
level=logging.INFO,
screen=True,
tofile=True)
util.setup_logger('base_val',
opt['path']['log'],
'val_' + opt['name'],
level=logging.INFO,
screen=True,
tofile=True)
logger = logging.getLogger('base')
logger_val = logging.getLogger('base_val')
logger.info(option.dict2str(opt))
# tensorboard logger
if opt['use_tb_logger'] and 'debug' not in opt['name']:
version = float(torch.__version__[0:3])
if version >= 1.1: # PyTorch 1.1
from torch.utils.tensorboard import SummaryWriter
else:
logger.info(
'You are using PyTorch {}. Tensorboard will use [tensorboardX]'
.format(version))
from tensorboardX import SummaryWriter
tb_logger = SummaryWriter(log_dir='../tb_logger/' + opt['name'])
else:
# config loggers. Before it, the log will not work
util.setup_logger('base',
opt['path']['log'],
'train_',
level=logging.INFO,
screen=True)
print("set train log")
util.setup_logger('base_val',
opt['path']['log'],
'val_',
level=logging.INFO,
screen=True)
print("set val log")
logger = logging.getLogger('base')
logger_val = logging.getLogger('base_val')
# convert to NoneDict, which returns None for missing keys
opt = option.dict_to_nonedict(opt)
#### random seed
seed = opt['train']['manual_seed']
if seed is None:
seed = random.randint(1, 10000)
if rank <= 0:
logger.info('Random seed: {}'.format(seed))
util.set_random_seed(seed)
torch.backends.cudnn.benchmark = True
# torch.backends.cudnn.deterministic = True
############################################
#
# create train and val dataloader
#
############################################
####
# dataset_ratio = 200 # enlarge the size of each epoch, todo: what it is
dataset_ratio = 1 # enlarge the size of each epoch, todo: what it is
for phase, dataset_opt in opt['datasets'].items():
if phase == 'train':
train_set = create_dataset(dataset_opt)
train_size = int(
math.ceil(len(train_set) / dataset_opt['batch_size']))
# total_iters = int(opt['train']['niter'])
# total_epochs = int(math.ceil(total_iters / train_size))
total_iters = train_size
total_epochs = int(opt['train']['epoch'])
if opt['dist']:
train_sampler = DistIterSampler(train_set, world_size, rank,
dataset_ratio)
# total_epochs = int(math.ceil(total_iters / (train_size * dataset_ratio)))
total_epochs = int(opt['train']['epoch'])
if opt['train']['enable'] == False:
total_epochs = 1
else:
train_sampler = None
train_loader = create_dataloader(train_set, dataset_opt, opt,
train_sampler)
if rank <= 0:
logger.info(
'Number of train images: {:,d}, iters: {:,d}'.format(
len(train_set), train_size))
logger.info('Total epochs needed: {:d} for iters {:,d}'.format(
total_epochs, total_iters))
elif phase == 'val':
val_set = create_dataset(dataset_opt)
val_loader = create_dataloader(val_set, dataset_opt, opt, None)
if rank <= 0:
logger.info('Number of val images in [{:s}]: {:d}'.format(
dataset_opt['name'], len(val_set)))
else:
raise NotImplementedError(
'Phase [{:s}] is not recognized.'.format(phase))
assert train_loader is not None
############################################
#
# create model
#
############################################
####
model = create_model(opt)
print("Model Created! ")
#### resume training
if resume_state:
logger.info('Resuming training from epoch: {}, iter: {}.'.format(
resume_state['epoch'], resume_state['iter']))
start_epoch = resume_state['epoch']
current_step = resume_state['iter']
model.resume_training(resume_state) # handle optimizers and schedulers
else:
current_step = 0
start_epoch = 0
print("Not Resume Training")
############################################
#
# training
#
############################################
####
####
logger.info('Start training from epoch: {:d}, iter: {:d}'.format(
start_epoch, current_step))
Avg_train_loss = AverageMeter() # total
if (opt['train']['pixel_criterion'] == 'cb+ssim'):
Avg_train_loss_pix = AverageMeter()
Avg_train_loss_ssim = AverageMeter()
elif (opt['train']['pixel_criterion'] == 'cb+ssim+vmaf'):
Avg_train_loss_pix = AverageMeter()
Avg_train_loss_ssim = AverageMeter()
Avg_train_loss_vmaf = AverageMeter()
elif (opt['train']['pixel_criterion'] == 'ssim'):
Avg_train_loss_ssim = AverageMeter()
elif (opt['train']['pixel_criterion'] == 'msssim'):
Avg_train_loss_msssim = AverageMeter()
elif (opt['train']['pixel_criterion'] == 'cb+msssim'):
Avg_train_loss_pix = AverageMeter()
Avg_train_loss_msssim = AverageMeter()
saved_total_loss = 10e10
saved_total_PSNR = -1
for epoch in range(start_epoch, total_epochs):
############################################
#
# Start a new epoch
#
############################################
# Turn into training mode
#model = model.train()
# reset total loss
Avg_train_loss.reset()
current_step = 0
if (opt['train']['pixel_criterion'] == 'cb+ssim'):
Avg_train_loss_pix.reset()
Avg_train_loss_ssim.reset()
elif (opt['train']['pixel_criterion'] == 'cb+ssim+vmaf'):
Avg_train_loss_pix.reset()
Avg_train_loss_ssim.reset()
Avg_train_loss_vmaf.reset()
elif (opt['train']['pixel_criterion'] == 'ssim'):
Avg_train_loss_ssim = AverageMeter()
elif (opt['train']['pixel_criterion'] == 'msssim'):
Avg_train_loss_msssim = AverageMeter()
elif (opt['train']['pixel_criterion'] == 'cb+msssim'):
Avg_train_loss_pix = AverageMeter()
Avg_train_loss_msssim = AverageMeter()
if opt['dist']:
train_sampler.set_epoch(epoch)
for train_idx, train_data in enumerate(train_loader):
if 'debug' in opt['name']:
img_dir = os.path.join(opt['path']['train_images'])
util.mkdir(img_dir)
LQ = train_data['LQs']
GT = train_data['GT']
GT_img = util.tensor2img(GT) # uint8
save_img_path = os.path.join(
img_dir, '{:4d}_{:s}.png'.format(train_idx, 'debug_GT'))
util.save_img(GT_img, save_img_path)
for i in range(5):
LQ_img = util.tensor2img(LQ[0, i, ...]) # uint8
save_img_path = os.path.join(
img_dir,
'{:4d}_{:s}_{:1d}.png'.format(train_idx, 'debug_LQ',
i))
util.save_img(LQ_img, save_img_path)
if (train_idx >= 3):
break
if opt['train']['enable'] == False:
message_train_loss = 'None'
break
current_step += 1
if current_step > total_iters:
print("Total Iteration Reached !")
break
#### update learning rate
if opt['train']['lr_scheme'] == 'ReduceLROnPlateau':
pass
else:
model.update_learning_rate(
current_step, warmup_iter=opt['train']['warmup_iter'])
#### training
model.feed_data(train_data)
# if opt['train']['lr_scheme'] == 'ReduceLROnPlateau':
# model.optimize_parameters_without_schudlue(current_step)
# else:
model.optimize_parameters(current_step)
if (opt['train']['pixel_criterion'] == 'cb+ssim'):
Avg_train_loss.update(model.log_dict['total_loss'], 1)
Avg_train_loss_pix.update(model.log_dict['l_pix'], 1)
Avg_train_loss_ssim.update(model.log_dict['ssim_loss'], 1)
elif (opt['train']['pixel_criterion'] == 'cb+ssim+vmaf'):
Avg_train_loss.update(model.log_dict['total_loss'], 1)
Avg_train_loss_pix.update(model.log_dict['l_pix'], 1)
Avg_train_loss_ssim.update(model.log_dict['ssim_loss'], 1)
Avg_train_loss_vmaf.update(model.log_dict['vmaf_loss'], 1)
elif (opt['train']['pixel_criterion'] == 'ssim'):
Avg_train_loss.update(model.log_dict['total_loss'], 1)
Avg_train_loss_ssim.update(model.log_dict['ssim_loss'], 1)
elif (opt['train']['pixel_criterion'] == 'msssim'):
Avg_train_loss.update(model.log_dict['total_loss'], 1)
Avg_train_loss_msssim.update(model.log_dict['msssim_loss'], 1)
elif (opt['train']['pixel_criterion'] == 'cb+msssim'):
Avg_train_loss.update(model.log_dict['total_loss'], 1)
Avg_train_loss_pix.update(model.log_dict['l_pix'], 1)
Avg_train_loss_msssim.update(model.log_dict['msssim_loss'], 1)
else:
Avg_train_loss.update(model.log_dict['l_pix'], 1)
# add total train loss
if (opt['train']['pixel_criterion'] == 'cb+ssim'):
message_train_loss = ' pix_avg_loss: {:.4e}'.format(
Avg_train_loss_pix.avg)
message_train_loss += ' ssim_avg_loss: {:.4e}'.format(
Avg_train_loss_ssim.avg)
message_train_loss += ' total_avg_loss: {:.4e}'.format(
Avg_train_loss.avg)
elif (opt['train']['pixel_criterion'] == 'cb+ssim+vmaf'):
message_train_loss = ' pix_avg_loss: {:.4e}'.format(
Avg_train_loss_pix.avg)
message_train_loss += ' ssim_avg_loss: {:.4e}'.format(
Avg_train_loss_ssim.avg)
message_train_loss += ' vmaf_avg_loss: {:.4e}'.format(
Avg_train_loss_vmaf.avg)
message_train_loss += ' total_avg_loss: {:.4e}'.format(
Avg_train_loss.avg)
elif (opt['train']['pixel_criterion'] == 'ssim'):
message_train_loss = ' ssim_avg_loss: {:.4e}'.format(
Avg_train_loss_ssim.avg)
message_train_loss += ' total_avg_loss: {:.4e}'.format(
Avg_train_loss.avg)
elif (opt['train']['pixel_criterion'] == 'msssim'):
message_train_loss = ' msssim_avg_loss: {:.4e}'.format(
Avg_train_loss_msssim.avg)
message_train_loss += ' total_avg_loss: {:.4e}'.format(
Avg_train_loss.avg)
elif (opt['train']['pixel_criterion'] == 'cb+msssim'):
message_train_loss = ' pix_avg_loss: {:.4e}'.format(
Avg_train_loss_pix.avg)
message_train_loss += ' msssim_avg_loss: {:.4e}'.format(
Avg_train_loss_msssim.avg)
message_train_loss += ' total_avg_loss: {:.4e}'.format(
Avg_train_loss.avg)
else:
message_train_loss = ' train_avg_loss: {:.4e}'.format(
Avg_train_loss.avg)
#### log
if current_step % opt['logger']['print_freq'] == 0:
logs = model.get_current_log()
message = '[epoch:{:3d}, iter:{:8,d}, lr:('.format(
epoch, current_step)
for v in model.get_current_learning_rate():
message += '{:.3e},'.format(v)
message += ')] '
for k, v in logs.items():
message += '{:s}: {:.4e} '.format(k, v)
# tensorboard logger
if opt['use_tb_logger'] and 'debug' not in opt['name']:
if rank <= 0:
tb_logger.add_scalar(k, v, current_step)
message += message_train_loss
if rank <= 0:
logger.info(message)
############################################
#
# end of one epoch, save epoch model
#
############################################
#### save models and training states
# if current_step % opt['logger']['save_checkpoint_freq'] == 0:
# if rank <= 0:
# logger.info('Saving models and training states.')
# model.save(current_step)
# model.save('latest')
# # model.save_training_state(epoch, current_step)
# # todo delete previous weights
# previous_step = current_step - opt['logger']['save_checkpoint_freq']
# save_filename = '{}_{}.pth'.format(previous_step, 'G')
# save_path = os.path.join(opt['path']['models'], save_filename)
# if os.path.exists(save_path):
# os.remove(save_path)
if epoch == 1:
save_filename = '{:04d}_{}.pth'.format(0, 'G')
save_path = os.path.join(opt['path']['models'], save_filename)
if os.path.exists(save_path):
os.remove(save_path)
save_filename = '{:04d}_{}.pth'.format(epoch - 1, 'G')
save_path = os.path.join(opt['path']['models'], save_filename)
if os.path.exists(save_path):
os.remove(save_path)
if rank <= 0:
logger.info('Saving models and training states.')
save_filename = '{:04d}'.format(epoch)
model.save(save_filename)
# model.save('latest')
# model.save_training_state(epoch, current_step)
############################################
#
# end of one epoch, do validation
#
############################################
#### validation
#if opt['datasets'].get('val', None) and current_step % opt['train']['val_freq'] == 0:
if opt['datasets'].get('val', None):
if opt['model'] in [
'sr', 'srgan'
] and rank <= 0: # image restoration validation
# does not support multi-GPU validation
pbar = util.ProgressBar(len(val_loader))
avg_psnr = 0.
idx = 0
for val_data in val_loader:
idx += 1
img_name = os.path.splitext(
os.path.basename(val_data['LQ_path'][0]))[0]
img_dir = os.path.join(opt['path']['val_images'], img_name)
util.mkdir(img_dir)
model.feed_data(val_data)
model.test()
visuals = model.get_current_visuals()
sr_img = util.tensor2img(visuals['rlt']) # uint8
gt_img = util.tensor2img(visuals['GT']) # uint8
# Save SR images for reference
save_img_path = os.path.join(
img_dir,
'{:s}_{:d}.png'.format(img_name, current_step))
#util.save_img(sr_img, save_img_path)
# calculate PSNR
sr_img, gt_img = util.crop_border([sr_img, gt_img],
opt['scale'])
avg_psnr += util.calculate_psnr(sr_img, gt_img)
pbar.update('Test {}'.format(img_name))
avg_psnr = avg_psnr / idx
# log
logger.info('# Validation # PSNR: {:.4e}'.format(avg_psnr))
# tensorboard logger
if opt['use_tb_logger'] and 'debug' not in opt['name']:
tb_logger.add_scalar('psnr', avg_psnr, current_step)
else: # video restoration validation
if opt['dist']:
# todo : multi-GPU testing
psnr_rlt = {} # with border and center frames
psnr_rlt_avg = {}
psnr_total_avg = 0.
ssim_rlt = {} # with border and center frames
ssim_rlt_avg = {}
ssim_total_avg = 0.
val_loss_rlt = {}
val_loss_rlt_avg = {}
val_loss_total_avg = 0.
if rank == 0:
pbar = util.ProgressBar(len(val_set))
for idx in range(rank, len(val_set), world_size):
print('idx', idx)
if 'debug' in opt['name']:
if (idx >= 3):
break
val_data = val_set[idx]
val_data['LQs'].unsqueeze_(0)
val_data['GT'].unsqueeze_(0)
folder = val_data['folder']
idx_d, max_idx = val_data['idx'].split('/')
idx_d, max_idx = int(idx_d), int(max_idx)
if psnr_rlt.get(folder, None) is None:
psnr_rlt[folder] = torch.zeros(max_idx,
dtype=torch.float32,
device='cuda')
if ssim_rlt.get(folder, None) is None:
ssim_rlt[folder] = torch.zeros(max_idx,
dtype=torch.float32,
device='cuda')
if val_loss_rlt.get(folder, None) is None:
val_loss_rlt[folder] = torch.zeros(
max_idx, dtype=torch.float32, device='cuda')
# tmp = torch.zeros(max_idx, dtype=torch.float32, device='cuda')
model.feed_data(val_data)
# model.test()
# model.test_stitch()
if opt['stitch'] == True:
model.test_stitch()
else:
model.test() # large GPU memory
# visuals = model.get_current_visuals()
visuals = model.get_current_visuals(
save=True,
name='{}_{}'.format(folder, idx),
save_path=opt['path']['val_images'])
rlt_img = util.tensor2img(visuals['rlt']) # uint8
gt_img = util.tensor2img(visuals['GT']) # uint8
# calculate PSNR
psnr = util.calculate_psnr(rlt_img, gt_img)
psnr_rlt[folder][idx_d] = psnr
# calculate SSIM
ssim = util.calculate_ssim(rlt_img, gt_img)
ssim_rlt[folder][idx_d] = ssim
# calculate Val loss
val_loss = model.get_loss()
val_loss_rlt[folder][idx_d] = val_loss
logger.info(
'{}_{:02d} PSNR: {:.4f}, SSIM: {:.4f}'.format(
folder, idx, psnr, ssim))
if rank == 0:
for _ in range(world_size):
pbar.update('Test {} - {}/{}'.format(
folder, idx_d, max_idx))
# # collect data
for _, v in psnr_rlt.items():
dist.reduce(v, 0)
for _, v in ssim_rlt.items():
dist.reduce(v, 0)
for _, v in val_loss_rlt.items():
dist.reduce(v, 0)
dist.barrier()
if rank == 0:
psnr_rlt_avg = {}
psnr_total_avg = 0.
for k, v in psnr_rlt.items():
psnr_rlt_avg[k] = torch.mean(v).cpu().item()
psnr_total_avg += psnr_rlt_avg[k]
psnr_total_avg /= len(psnr_rlt)
log_s = '# Validation # PSNR: {:.4e}:'.format(
psnr_total_avg)
for k, v in psnr_rlt_avg.items():
log_s += ' {}: {:.4e}'.format(k, v)
logger.info(log_s)
# ssim
ssim_rlt_avg = {}
ssim_total_avg = 0.
for k, v in ssim_rlt.items():
ssim_rlt_avg[k] = torch.mean(v).cpu().item()
ssim_total_avg += ssim_rlt_avg[k]
ssim_total_avg /= len(ssim_rlt)
log_s = '# Validation # PSNR: {:.4e}:'.format(
ssim_total_avg)
for k, v in ssim_rlt_avg.items():
log_s += ' {}: {:.4e}'.format(k, v)
logger.info(log_s)
# added
val_loss_rlt_avg = {}
val_loss_total_avg = 0.
for k, v in val_loss_rlt.items():
val_loss_rlt_avg[k] = torch.mean(v).cpu().item()
val_loss_total_avg += val_loss_rlt_avg[k]
val_loss_total_avg /= len(val_loss_rlt)
log_l = '# Validation # Loss: {:.4e}:'.format(
val_loss_total_avg)
for k, v in val_loss_rlt_avg.items():
log_l += ' {}: {:.4e}'.format(k, v)
logger.info(log_l)
message = ''
for v in model.get_current_learning_rate():
message += '{:.5e}'.format(v)
logger_val.info(
'Epoch {:02d}, LR {:s}, PSNR {:.4f}, SSIM {:.4f} Train {:s}, Val Total Loss {:.4e}'
.format(epoch, message, psnr_total_avg,
ssim_total_avg, message_train_loss,
val_loss_total_avg))
if opt['use_tb_logger'] and 'debug' not in opt['name']:
tb_logger.add_scalar('psnr_avg', psnr_total_avg,
current_step)
for k, v in psnr_rlt_avg.items():
tb_logger.add_scalar(k, v, current_step)
# add val loss
tb_logger.add_scalar('val_loss_avg',
val_loss_total_avg,
current_step)
for k, v in val_loss_rlt_avg.items():
tb_logger.add_scalar(k, v, current_step)
else: # Todo: our function One GPU
pbar = util.ProgressBar(len(val_loader))
psnr_rlt = {} # with border and center frames
psnr_rlt_avg = {}
psnr_total_avg = 0.
ssim_rlt = {} # with border and center frames
ssim_rlt_avg = {}
ssim_total_avg = 0.
val_loss_rlt = {}
val_loss_rlt_avg = {}
val_loss_total_avg = 0.
for val_inx, val_data in enumerate(val_loader):
if 'debug' in opt['name']:
if (val_inx >= 5):
break
folder = val_data['folder'][0]
# idx_d = val_data['idx'].item()
idx_d = val_data['idx']
# border = val_data['border'].item()
if psnr_rlt.get(folder, None) is None:
psnr_rlt[folder] = []
if ssim_rlt.get(folder, None) is None:
ssim_rlt[folder] = []
if val_loss_rlt.get(folder, None) is None:
val_loss_rlt[folder] = []
# process the black blank [B N C H W]
print(val_data['LQs'].size())
H_S = val_data['LQs'].size(3) # 540
W_S = val_data['LQs'].size(4) # 960
print(H_S)
print(W_S)
blank_1_S = 0
blank_2_S = 0
print(val_data['LQs'][0, 2, 0, :, :].size())
for i in range(H_S):
if not sum(val_data['LQs'][0, 2, 0, i, :]) == 0:
blank_1_S = i - 1
# assert not sum(data_S[:, :, 0][i+1]) == 0
break
for i in range(H_S):
if not sum(val_data['LQs'][0, 2, 0, :,
H_S - i - 1]) == 0:
blank_2_S = (H_S - 1) - i - 1
# assert not sum(data_S[:, :, 0][blank_2_S-1]) == 0
break
print('LQ :', blank_1_S, blank_2_S)
if blank_1_S == -1:
print('LQ has no blank')
blank_1_S = 0
blank_2_S = H_S
# val_data['LQs'] = val_data['LQs'][:,:,:,blank_1_S:blank_2_S,:]
print("LQ", val_data['LQs'].size())
# end of process the black blank
model.feed_data(val_data)
if opt['stitch'] == True:
model.test_stitch()
else:
model.test() # large GPU memory
# process blank
blank_1_L = blank_1_S << 2
blank_2_L = blank_2_S << 2
print(blank_1_L, blank_2_L)
print(model.fake_H.size())
if not blank_1_S == 0:
# model.fake_H = model.fake_H[:,:,blank_1_L:blank_2_L,:]
model.fake_H[:, :, 0:blank_1_L, :] = 0
model.fake_H[:, :, blank_2_L:H_S, :] = 0
# end of # process blank
visuals = model.get_current_visuals(
save=True,
name='{}_{:02d}'.format(folder, val_inx),
save_path=opt['path']['val_images'])
rlt_img = util.tensor2img(visuals['rlt']) # uint8
gt_img = util.tensor2img(visuals['GT']) # uint8
# calculate PSNR
psnr = util.calculate_psnr(rlt_img, gt_img)
psnr_rlt[folder].append(psnr)
# calculate SSIM
ssim = util.calculate_ssim(rlt_img, gt_img)
ssim_rlt[folder].append(ssim)
# val loss
val_loss = model.get_loss()
val_loss_rlt[folder].append(val_loss.item())
logger.info(
'{}_{:02d} PSNR: {:.4f}, SSIM: {:.4f}'.format(
folder, val_inx, psnr, ssim))
pbar.update('Test {} - {}'.format(folder, idx_d))
# average PSNR
for k, v in psnr_rlt.items():
psnr_rlt_avg[k] = sum(v) / len(v)
psnr_total_avg += psnr_rlt_avg[k]
psnr_total_avg /= len(psnr_rlt)
log_s = '# Validation # PSNR: {:.4e}:'.format(
psnr_total_avg)
for k, v in psnr_rlt_avg.items():
log_s += ' {}: {:.4e}'.format(k, v)
logger.info(log_s)
# average SSIM
for k, v in ssim_rlt.items():
ssim_rlt_avg[k] = sum(v) / len(v)
ssim_total_avg += ssim_rlt_avg[k]
ssim_total_avg /= len(ssim_rlt)
log_s = '# Validation # SSIM: {:.4e}:'.format(
ssim_total_avg)
for k, v in ssim_rlt_avg.items():
log_s += ' {}: {:.4e}'.format(k, v)
logger.info(log_s)
# average VMAF
# average Val LOSS
for k, v in val_loss_rlt.items():
val_loss_rlt_avg[k] = sum(v) / len(v)
val_loss_total_avg += val_loss_rlt_avg[k]
val_loss_total_avg /= len(val_loss_rlt)
log_l = '# Validation # Loss: {:.4e}:'.format(
val_loss_total_avg)
for k, v in val_loss_rlt_avg.items():
log_l += ' {}: {:.4e}'.format(k, v)
logger.info(log_l)
# toal validation log
message = ''
for v in model.get_current_learning_rate():
message += '{:.5e}'.format(v)
logger_val.info(
'Epoch {:02d}, LR {:s}, PSNR {:.4f}, SSIM {:.4f} Train {:s}, Val Total Loss {:.4e}'
.format(epoch, message, psnr_total_avg, ssim_total_avg,
message_train_loss, val_loss_total_avg))
# end add
if opt['use_tb_logger'] and 'debug' not in opt['name']:
tb_logger.add_scalar('psnr_avg', psnr_total_avg,
current_step)
for k, v in psnr_rlt_avg.items():
tb_logger.add_scalar(k, v, current_step)
# tb_logger.add_scalar('ssim_avg', ssim_total_avg, current_step)
# for k, v in ssim_rlt_avg.items():
# tb_logger.add_scalar(k, v, current_step)
# add val loss
tb_logger.add_scalar('val_loss_avg',
val_loss_total_avg, current_step)
for k, v in val_loss_rlt_avg.items():
tb_logger.add_scalar(k, v, current_step)
############################################
#
# end of validation, save model
#
############################################
#
logger.info("Finished an epoch, Check and Save the model weights")
# we check the validation loss instead of training loss. OK~
if saved_total_loss >= val_loss_total_avg:
saved_total_loss = val_loss_total_avg
#torch.save(model.state_dict(), args.save_path + "/best" + ".pth")
model.save('best')
logger.info(
"Best Weights updated for decreased validation loss")
else:
logger.info(
"Weights Not updated for undecreased validation loss")
if saved_total_PSNR <= psnr_total_avg:
saved_total_PSNR = psnr_total_avg
model.save('bestPSNR')
logger.info(
"Best Weights updated for increased validation PSNR")
else:
logger.info(
"Weights Not updated for unincreased validation PSNR")
############################################
#
# end of one epoch, schedule LR
#
############################################
# add scheduler todo
if opt['train']['lr_scheme'] == 'ReduceLROnPlateau':
for scheduler in model.schedulers:
# scheduler.step(val_loss_total_avg)
scheduler.step(val_loss_total_avg)
if rank <= 0:
logger.info('Saving the final model.')
model.save('last')
logger.info('End of training.')
tb_logger.close()
def psnr_main(opt,
train_loader,
val_loader,
train_sampler,
logger,
resume_state=None,
tb_logger=None,
rank=-1):
# create model
model = create_model(opt)
# resume training
if resume_state:
logger.info('Resuming training from epoch: {}, iter: {}.'.format(
resume_state['epoch'], resume_state['iter']))
start_epoch = resume_state['epoch']
current_step = resume_state['iter']
model.resume_training(resume_state) # handle optimizers and schedulers
else:
current_step = 0
start_epoch = 0
logger.info('Start training from epoch: {:d}, iter: {:d}'.format(
start_epoch, current_step))
total_epochs = int(opt['train']['nepochs'])
best_psnr = 0
patience = 0
all_results = []
for epoch in range(start_epoch, total_epochs):
if opt['dist']:
train_sampler.set_epoch(epoch)
for batch_num, train_data in enumerate(train_loader):
current_step += 1
# training
model.feed_data(train_data)
model.optimize_parameters(current_step)
progress_bar(batch_num, len(train_loader), msg=None)
# log
if epoch % opt['logger']['print_freq'] == 0:
logs = model.get_current_log()
message = '<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> '.format(
epoch, current_step, model.get_current_learning_rate())
for k, v in logs.items():
message += '{:s}: {:.4e} '.format(k, v)
# tensorboard logger
if opt['use_tb_logger'] and 'debug' not in opt['name']:
if rank <= 0:
tb_logger.add_scalar(k, v, current_step)
if rank <= 0:
logger.info(message)
# batched validation
if epoch % opt['train']['val_freq'] == 0 and rank <= 0:
avg_psnr = 0.0
idx = 0
for batch_num, val_data in enumerate(val_loader):
img_name = os.path.splitext(
os.path.basename(val_data['LQ_path'][0]))[0]
img_dir = os.path.join(opt['path']['val_images'], img_name)
util.mkdir(img_dir)
model.feed_data(val_data)
model.test()
# calculate PSNR
item_psnr = util.tensor_psnr(model.real_H, model.fake_H)
if math.isfinite(item_psnr):
avg_psnr += item_psnr
idx += 1
progress_bar(batch_num, len(val_loader), msg=None)
avg_psnr = avg_psnr / idx
all_results.append(avg_psnr)
if avg_psnr < best_psnr:
patience += 1
if patience == opt['train']['epoch_patience']:
model.update_learning_rate(opt['train']['lr_decay'])
print(
"no improvement, final patience, updating learning rate to {}"
.format(model.get_current_learning_rate()))
patience = 0
else:
print("no improvement, patience {} out of {}".format(
patience, opt['train']['epoch_patience']))
if model.get_current_learning_rate() < opt['train']['min_lr']:
break
else:
best_psnr = avg_psnr
if rank <= 0:
logger.info('Saving models and training states.')
model.save('latest')
# log
logger.info('# Validation # PSNR: {:.4e}'.format(avg_psnr))
logger_val = logging.getLogger('val') # validation logger
logger_val.info(
'<epoch:{:3d}, iter:{:8,d}> psnr: {:.4e} (best: {:.4e})'.
format(epoch, current_step, avg_psnr, best_psnr))
# tensorboard logger
if opt['use_tb_logger'] and 'debug' not in opt['name']:
tb_logger.add_scalar('psnr', avg_psnr, current_step)
if rank <= 0:
logger.info('End of training.')
json.dump(all_results,
open(
os.path.join(opt['path']['log'],
'validation_results.json'), 'w'),
indent=2)
def main():
#### options
parser = argparse.ArgumentParser()
parser.add_argument('-opt',
type=str,
default='options/test/test_KPSAGAN.yml',
help='Path to option YMAL file.')
parser.add_argument('--launcher',
choices=['none', 'pytorch'],
default='none',
help='job launcher')
parser.add_argument('--local_rank', type=int, default=0)
args = parser.parse_args()
opt = option.parse(args.opt, is_train=False)
#### distributed training settings
if args.launcher == 'none': # disabled distributed training
opt['dist'] = False
rank = -1
print('Disabled distributed training.')
else:
opt['dist'] = True
init_dist()
world_size = torch.distributed.get_world_size()
rank = torch.distributed.get_rank()
#### loading resume state if exists
if opt['path'].get('resume_state', None):
# distributed resuming: all load into default GPU
device_id = torch.cuda.current_device()
resume_state = torch.load(
opt['path']['resume_state'],
map_location=lambda storage, loc: storage.cuda(device_id))
option.check_resume(opt, resume_state['iter']) # check resume options
else:
resume_state = None
#### mkdir and loggers
if rank <= 0: # normal training (rank -1) OR distributed training (rank 0)
if resume_state is None:
# util.mkdir_and_rename(
#opt['path']['experiments_root']) # rename experiment folder if exists
util.mkdirs(
(path for key, path in opt['path'].items()
if not key == 'experiments_root'
and 'pretrain_model' not in key and 'resume' not in key))
# config loggers. Before it, the log will not work
util.setup_logger('base',
opt['path']['log'],
'train_' + opt['name'],
level=logging.INFO,
screen=True,
tofile=True)
util.setup_logger('val',
opt['path']['log'],
'val_' + opt['name'],
level=logging.INFO,
screen=True,
tofile=True)
logger = logging.getLogger('base')
logger.info(option.dict2str(opt))
else:
util.setup_logger('base',
opt['path']['log'],
'train',
level=logging.INFO,
screen=True)
logger = logging.getLogger('base')
# convert to NoneDict, which returns None for missing keys
opt = option.dict_to_nonedict(opt)
torch.backends.cudnn.benckmark = True
# torch.backends.cudnn.deterministic = True
#### create train and val dataloader
dataset_ratio = 200 # enlarge the size of each epoch
for phase, dataset_opt in opt['datasets'].items():
val_set = create_dataset(dataset_opt)
val_loader = create_dataloader(val_set, dataset_opt, opt, None)
if rank <= 0:
logger.info('Number of val images in [{:s}]: {:d}'.format(
dataset_opt['name'], len(val_set)))
#### create model
model = create_model(opt)
avg_psnr = 0.0
idx = 0
dataset_dir = '/srv/wuyichao/Super-Resolution/KPSAGAN/BasicSR-master/BasicSR-master-c/result_600000/'
util.mkdir(dataset_dir)
for val_data in val_loader:
idx += 1
img_name = os.path.splitext(os.path.basename(
val_data['LQ_path'][0]))[0]
logger.info(img_name)
#img_dir = os.path.join(opt['path']['val_images'], img_name)
#util.mkdir(img_dir)
model.feed_data(val_data)
model.test()
visuals = model.get_current_visuals()
sr_img = util.tensor2img(visuals['SR']) # uint8
gt_img = util.tensor2img(visuals['GT']) # uint8
# save images
suffix = 'cut'
#opt['suffix']
if suffix:
save_img_path = osp.join(dataset_dir, img_name + suffix + '.png')
else:
save_img_path = osp.join(dataset_dir, img_name + '.png')
util.save_img(sr_img, save_img_path)
# calculate PSNR
crop_size = opt['scale']
gt_img = gt_img / 255.
sr_img = sr_img / 255.
cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size, :]
cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size, :]
avg_psnr += util.calculate_psnr(cropped_sr_img * 255,
cropped_gt_img * 255)
avg_psnr = avg_psnr / idx
# log
logger.info('# Validation # PSNR: {:.4e}'.format(avg_psnr))
logger_val = logging.getLogger('val') # validation logger
logger_val.info('psnr: {:.4e}'.format(avg_psnr))
def final_weights_file(self):
path = "weights/{}".format(self.cnf['name'])
mkdir(path)
return os.path.join(path, 'weights_final.pkl')
def retrain_weights_file(self):
path = "weights/{}/retrain".format(self.cnf['name'])
mkdir(path)
return os.path.join(path, 'weights.pkl')
def main():
dataset = 'REDS' # REDS | Vimeo90K | DIV2K800_sub
opt = {}
opt['dist'] = False
opt['gpu_ids'] = [0]
if dataset == 'REDS':
opt['name'] = 'test_REDS'
opt['dataroot_GT'] = '../../datasets/REDS/train_sharp_wval.lmdb'
opt['dataroot_LQ'] = '../../datasets/REDS/train_sharp_bicubic_wval.lmdb'
opt['mode'] = 'REDS'
opt['N_frames'] = 5
opt['phase'] = 'train'
opt['use_shuffle'] = True
opt['n_workers'] = 8
opt['batch_size'] = 16
opt['GT_size'] = 256
opt['LQ_size'] = 64
opt['scale'] = 4
opt['use_flip'] = True
opt['use_rot'] = True
opt['interval_list'] = [1]
opt['random_reverse'] = False
opt['border_mode'] = False
opt['cache_keys'] = None
opt['data_type'] = 'lmdb' # img | lmdb | mc
elif dataset == 'Vimeo90K':
opt['name'] = 'test_Vimeo90K'
opt['dataroot_GT'] = '../../datasets/vimeo90k/vimeo90k_train_GT.lmdb'
opt['dataroot_LQ'] = '../../datasets/vimeo90k/vimeo90k_train_LR7frames.lmdb'
opt['mode'] = 'Vimeo90K'
opt['N_frames'] = 7
opt['phase'] = 'train'
opt['use_shuffle'] = True
opt['n_workers'] = 8
opt['batch_size'] = 16
opt['GT_size'] = 256
opt['LQ_size'] = 64
opt['scale'] = 4
opt['use_flip'] = True
opt['use_rot'] = True
opt['interval_list'] = [1]
opt['random_reverse'] = False
opt['border_mode'] = False
opt['cache_keys'] = None
opt['data_type'] = 'lmdb' # img | lmdb | mc
elif dataset == 'DIV2K800_sub':
opt['name'] = 'DIV2K800'
opt['dataroot_GT'] = '../../datasets/DIV2K/DIV2K800_sub.lmdb'
opt['dataroot_LQ'] = '../../datasets/DIV2K/DIV2K800_sub_bicLRx4.lmdb'
opt['mode'] = 'LQGT'
opt['phase'] = 'train'
opt['use_shuffle'] = True
opt['n_workers'] = 8
opt['batch_size'] = 16
opt['GT_size'] = 128
opt['scale'] = 4
opt['use_flip'] = True
opt['use_rot'] = True
opt['color'] = 'RGB'
opt['data_type'] = 'lmdb' # img | lmdb
else:
raise ValueError('Please implement by yourself.')
util.mkdir('tmp')
train_set = create_dataset(opt)
train_loader = create_dataloader(train_set, opt, opt, None)
nrow = int(math.sqrt(opt['batch_size']))
padding = 2 if opt['phase'] == 'train' else 0
print('start...')
for i, data in enumerate(train_loader):
if i > 5:
break
print(i)
if dataset == 'REDS' or dataset == 'Vimeo90K':
LQs = data['LQs']
else:
LQ = data['LQ']
GT = data['GT']
if dataset == 'REDS' or dataset == 'Vimeo90K':
for j in range(LQs.size(1)):
torchvision.utils.save_image(LQs[:, j, :, :, :],
'tmp/LQ_{:03d}_{}.png'.format(i, j), nrow=nrow,
padding=padding, normalize=False)
else:
torchvision.utils.save_image(LQ, 'tmp/LQ_{:03d}.png'.format(i), nrow=nrow,
padding=padding, normalize=False)
torchvision.utils.save_image(GT, 'tmp/GT_{:03d}.png'.format(i), nrow=nrow, padding=padding,
normalize=False)
def train_dir(self): return ut.mkdir(pj(self.resdir, 'training'))
def main(config_path, trained_model_path):
restore_epoch = None # if None, restore the final epoch
# Read a config file (.yml)
with open(config_path, "r") as f:
config = yaml.load(f)
corpus = config['corpus']
feature = config['feature']
param = config['param']
if corpus['label_type'] == 'phone':
if corpus['social_signal_type'] in ['insert', 'insert3']:
output_size = 41
elif corpus['social_signal_type'] == 'insert2':
output_size = 44
elif corpus['social_signal_type'] == 'remove':
output_size = 38
elif corpus['label_type'] == 'character':
if corpus['social_signal_type'] in ['insert', 'insert3']:
output_size = 150
elif corpus['social_signal_type'] == 'insert2':
output_size = 153
elif corpus['social_signal_type'] == 'remove':
output_size = 147
# Load model
CTCModel = load(model_type=config['model_name'])
network = CTCModel(batch_size=param['batch_size'],
input_size=feature['input_size'] * feature['num_stack'],
num_cell=param['num_cell'],
num_layer=param['num_layer'],
output_size=output_size,
clip_gradients=param['clip_grad'],
clip_activation=param['clip_activation'],
dropout_ratio_input=param['dropout_input'],
dropout_ratio_hidden=param['dropout_hidden'],
num_proj=param['num_proj'],
weight_decay=param['weight_decay'])
network.model_name = config['model_name'].upper()
network.model_name += '_' + str(param['num_cell'])
network.model_name += '_' + str(param['num_layer'])
network.model_name += '_' + param['optimizer']
network.model_name += '_lr' + str(param['learning_rate'])
if feature['num_stack'] != 1:
network.model_name += '_stack' + str(feature['num_stack'])
network.model_name += '_transfer_' + corpus['transfer_data_size']
# Set save path
network.model_dir = mkdir('/n/sd8/inaguma/result/csj/dialog/')
network.model_dir = join(network.model_dir, 'ctc')
network.model_dir = join(network.model_dir, corpus['label_type'])
network.model_dir = join(network.model_dir, corpus['social_signal_type'])
network.model_dir = join(network.model_dir, network.model_name)
# Reset model directory
if not os.path.isfile(os.path.join(network.model_dir, 'complete.txt')):
tf.gfile.DeleteRecursively(network.model_dir)
tf.gfile.MakeDirs(network.model_dir)
else:
raise ValueError('File exists.')
# Set process name
setproctitle('ctc_csj_dialog_' + corpus['label_type'] + '_' +
param['optimizer'] + '_' + corpus['social_signal_type'] +
'_transfer_' + corpus['transfer_data_size'])
# Save config file
shutil.copyfile(config_path, os.path.join(network.model_dir, 'config.yml'))
sys.stdout = open(os.path.join(network.model_dir, 'train.log'), 'w')
print(network.model_name)
do_fine_tune(network=network,
optimizer=param['optimizer'],
learning_rate=param['learning_rate'],
batch_size=param['batch_size'],
epoch_num=param['num_epoch'],
label_type=corpus['label_type'],
num_stack=feature['num_stack'],
num_skip=feature['num_skip'],
social_signal_type=corpus['social_signal_type'],
trained_model_path=trained_model_path,
restore_epoch=restore_epoch)
sys.stdout = sys.__stdout__
from datetime import datetime
import pprint
import os
import argparse
import numpy as np
import torchvision.models as models
from utils.util import mkdir
from data import FEATURE_DIR
mkdir(FEATURE_DIR)
class Config(object):
def __init__(self, layers, cnf=None):
self.layers = layersGlobalPoolLayer
self.cnf = cnf
pprint.pprint(cnf)
def get(self, k, default=None):
return self.cnf.get(k, default)
@property
def weights_epoch(self):
path = "weights/{}/epochs".format(self.cnf['name'])
mkdir(path)
return os.path.join(path, '{epoch}_{timestamp}_{loss}.pkl')
@property
def weights_best(self):
path = "weights/{}/best".format(self.cnf['name'])
mkdir(path)