def train(gpu, config):
dist.init_process_group(backend='nccl',
init_method='env://',
world_size=config['num_gpus'],
rank=gpu)
torch.cuda.set_device(gpu)
"""
@ build the dataset for training
"""
dataset = get_data(config)
trainset = dataset(config, "train")
testset = dataset(config, "test")
sampler_train = DistributedSampler(trainset,
num_replicas=config['num_gpus'],
rank=gpu)
sampler_val = DistributedSampler(testset,
num_replicas=config['num_gpus'],
rank=gpu)
batch_size = config['batch_size']
loader_train = DataLoader(dataset=trainset,
batch_size=batch_size,
shuffle=False,
num_workers=config['num_threads'],
pin_memory=True,
sampler=sampler_train,
drop_last=True)
loader_val = DataLoader(dataset=testset,
batch_size=1,
shuffle=False,
num_workers=1,
pin_memory=True,
sampler=sampler_val,
drop_last=True)
model = UNet(config["in_channels"],
config["out_channels"],
post_processing=True)
model.cuda(gpu)
mask_sampling = masksamplingv2()
""" @ init parameter
"""
save_folder = os.path.join(
config['save_root'], 'batch_{}_lr_{}'.format(config['batch_size'],
config['lr']))
best_epoch = 0
lowest_loss = 0.
resume = 0
print('=>Save folder: {}\n'.format(save_folder))
if not os.path.exists(save_folder):
os.makedirs(save_folder)
optimizer = define_optim(config['optimizer'], model.parameters(),
float(config['lr']), 0)
criterion_1 = define_loss(config['loss_type'])
criterion_2 = define_loss("Multimse")
scheduler = define_scheduler(optimizer, config)
"""
@ justify the resume model
"""
if config['resume'] != 'None':
checkpoint = torch.load(config['resume'],
map_location=torch.device('cpu'))
resume = checkpoint['epoch']
lowest_loss = checkpoint['loss']
best_epoch = checkpoint['best_epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['scheduler'])
model, optimizer = amp.initialize(model,
optimizer,
opt_level='O0',
verbosity=0)
amp.load_state_dict(checkpoint['amp'])
print("=> loaded checkpoint '{}' (epoch {})".format(
resume, checkpoint['epoch']))
del checkpoint
log_file = 'log_train_start_{}.txt'.format(resume)
"""
@ convert model to multi-gpus modes for training
"""
model = apex.parallel.convert_syncbn_model(model)
if config['resume'] == 'None':
model, optimizer = amp.initialize(model,
optimizer,
opt_level='O0',
verbosity=0)
model = DDP(model)
if gpu == 0:
sys.stdout = Logger(os.path.join(save_folder, log_file))
print("Number of parameters in model is {:.3f}M".format(
sum(tensor.numel() for tensor in model.parameters()) / 1e6))
"""
@ start to train
"""
for epoch in range(resume + 1, config['epoches'] + 1):
print('=> Starch Epoch {}\n'.format(epoch))
print(datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
print('learning rate is set to {}.\n'.format(
optimizer.param_groups[0]['lr']))
model.train()
sampler_train.set_epoch(epoch)
batch_time = AverageMeter()
losses = AverageMeter()
metric_train = Metrics()
rmse_train = AverageMeter()
mae_train = AverageMeter()
time_snap = time.time()
for i, inputs in tqdm(enumerate(loader_train)):
gt, noise = inputs['gt'].cuda(gpu), inputs['noise'].cuda(gpu)
optimizer.zero_grad()
""" update the train inputs
"""
# patten = np.random.randint(0, 4, 1)
patten = torch.randint(0, 8, (1, ))
redinput, blueinput = mask_sampling(noise, patten)
# redinput, blueinput = generator(noise, mask1, mask2)
output = model(redinput)
loss = criterion_1(output, blueinput)
fulloutput = model(noise)
redoutput, blueoutput = mask_sampling(fulloutput, patten)
# redoutput, blueoutput = generator(fulloutput, mask1, mask2)
loss2 = criterion_2(output, blueinput, redoutput, blueoutput)
losssum = config["gamma"] * loss2 + loss
with amp.scale_loss(losssum, optimizer) as scaled_loss:
scaled_loss.backward()
optimizer.step()
"@ map-reduce tensor"
rt = reduce_tensor(losssum.data)
torch.cuda.synchronize()
losses.update(rt.item(), loader_train.batch_size)
metric_train.calculate(fulloutput.detach(), gt)
rmse_train.update(metric_train.get_metric('mse'), metric_train.num)
mae_train.update(metric_train.get_metric('mae'), metric_train.num)
batch_time.update(time.time() - time_snap)
time_snap = time.time()
if (i + 1) % config['print_freq'] == 0:
if gpu == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.6f} ({loss.avg:.6f})\t'
'Metric {rmse_train.val:.6f} ({rmse_train.avg:.6f})'.
format(epoch,
i + 1,
len(loader_train),
batch_time=batch_time,
loss=losses,
rmse_train=rmse_train))
if (i + 1) % config['save_freq'] == 0:
print('=> Start sub-selection validation set')
rmse, mae = val(model, loader_val, epoch, gpu)
model.train()
if gpu == 0:
print("===> Average RMSE score on selection set is {:.6f}".
format(rmse))
print("===> Average MAE score on selection set is {:.6f}".
format(mae))
print(
"===> Last best score was RMSE of {:.6f} in epoch {}".
format(lowest_loss, best_epoch))
if rmse > lowest_loss:
lowest_loss = rmse
best_epoch = epoch
states = {
'epoch': epoch,
'best_epoch': best_epoch,
'loss': lowest_loss,
'state_dict': model.module.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'amp': amp.state_dict()
}
save_checkpoints(states, save_folder, epoch, gpu, True)
# save checkpoints
print('=> Start selection validation set')
rmse, mae = val(model, loader_val, epoch, gpu)
model.train()
if gpu == 0:
print("===> Average RMSE score on selection set is {:.6f}".format(
rmse))
print("===> Average MAE score on selection set is {:.6f}".format(
mae))
print("===> Last best score was RMSE of {:.6f} in epoch {}".format(
lowest_loss, best_epoch))
if rmse > lowest_loss:
best_epoch = epoch
lowest_loss = rmse
states = {
'epoch': epoch,
'best_epoch': best_epoch,
'loss': lowest_loss,
'state_dict': model.module.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'amp': amp.state_dict()
}
save_checkpoints(states, save_folder, epoch, gpu, True)
if config['lr_policy'] == 'plateau':
scheduler.step(rmse)
else:
scheduler.step()
# if (epoch) % 10 == 0:
# config["gamma"] += 0.5
print('=>> the model training finish!')
def main():
global args
args = parser.parse_args()
if args.num_samples == 0:
args.num_samples = None
if args.val_batch_size is None:
args.val_batch_size = args.batch_size
if args.seed:
random.seed(args.seed)
torch.manual_seed(args.seed)
# torch.backends.cudnn.deterministic = True
# warnings.warn('You have chosen to seed training. '
# 'This will turn on the CUDNN deterministic setting, '
# 'which can slow down your training considerably! '
# 'You may see unexpected behavior when restarting from checkpoints.')
# For distributed training
# init_distributed_mode(args)
if not args.no_cuda and not torch.cuda.is_available():
raise Exception("No gpu available for usage")
torch.backends.cudnn.benchmark = args.cudnn
# Init model
channels_in = 1 if args.input_type == 'depth' else 4
model = Models.define_model(mod=args.mod,
in_channels=channels_in,
thres=args.thres)
define_init_weights(model, args.weight_init)
# Load on gpu before passing params to optimizer
if not args.no_cuda:
if not args.multi:
model = model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
# model.cuda()
# model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
# model = model.module
save_id = '{}_{}_{}_{}_{}_batch{}_pretrain{}_wlid{}_wrgb{}_wguide{}_wpred{}_patience{}_num_samples{}_multi{}'.\
format(args.mod, args.optimizer, args.loss_criterion,
args.learning_rate,
args.input_type,
args.batch_size,
args.pretrained, args.wlid, args.wrgb, args.wguide, args.wpred,
args.lr_decay_iters, args.num_samples, args.multi)
# INIT optimizer/scheduler/loss criterion
optimizer = define_optim(args.optimizer, model.parameters(),
args.learning_rate, args.weight_decay)
scheduler = define_scheduler(optimizer, args)
# Optional to use different losses
criterion_local = define_loss(args.loss_criterion)
criterion_lidar = define_loss(args.loss_criterion)
criterion_rgb = define_loss(args.loss_criterion)
criterion_guide = define_loss(args.loss_criterion)
# INIT dataset
dataset = Datasets.define_dataset(args.dataset, args.data_path,
args.input_type, args.side_selection)
dataset.prepare_dataset()
train_loader, valid_loader, valid_selection_loader = get_loader(
args, dataset)
# Resume training
best_epoch = 0
lowest_loss = np.inf
args.save_path = os.path.join(args.save_path, save_id)
mkdir_if_missing(args.save_path)
log_file_name = 'log_train_start_0.txt'
args.resume = first_run(args.save_path)
if args.resume and not args.test_mode and not args.evaluate:
path = os.path.join(
args.save_path,
'checkpoint_model_epoch_{}.pth.tar'.format(int(args.resume)))
if os.path.isfile(path):
log_file_name = 'log_train_start_{}.txt'.format(args.resume)
# stdout
sys.stdout = Logger(os.path.join(args.save_path, log_file_name))
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(path)
args.start_epoch = checkpoint['epoch']
lowest_loss = checkpoint['loss']
best_epoch = checkpoint['best epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
log_file_name = 'log_train_start_0.txt'
# stdout
sys.stdout = Logger(os.path.join(args.save_path, log_file_name))
print("=> no checkpoint found at '{}'".format(path))
# Only evaluate
elif args.evaluate:
print("Evaluate only")
best_file_lst = glob.glob(os.path.join(args.save_path, 'model_best*'))
if len(best_file_lst) != 0:
best_file_name = best_file_lst[0]
print(best_file_name)
if os.path.isfile(best_file_name):
sys.stdout = Logger(
os.path.join(args.save_path, 'Evaluate.txt'))
print("=> loading checkpoint '{}'".format(best_file_name))
checkpoint = torch.load(best_file_name)
model.load_state_dict(checkpoint['state_dict'])
else:
print("=> no checkpoint found at '{}'".format(best_file_name))
else:
print("=> no checkpoint found at due to empy list in folder {}".
format(args.save_path))
validate(valid_selection_loader, model, criterion_lidar, criterion_rgb,
criterion_local, criterion_guide)
return
# Start training from clean slate
else:
# Redirect stdout
sys.stdout = Logger(os.path.join(args.save_path, log_file_name))
# INIT MODEL
print(40 * "=" + "\nArgs:{}\n".format(args) + 40 * "=")
print("Init model: '{}'".format(args.mod))
print("Number of parameters in model {} is {:.3f}M".format(
args.mod.upper(),
sum(tensor.numel() for tensor in model.parameters()) / 1e6))
# Load pretrained state for cityscapes in GLOBAL net
if args.pretrained and not args.resume:
if not args.load_external_mod:
if not args.multi:
target_state = model.depthnet.state_dict()
else:
target_state = model.module.depthnet.state_dict()
check = torch.load('erfnet_pretrained.pth')
for name, val in check.items():
# Exclude multi GPU prefix
mono_name = name[7:]
if mono_name not in target_state:
continue
try:
target_state[mono_name].copy_(val)
except RuntimeError:
continue
print('Successfully loaded pretrained model')
else:
check = torch.load('external_mod.pth.tar')
lowest_loss_load = check['loss']
target_state = model.state_dict()
for name, val in check['state_dict'].items():
if name not in target_state:
continue
try:
target_state[name].copy_(val)
except RuntimeError:
continue
print("=> loaded EXTERNAL checkpoint with best rmse {}".format(
lowest_loss_load))
# Start training
for epoch in range(args.start_epoch, args.nepochs):
print("\n => Start EPOCH {}".format(epoch + 1))
print(datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
print(args.save_path)
# Adjust learning rate
if args.lr_policy is not None and args.lr_policy != 'plateau':
scheduler.step()
lr = optimizer.param_groups[0]['lr']
print('lr is set to {}'.format(lr))
# Define container objects
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
score_train = AverageMeter()
score_train_1 = AverageMeter()
metric_train = Metrics(max_depth=args.max_depth,
disp=args.use_disp,
normal=args.normal)
# Train model for args.nepochs
model.train()
# compute timing
end = time.time()
# Load dataset
for i, (input, gt) in tqdm(enumerate(train_loader)):
# Time dataloader
data_time.update(time.time() - end)
# Put inputs on gpu if possible
if not args.no_cuda:
input, gt = input.cuda(), gt.cuda()
prediction, lidar_out, precise, guide = model(input, epoch)
loss = criterion_local(prediction, gt)
loss_lidar = criterion_lidar(lidar_out, gt)
loss_rgb = criterion_rgb(precise, gt)
loss_guide = criterion_guide(guide, gt)
loss = args.wpred * loss + args.wlid * loss_lidar + args.wrgb * loss_rgb + args.wguide * loss_guide
losses.update(loss.item(), input.size(0))
metric_train.calculate(prediction[:, 0:1].detach(), gt.detach())
score_train.update(metric_train.get_metric(args.metric),
metric_train.num)
score_train_1.update(metric_train.get_metric(args.metric_1),
metric_train.num)
# Clip gradients (usefull for instabilities or mistakes in ground truth)
if args.clip_grad_norm != 0:
nn.utils.clip_grad_norm(model.parameters(),
args.clip_grad_norm)
# Setup backward pass
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Time trainig iteration
batch_time.update(time.time() - end)
end = time.time()
# Print info
if (i + 1) % args.print_freq == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Metric {score.val:.4f} ({score.avg:.4f})'.format(
epoch + 1,
i + 1,
len(train_loader),
batch_time=batch_time,
loss=losses,
score=score_train))
print("===> Average RMSE score on training set is {:.4f}".format(
score_train.avg))
print("===> Average MAE score on training set is {:.4f}".format(
score_train_1.avg))
# Evaulate model on validation set
print("=> Start validation set")
score_valid, score_valid_1, losses_valid = validate(
valid_loader, model, criterion_lidar, criterion_rgb,
criterion_local, criterion_guide, epoch)
print("===> Average RMSE score on validation set is {:.4f}".format(
score_valid))
print("===> Average MAE score on validation set is {:.4f}".format(
score_valid_1))
# Evaluate model on selected validation set
if args.subset is None:
print("=> Start selection validation set")
score_selection, score_selection_1, losses_selection = validate(
valid_selection_loader, model, criterion_lidar, criterion_rgb,
criterion_local, criterion_guide, epoch)
total_score = score_selection
print("===> Average RMSE score on selection set is {:.4f}".format(
score_selection))
print("===> Average MAE score on selection set is {:.4f}".format(
score_selection_1))
else:
total_score = score_valid
print("===> Last best score was RMSE of {:.4f} in epoch {}".format(
lowest_loss, best_epoch))
# Adjust lr if loss plateaued
if args.lr_policy == 'plateau':
scheduler.step(total_score)
lr = optimizer.param_groups[0]['lr']
print('LR plateaued, hence is set to {}'.format(lr))
# File to keep latest epoch
with open(os.path.join(args.save_path, 'first_run.txt'), 'w') as f:
f.write(str(epoch))
# Save model
to_save = False
if total_score < lowest_loss:
to_save = True
best_epoch = epoch + 1
lowest_loss = total_score
save_checkpoint(
{
'epoch': epoch + 1,
'best epoch': best_epoch,
'arch': args.mod,
'state_dict': model.state_dict(),
'loss': lowest_loss,
'optimizer': optimizer.state_dict()
}, to_save, epoch)
if not args.no_tb:
writer.close()
def main():
global args
args = parser.parse_args()
if args.num_samples == 0:
# Use all lidar points
args.num_samples = None
else:
args.data_path = "" # path to precomputed 500 samples
assert args.num_samples == 500
print("changed path to samples500 dataset")
if args.val_batch_size is None:
args.val_batch_size = args.batch_size
if args.seed:
random.seed(args.seed)
torch.manual_seed(args.seed)
# init_distributed_mode(args)
if not args.no_cuda and not torch.cuda.is_available():
raise Exception("No gpu available for usage")
torch.backends.cudnn.benchmark = args.cudnn
# Init model
args.channels_in = 3 if args.input_type == 'rgb' else 4
model = Models.define_model(args.mod, args)
# define_init_weights(model, args.weight_init)
# Load on gpu before passing params to optimizer
if not args.no_cuda:
if not args.multi:
model = model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
# model.cuda()
# model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
# model = model.module
save_id = '{}_{}_{}_{}_{}_batch{}_pretrain{}_wrgb{}_drop{}_patience{}_num_samples{}_multi{}_submod{}'.\
format(args.mod, args.optimizer, args.loss_criterion,
args.learning_rate,
args.input_type,
args.batch_size,
args.pretrained, args.wrgb, args.drop,
args.lr_decay_iters, args.num_samples, args.multi, args.submod)
# INIT optimizer/scheduler/loss criterion
optimizer = define_optim(args.optimizer, model.parameters(),
args.learning_rate, args.weight_decay)
scheduler = define_scheduler(optimizer, args)
# Optional to use different losses
criterion_local = define_loss(args.loss_criterion)
criterion_rgb = define_loss(args.loss_criterion)
# INIT dataset
dataset = Datasets.define_dataset(args.dataset, args.data_path,
args.input_type)
dataset.prepare_dataset()
train_loader, _, valid_loader, valid_selection_loader = get_loader(
args, dataset)
# Resume training
best_epoch = 0
lowest_loss = np.inf
args.save_path = os.path.join(args.save_path, save_id)
mkdir_if_missing(args.save_path)
log_file_name = 'log_train_start_0.txt'
args.resume = first_run(args.save_path)
if args.resume and not args.test_mode and not args.evaluate:
path = os.path.join(
args.save_path,
'checkpoint_model_epoch_{}.pth.tar'.format(int(args.resume)))
if os.path.isfile(path):
log_file_name = 'log_train_start_{}.txt'.format(args.resume)
# stdout
sys.stdout = Logger(os.path.join(args.save_path, log_file_name))
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(path)
args.start_epoch = checkpoint['epoch']
lowest_loss = checkpoint['loss']
best_epoch = checkpoint['best epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
log_file_name = 'log_train_start_0.txt'
# stdout
sys.stdout = Logger(os.path.join(args.save_path, log_file_name))
print("=> no checkpoint found at '{}'".format(path))
# Only evaluate
elif args.evaluate:
print("Evaluate only")
best_file_lst = glob.glob(os.path.join(args.save_path, 'model_best*'))
if len(best_file_lst) != 0:
best_file_name = best_file_lst[0]
print(best_file_name)
if os.path.isfile(best_file_name):
sys.stdout = Logger(
os.path.join(args.save_path, 'Evaluate.txt'))
print("=> loading checkpoint '{}'".format(best_file_name))
checkpoint = torch.load(best_file_name)
model.load_state_dict(checkpoint['state_dict'])
else:
print("=> no checkpoint found at '{}'".format(best_file_name))
else:
print("=> no checkpoint found at due to empy list in folder {}".
format(args.save_path))
validate(valid_selection_loader, model, criterion_global,
criterion_local)
return
# Start training from clean slate
else:
# Redirect stdout
sys.stdout = Logger(os.path.join(args.save_path, log_file_name))
# INIT MODEL
print(40 * "=" + "\nArgs:{}\n".format(args) + 40 * "=")
print("Init model: '{}'".format(args.mod))
print("Number of parameters in model {} is {:.3f}M".format(
args.mod.upper(),
sum(tensor.numel() for tensor in model.parameters()) / 1e6))
# Load pretrained state for cityscapes in GLOBAL net
# if args.pretrained and not args.resume:
# if not args.multi:
# target_state = model.depthnet.state_dict()
# else:
# target_state = model.module.depthnet.state_dict()
# check = torch.load('erfnet_pretrained.pth')
# for name, val in check.items():
# # Exclude multi GPU prefix
# mono_name = name[7:]
# if mono_name not in target_state:
# continue
# try:
# target_state[mono_name].copy_(val)
# except RuntimeError:
# continue
# print('Successfully loaded pretrained model')
# Create summary writer
log_path = os.path.join(args.save_path, "logs")
if not os.path.exists(log_path):
os.makedirs(log_path)
logger = SummaryWriter(log_path)
with open(os.path.join(args.save_path, 'commandline_args.txt'), 'w') as f:
json.dump(args.__dict__, f, indent=2)
# Start training
for epoch in range(args.start_epoch, args.nepochs):
print("\n => Start EPOCH {}".format(epoch + 1))
print(datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
print(args.save_path)
# Adjust learning rate
if args.lr_policy is not None and args.lr_policy != 'plateau':
scheduler.step()
lr = optimizer.param_groups[0]['lr']
print('lr is set to {}'.format(lr))
# Define container objects
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
score_train = AverageMeter()
score_train_1 = AverageMeter()
metric_train = Metrics(max_depth=args.max_depth, disp=args.use_disp)
# Train model for args.nepochs
model.train()
# compute timing
end = time.time()
# Load dataset
for i, (input, gt) in tqdm(enumerate(train_loader)):
# Time dataloader
data_time.update(time.time() - end)
# Put inputs on gpu if possible
if not args.no_cuda:
input, gt = input.cuda(), gt.cuda()
prediction, hidden = model(input, hidden=(None, None))
if 'mod' in args.mod or 'stacked' in args.mod:
loss = criterion_local(prediction[0], gt)
loss_rgb = criterion_rgb(prediction[1], gt)
loss += args.wrgb * loss_rgb
prediction = prediction[0]
else:
loss = criterion_local(prediction, gt)
losses.update(loss.item(), input.size(0))
metric_train.calculate(prediction[:, 0:1].detach(), gt.detach())
score_train.update(metric_train.get_metric(args.metric),
metric_train.num)
score_train_1.update(metric_train.get_metric(args.metric_1),
metric_train.num)
# Clip gradients (usefull for instabilities or mistakes in ground truth)
if args.clip_grad_norm != 0:
nn.utils.clip_grad_norm(model.parameters(),
args.clip_grad_norm)
# Setup backward pass
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Time trainig iteration
batch_time.update(time.time() - end)
end = time.time()
# Print info
if (i + 1) % args.print_freq == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Metric {score.val:.4f} ({score.avg:.4f})'.format(
epoch + 1,
i + 1,
len(train_loader),
batch_time=batch_time,
loss=losses,
score=score_train))
batch_num = len(train_loader)
if ((i + 1) % 10 == 0) and (logger is not None):
current_step = epoch * batch_num + i
# Add scalar summaries
logger.add_scalar('Train_loss/Loss', loss.item(), current_step)
# utils.record_scalar_summary(result, average_meter, current_step, logger, "Train")
# # Add system info
# logger.add_scalar('System/gpu_time', average_meter.average().gpu_time, current_step)
# logger.add_scalar('System/data_time', average_meter.average().data_time, current_step)
if ((i + 1) % 200 == 0):
# Add some image summary
if args.modality == "rgb":
input_images = input.cpu()
else:
input_images = input[:, :3, :, :].cpu()
input_depth = torch.unsqueeze(input[:, 3, :, :],
dim=1).cpu()
rgb_grid = make_grid(input_images[0:6, :, :, :],
nrow=3,
normalize=True),
target_grid = make_grid(gt.cpu()[0:6, :, :, :],
nrow=3,
normalize=True)
pred_grid = make_grid(prediction.cpu()[0:6, :, :, :],
nrow=3,
normalize=True)
logger.add_image('Train/RGB', rgb_grid[0].data.numpy(),
current_step)
logger.add_image('Train/Depth_gt',
target_grid.data.numpy(), current_step)
logger.add_image('Train/Depth_pred',
pred_grid.data.numpy(), current_step)
if args.modality == "rgbd":
depth_grid = make_grid(input_depth[0:6, :, :, :],
nrow=3,
normalize=True)
logger.add_image('Train/Depth_input',
depth_grid.data.numpy(), current_step)
print("===> Average RMSE score on training set is {:.4f}".format(
score_train.avg))
print("===> Average MAE score on training set is {:.4f}".format(
score_train_1.avg))
# Evaulate model on validation set
print("=> Start validation set")
score_valid, score_valid_1, losses_valid = validate(
valid_loader, model, criterion_rgb, criterion_local, epoch)
print("===> Average RMSE score on validation set is {:.4f}".format(
score_valid))
print("===> Average MAE score on validation set is {:.4f}".format(
score_valid_1))
# Evaluate model on selected validation set
if args.subset is None:
print("=> Start selection validation set")
score_selection, score_selection_1, losses_selection = validate(
valid_selection_loader, model, criterion_rgb, criterion_local,
epoch)
total_score = score_selection
print("===> Average RMSE score on selection set is {:.4f}".format(
score_selection))
print("===> Average MAE score on selection set is {:.4f}".format(
score_selection_1))
else:
total_score = score_valid
print("===> Last best score was RMSE of {:.4f} in epoch {}".format(
lowest_loss, best_epoch))
# Adjust lr if loss plateaued
if args.lr_policy == 'plateau':
scheduler.step(total_score)
lr = optimizer.param_groups[0]['lr']
print('LR plateaued, hence is set to {}'.format(lr))
# File to keep latest epoch
with open(os.path.join(args.save_path, 'first_run.txt'), 'w') as f:
f.write(str(epoch))
# Save model
to_save = False
if total_score < lowest_loss:
to_save = True
best_epoch = epoch + 1
lowest_loss = total_score
save_checkpoint(
{
'epoch': epoch + 1,
'best epoch': best_epoch,
'arch': args.mod,
'state_dict': model.state_dict(),
'loss': lowest_loss,
'optimizer': optimizer.state_dict()
}, to_save, epoch)
if not args.no_tb:
writer.close()