## data loader for val set, which is completely labeled
val_set = pipeline.IntrinsicDataset(args.data_path,
args.val_sets,
args.val_intrinsics,
inds=range(0,
args.num_val * args.val_offset,
args.val_offset),
array=args.unlabeled_array,
size_per_dataset=args.set_size)
val_loader = torch.utils.data.DataLoader(val_set,
batch_size=args.batch_size,
num_workers=args.loader_threads,
shuffle=False)
## print out error plots after every epoch for every prediction
logger = pipeline.Logger(
['recon', 'refl', 'depth', 'shape', 'lights', 'shading'], args.save_path)
param_updater = pipeline.ParamUpdater(args.transfer)
for epoch in range(args.num_epochs):
print '<Main> Epoch {}'.format(epoch)
if param_updater.check(epoch):
## update which parameters are updated
transfer = param_updater.refresh(epoch)
print 'Updating params: ', epoch, transfer
## get a new trainer with different learnable parameters
trainer = pipeline.ComposerTrainer(model,
train_loader,
args.lr,
args.lights_mult,
args.un_mult,
batch_size=args.batch_size,
num_workers=args.loaders,
shuffle=True)
val_set = pipeline.IntrinsicDataset(args.data_path,
args.val_sets,
args.intrinsics,
array=args.array,
size_per_dataset=args.num_val)
val_loader = torch.utils.data.DataLoader(val_set,
batch_size=args.batch_size,
num_workers=args.loaders,
shuffle=False)
## plots errors for the intrinsic image predictions
logger = pipeline.Logger(['refl', 'shape', 'lights'], args.save_path)
trainer = pipeline.DecomposerTrainer(model, train_loader, args.lr,
args.lights_mult)
for epoch in range(args.num_epochs):
print('<Main> Epoch {}'.format(epoch))
## save model state
state = model.state_dict()
torch.save(state, os.path.join(args.save_path, 'state.t7'))
## get losses and save visualization on val images
val_losses = pipeline.visualize_decomposer(model, val_loader,
args.save_path, epoch)
## data loader for val set, which is completely labeled
val_set = pipeline.IntrinsicDataset(
args.data_path,
args.val_sets,
args.val_intrinsics,
inds=range(0, args.num_val * args.val_offset, args.val_offset),
array=args.unlabeled_array,
size_per_dataset=args.num_val,
)
val_loader = torch.utils.data.DataLoader(
val_set, batch_size=args.batch_size, num_workers=args.loader_threads, shuffle=False
)
## print out error plots after every epoch for every prediction
logger = pipeline.Logger(
["recon", "refl", "depth", "shape", "lights", "shading"], args.save_path
)
param_updater = pipeline.ParamUpdater(args.transfer)
for epoch in range(args.num_epochs):
print("<Main> Epoch {}".format(epoch))
if param_updater.check(epoch):
## update which parameters are updated
transfer = param_updater.refresh(epoch)
print("Updating params: ", epoch, transfer)
## get a new trainer with different learnable parameters
trainer = pipeline.ComposerTrainer(
model,
train_loader,
args.lr,