def train(train_loader, pfld_backbone, auxiliarynet, criterion, optimizer,
cur_epoch):
losses = AverageMeter()
for img, landmark_gt, euler_angle_gt in train_loader:
img.requires_grad = False
img = img.cuda(non_blocking=True)
landmark_gt.requires_grad = False
landmark_gt = landmark_gt.cuda(non_blocking=True)
euler_angle_gt.requires_grad = False
euler_angle_gt = euler_angle_gt.cuda(non_blocking=True)
pfld_backbone = pfld_backbone.cuda()
auxiliarynet = auxiliarynet.cuda()
features, landmarks = pfld_backbone(img)
angle = auxiliarynet(features)
weighted_loss, loss = criterion(landmark_gt, euler_angle_gt, angle,
landmarks, args.train_batchsize)
optimizer.zero_grad()
weighted_loss.backward()
optimizer.step()
losses.update(loss.item())
return weighted_loss, loss
def train(train_loader, linear_backbone, criterion, optimizer, cur_epoch):
losses = AverageMeter()
for samples in train_loader:
img = samples['image']
landmark_gt = samples['landmarks']
img.requires_grad = False
img = img.cuda(non_blocking=True)
landmark_gt.requires_grad = False
landmark_gt = landmark_gt.cuda(non_blocking=True)
linear_backbone = linear_backbone.cuda()
landmarks = linear_backbone(img)
loss = criterion(landmark_gt, landmarks, args.train_batchsize)
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses.update(loss.item())
return loss
def validate(loader,
model,
criterion_lidar,
criterion_rgb,
criterion_local,
criterion_guide,
epoch=0):
# batch_time = AverageMeter()
losses = AverageMeter()
metric = Metrics(max_depth=args.max_depth,
disp=args.use_disp,
normal=args.normal)
score = AverageMeter()
score_1 = AverageMeter()
# Evaluate model
model.eval()
# Only forward pass, hence no grads needed
with torch.no_grad():
# end = time.time()
for i, (input, gt) in tqdm(enumerate(loader)):
if not args.no_cuda:
input, gt = input.cuda(non_blocking=True), gt.cuda(
non_blocking=True)
prediction, lidar_out, precise, guide = model(input, epoch)
loss = criterion_local(prediction, gt, epoch)
loss_lidar = criterion_lidar(lidar_out, gt, epoch)
loss_rgb = criterion_rgb(precise, gt, epoch)
loss_guide = criterion_guide(guide, gt, epoch)
loss = args.wpred * loss + args.wlid * loss_lidar + args.wrgb * loss_rgb + args.wguide * loss_guide
losses.update(loss.item(), input.size(0))
metric.calculate(prediction[:, 0:1], gt)
score.update(metric.get_metric(args.metric), metric.num)
score_1.update(metric.get_metric(args.metric_1), metric.num)
if (i + 1) % args.print_freq == 0:
print('Test: [{0}/{1}]\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Metric {score.val:.4f} ({score.avg:.4f})'.format(
i + 1, len(loader), loss=losses, score=score))
if args.evaluate:
print("===> Average RMSE score on validation set is {:.4f}".format(
score.avg))
print("===> Average MAE score on validation set is {:.4f}".format(
score_1.avg))
return score.avg, score_1.avg, losses.avg
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 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 test_epoch(self, epoch, data_loader, model, criterion, opt, logger):
print('test at epoch {}'.format(epoch))
model.eval()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
accuracies = AverageMeter()
end_time = time.time()
for i, (inputs, targets, _) in enumerate(data_loader):
data_time.update(time.time() - end_time)
if not opt.no_cuda:
targets = targets.cuda(non_blocking=True)
with torch.no_grad():
inputs = Variable(inputs)
targets = Variable(targets)
outputs = model(inputs)
loss = criterion(outputs, targets)
acc = calculate_accuracy(outputs, targets)
losses.update(loss.data, inputs.size(0))
accuracies.update(acc, inputs.size(0))
batch_time.update(time.time() - end_time)
end_time = time.time()
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc {acc.val:.3f} ({acc.avg:.3f})'.format(
epoch,
i + 1,
len(data_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=accuracies))
logger.log({
'epoch': epoch,
'loss': losses.avg,
'acc': accuracies.avg
})
return losses.avg
def val_epoch(self, epoch, data_loader, model, criterion, opt, logger):
print('validation at epoch {}'.format(epoch))
model.eval()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
accuracies = AverageMeter()
end_time = time.time()
confusion_matrix = np.zeros((opt.n_classes, opt.n_classes))
confidence_for_each_validation = {}
###########################################################################
# pdb.set_trace()
for i, (inputs, targets, paths) in enumerate(data_loader):
data_time.update(time.time() - end_time)
targets = targets.cuda(non_blocking=True)
with torch.no_grad():
inputs = Variable(inputs)
targets = Variable(targets)
outputs = model(inputs)
loss = criterion(outputs, targets)
acc = calculate_accuracy(outputs, targets)
######## temp line, needs to be removed##################################
for j in range(len(targets)):
key = paths[j].split('/')[-1]
confidence_for_each_validation[key] = [
x.item() for x in outputs[j]
]
rows = [int(x) for x in targets]
columns = [int(x) for x in np.argmax(outputs.data.cpu(), 1)]
assert len(rows) == len(columns)
for idx in range(len(rows)):
confusion_matrix[rows[idx]][columns[idx]] += 1
###########################################################################
losses.update(loss.data, inputs.size(0))
accuracies.update(acc, inputs.size(0))
batch_time.update(time.time() - end_time)
end_time = time.time()
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc {acc.val:.3f} ({acc.avg:.3f})'.format(
epoch,
i + 1,
len(data_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=accuracies))
######### temp line, needs to be removed##################################
# print(confusion_matrix)
confusion_matrix = pd.DataFrame(confusion_matrix)
# confusion_matrix.to_csv(file)
confidence_matrix = pd.DataFrame.from_dict(
confidence_for_each_validation, orient='index')
# confidence_matrix.to_csv('confidence_matrix.csv')
######### temp line, needs to be removed##################################
logger.log({'epoch': epoch, 'loss': losses.avg, 'acc': accuracies.avg})
return losses.avg, confusion_matrix, confidence_matrix
def train_epoch(self, epoch, data_loader, model, criterion, optimizer, opt,
epoch_logger, batch_logger):
print('train at epoch {}'.format(epoch))
model.train()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
accuracies = AverageMeter()
end_time = time.time()
for i, (inputs, targets, _) in enumerate(data_loader):
data_time.update(time.time() - end_time)
targets = targets.cuda(non_blocking=True)
inputs = Variable(inputs)
targets = Variable(targets)
outputs = model(inputs)
loss = criterion(outputs, targets)
acc = calculate_accuracy(outputs, targets)
losses.update(loss.data, inputs.size(0))
accuracies.update(acc, inputs.size(0))
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_time.update(time.time() - end_time)
end_time = time.time()
batch_logger.log({
'epoch': epoch,
'batch': i + 1,
'iter': (epoch - 1) * len(data_loader) + (i + 1),
'loss': losses.val,
'acc': accuracies.val,
'lr': optimizer.param_groups[0]['lr']
})
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc {acc.val:.3f} ({acc.avg:.3f})'.format(
epoch,
i + 1,
len(data_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=accuracies))
epoch_logger.log({
'epoch': epoch,
'loss': losses.avg,
'acc': accuracies.avg,
'lr': optimizer.param_groups[0]['lr']
})
if epoch % opt.checkpoint == 0:
save_file_path = os.path.join(opt.Results_directory,
'save_{}.pth'.format(epoch))
states = {
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
}
torch.save(states, save_file_path)
def validate(loader, model, criterion_rgb, criterion_local, epoch=0):
# batch_time = AverageMeter()
losses = AverageMeter()
metric = Metrics(max_depth=args.max_depth, disp=args.use_disp)
score = AverageMeter()
score_1 = AverageMeter()
loss_rgb = torch.zeros(1)
# Evaluate model
model.eval()
# Only forward pass, hence no grads needed
with torch.no_grad():
# end = time.time()
for i, (input, gt) in tqdm(enumerate(loader)):
if not args.no_cuda:
input, gt = input.cuda(non_blocking=True), gt.cuda(
non_blocking=True)
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.calculate(prediction[:, 0:1], gt)
score.update(metric.get_metric(args.metric), metric.num)
score_1.update(metric.get_metric(args.metric_1), metric.num)
# save 8 images for visualization
skip = 50
if args.modality == 'd':
img_merge = None
else:
if args.modality == 'rgb':
rgb = input
elif args.modality == 'rgbd':
rgb = input[:, :3, :, :]
depth = input[:, 3, :, :]
if i == 0:
if args.modality == 'rgbd':
img_merge = merge_into_row_with_gt(
rgb, depth, gt, prediction)
else:
img_merge = merge_into_row(rgb, gt, prediction)
elif (i < 8 * skip) and (i % skip == 0):
if args.modality == 'rgbd':
row = merge_into_row_with_gt(rgb, depth, gt,
prediction)
else:
row = merge_into_row(rgb, gt, prediction)
img_merge = add_row(img_merge, row)
elif i == 8 * skip:
filename = args.save_path + '/comparison_' + str(
epoch) + '.png'
save_image(img_merge, filename)
if (i + 1) % args.print_freq == 0:
print('Test: [{0}/{1}]\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Metric {score.val:.4f} ({score.avg:.4f})'.format(
i + 1, len(loader), loss=losses, score=score))
if args.evaluate:
print("===> Average RMSE score on validation set is {:.4f}".format(
score.avg))
print("===> Average MAE score on validation set is {:.4f}".format(
score_1.avg))
return score.avg, score_1.avg, losses.avg
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()
def main():
global args
args = parser.parse_args()
if args.num_samples == 0:
args.num_samples = None
if args.cuda and not torch.cuda.is_available():
raise Exception("No gpu available for usage")
# Init model
channels_in = 1 if args.input_type == 'depth' else 4
model = Models.define_model(mod=args.mod, in_channels=channels_in)
if args.mod == 'mod':
define_init_weights(model, args.weight_init)
# Load on gpu before passing params to optimizer
if args.cuda:
model = model.cuda()
save_id = '{}_{}_{}_{}_batch{}_pretrain{}_wlid{}_wrgb{}_wguide{}_wpred{}_num_samples{}'.\
format(args.mod, args.loss_criterion_source,
args.learning_rate,
args.input_type,
args.batch_size,
args.load_path!='', args.wlid, args.wrgb, args.wguide, args.wpred,
args.num_samples)
optimizer = torch.optim.Adam(model.parameters(),
lr=args.learning_rate,
weight_decay=args.weight_decay)
# Optional to use different losses
criterion_source = define_loss(args.loss_criterion_source)
criterion_target = define_loss(args.loss_criterion_target)
# INIT KITTI dataset
print('Load KITTI')
dataset = Datasets.define_dataset('kitti', args.data_path_target,
args.input_type)
dataset.prepare_dataset()
train_loader = get_loader(args, dataset, only_train=True)
# INIT Carla dataset
print('Load Carla')
dataset = Datasets.define_dataset('carla', args.data_path_source,
args.input_type)
dataset.prepare_dataset()
# The sparsification of the data and projection from the LiDAR reference
# frame to the RGB camera explained in the paper happens in the dataloader
train_loader_carla = get_loader(args,
dataset,
is_carla=True,
only_train=True)
train_loader_iter = iter(train_loader)
# Resume training
if args.save_name == '':
args.save_path = os.path.join(args.save_path, save_id)
else:
args.save_path = os.path.join(args.save_path, args.save_name)
if os.path.exists(args.save_path):
raise Exception('Save path already exists')
mkdir_if_missing(args.save_path)
# 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
if args.load_path != '':
print("=> loading checkpoint {:s}".format(args.load_path))
check = torch.load(
args.load_path,
map_location=lambda storage, loc: storage)['state_dict']
model.load_state_dict(check)
if args.use_image_translation:
image_trans_net = ResnetGeneratorCycle(3, 3, 64, n_blocks=9)
state_dict = torch.load('./image_translation_weights.pth')
image_trans_net.load_state_dict(state_dict)
image_trans_net.eval()
if args.cuda:
image_trans_net = image_trans_net.cuda()
# Start training
global_step = 0
for epoch in range(args.start_epoch, args.nepochs):
print("\n => Start EPOCH {}".format(epoch + 1))
# Define container objects
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
score_train_rmse = AverageMeter()
score_train_mae = AverageMeter()
metric_train = Metrics(max_depth=args.max_depth)
# Train model for args.nepochs
model.train()
# compute timing
end = time.time()
for i, (input, gt, filepath) in tqdm(enumerate(train_loader_carla)):
# Time dataloader
data_time.update(time.time() - end)
loss_extra = 0
# Put inputs on gpu if possible
if args.cuda:
input, gt = input.cuda(), gt.cuda()
# The LiDAR depths have large regions where no input depth is given
# We remove all of the GT in the synthetic data where no input information is given
# in a NxN window around the GT point (we set N=41) to avoid the model trying to estimate
# depth for areas without any input guidance
input_depth = input[:, 0:1]
input_depth, gt = filter_data(input_depth,
gt,
max_depth=args.max_depth)
input[:, 0:1] = input_depth
### Load target set (KITTI) data
if args.train_target:
try:
input_target, gt_target, filepath_t = next(
train_loader_iter)
except:
train_loader_iter = iter(train_loader)
input_target, gt_target, filepath_t = next(
train_loader_iter)
if args.cuda:
input_target, gt_target = input_target.cuda(
), gt_target.cuda()
if args.use_image_translation:
# The CycleGAN model was trained with inputs in the range of [-1, 1]
with torch.no_grad():
rgb_trans = image_trans_net(input[:, 1:] / 128.5 - 1)
rgb_trans = 128.5 * (rgb_trans + 1)
rgb_trans = rgb_trans.clamp(0, 255)
input = torch.cat([input[:, :1], rgb_trans], 1)
if args.train_target:
input_joint = torch.cat([input, input_target])
prediction, lidar_out, precise, guide = model(
input_joint, epoch)
# We separate predictions from the target domain and source domain
prediction_target, lidar_out_target, precise_target, guide_target = prediction[
args.batch_size:], lidar_out[args.batch_size:], precise[
args.batch_size:], guide[args.batch_size:]
prediction, lidar_out, precise, guide = prediction[:args.
batch_size], lidar_out[:
args
.
batch_size], precise[:
args
.
batch_size], guide[:
args
.
batch_size]
else:
prediction, lidar_out, precise, guide = model(input, epoch)
# We compute the loss for the source domain data
loss = criterion_source(prediction, gt)
loss_lidar = criterion_source(lidar_out, gt)
loss_rgb = criterion_source(precise, gt)
loss_guide = criterion_source(guide, gt)
loss = args.wpred * loss + args.wlid * loss_lidar + args.wrgb * loss_rgb + args.wguide * loss_guide
if args.train_target:
loss_target = 0
# We filter the input data for supervision as explained in the paper
filtered_sparse_data = filter_sparse_guidance(
input_target[:, :1], args.filter_window, args.filter_th)
# We compute the loss for the target domain data
loss_target += args.wpred * (criterion_target(
prediction_target, filtered_sparse_data))
loss_target += args.wlid * (criterion_target(
lidar_out_target, filtered_sparse_data))
loss_target += args.wrgb * (criterion_target(
precise_target, filtered_sparse_data))
loss_target += args.wguide * (criterion_target(
guide_target, filtered_sparse_data))
loss = loss + loss_target
metric_train.calculate(prediction[:, 0:1].detach(), gt.detach())
score_train_rmse.update(metric_train.get_metric('rmse'),
metric_train.num)
score_train_mae.update(metric_train.get_metric('mae'),
metric_train.num)
losses.update(loss.item(), input.size(0))
# Optimization step
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_time.update(time.time() - end)
end = time.time()
global_step += 1
# 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'
'RMSE Train {score.val:.4f} ({score.avg:.4f})'.format(
epoch + 1,
i + 1,
len(train_loader_carla),
batch_time=batch_time,
loss=losses,
score=score_train_rmse))
if global_step == args.n_training_iterations:
dict_save = {
'epoch': epoch + 1,
'arch': args.mod,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
}
save_checkpoint(dict_save, False, epoch + 1, global_step)
return 1
print("===> Average RMSE score on training set is {:.4f}".format(
score_train_rmse.avg))
print("===> Average MAE score on training set is {:.4f}".format(
score_train_mae.avg))
dict_save = {
'epoch': epoch + 1,
'arch': args.mod,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
}
save_checkpoint(dict_save, False, epoch + 1)
def eval_32bit(model, test_loader):
device = model.device
criterion = model.criterion
model.eval()
batch_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
end = time.time()
for i, data in enumerate(tqdm(test_loader)):
image = data[0].type(torch.FloatTensor).to(device)
label = data[1].type(torch.LongTensor).to(device)
pred_label = model(image)
loss = criterion(pred_label, label)
# measure accuracy and record loss
prec1, prec5 = accuracy(pred_label.data, label.data, topk=(1, 5))
losses.update(loss.item(), image.size(0))
top1.update(prec1.item(), image.size(0))
top5.update(prec5.item(), image.size(0))
# timing
batch_time.update(time.time() - end)
end = time.time()
print('Loss: {:.3f} | Acc1: {:.3f}% | Acc5: {:.3f}%'.format(
losses.avg, top1.avg, top5.avg))
acc = top1.avg
loss = losses.avg
return acc, loss
def validate(loader, model, criterion_rgb, criterion_local, epoch=0):
# batch_time = AverageMeter()
losses = AverageMeter()
metric = Metrics(max_depth=args.max_depth, disp=args.use_disp, normal=args.normal)
score = AverageMeter()
score_1 = AverageMeter()
loss_rgb = torch.zeros(1)
# Evaluate model
model.eval()
# Only forward pass, hence no grads needed
with torch.no_grad():
# end = time.time()
for i, (input, gt) in enumerate(loader):
if not args.no_cuda:
input, gt = input.cuda(non_blocking=True), gt.cuda(non_blocking=True)
prediction = model(input)
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.calculate(prediction[:, 0:1], gt)
score.update(metric.get_metric(args.metric), metric.num)
score_1.update(metric.get_metric(args.metric_1), metric.num)
if (i + 1) % args.print_freq == 0:
print('Test: [{0}/{1}]\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Metric {score.val:.4f} ({score.avg:.4f})'.format(
i+1, len(loader), loss=losses,
score=score))
# Synchronization needed
if args.world_size>1:
score.synchronize_between_processes()
score_1.synchronize_between_processes()
if args.evaluate:
print("===> Average RMSE score on validation set is {:.4f}".format(score.avg))
print("===> Average MAE score on validation set is {:.4f}".format(score_1.avg))
return score.avg, score_1.avg, losses.avg
def validate(val_loader, model, criterion):
batch_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
# switch to evaluate mode
model.eval()
with torch.no_grad():
target_index_output, target_index_target = list(), list()
end = time.time()
for i, (input, target) in enumerate(val_loader):
input = input.cuda(non_blocking=True)
target = target.cuda(non_blocking=True)
# compute output
output = model(input)
loss = criterion(output, target)
# measure accuracy and record loss
prec1 = accuracy(output, target)
losses.update(loss.item(), input.size(0))
top1.update(prec1[0], input.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
# for auroc get value from target index
output_cpu = output.cpu().data.numpy()
output_cpu = np.array(
[softmax(out)[args.target_index] for out in output_cpu])
target_index_output.extend(output_cpu.astype(np.float))
target_index_target.extend(
np.equal(target.cpu().data.numpy(),
args.target_index).astype(np.int))
# --------------------------------------
if i % args.print_freq == 0:
print('Test: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Prec@1 {top1.val:.3f} ({top1.avg:.3f})'.format(
i,
len(val_loader),
batch_time=batch_time,
loss=losses,
top1=top1))
auc, roc = compute_auroc(target_index_output, target_index_target)
print(' * Prec@1 {top1.avg:.3f}'.format(top1=top1))
save_auroc(auc, roc, os.path.join(args.result, 'rocgraph' + '.png'))
return top1.avg
def train(train_loader, model, criterion, optimizer, epoch):
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
# switch to train mode
model.train()
end = time.time()
for i, (input, target) in enumerate(train_loader):
# measure data loading time
data_time.update(time.time() - end)
input = input.cuda(non_blocking=True)
target = target.cuda(non_blocking=True)
# compute output
output = model(input)
loss = criterion(output, target)
# measure accuracy and record lossdks
prec1 = accuracy(output, target)
losses.update(loss.item(), input.size(0))
top1.update(prec1[0], input.size(0))
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % args.print_freq == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Prec@1 {top1.val:.3f} ({top1.avg:.3f})'.format(
epoch,
i,
len(train_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
top1=top1))
savelogloss = save_loss_log(epoch,
losses.avg,
args.result,
filename='losslog.txt')
save_log_graph(log_file=savelogloss)