batch_size=args.batch_size,
num_workers=args.num_workers)
val_loader = DataLoader(val_dataset,
batch_size=args.batch_size,
num_workers=args.num_workers)
dset_loaders = {'train': train_loader, 'val': val_loader}
dset_sizes = {'train': len(train_dataset), 'val': len(val_dataset)}
optimizer = getattr(optim, args.optimizer)(net.parameters(), args.lr)
scheduler = ReduceLROnPlateau(optimizer,
'min',
patience=2,
verbose=True)
model.compile(loss, optimizer, scheduler=scheduler)
folder_checkpoint = "model_checkpoints/" + args.env_name.replace(
' ', '_') + '/'
if not os.path.exists(folder_checkpoint):
os.mkdir(folder_checkpoint)
model_checkpoint = folder_checkpoint + "weights.{epoch:02d}-train_loss:{train_loss:.2f}" \
+ "-train_acc:{train_acc:.2f}-val_loss:{val_loss:.2f}-val_acc:{val_acc:.2f}.pkl"
monitor_outputs = [0, 1, 2]
display_acc = 1
loss_weights = None
dset_loaders = {'train_two_class': train_loader, 'val': val_loader}
dset_sizes = {
'train_two_class': len(train_dataset),
'val': len(val_dataset)
}
# dict_param = [{'params': x.parameters()} for x in net.phi_r]
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(
net.parameters(),
lr=1e-3,
# momentum=0.9
)
model_checkpoint = "/scratch/datasets/vrd/weights.{epoch:02d}-train_loss:{train_loss:.2f}-train_acc:{train_acc:.2f}" \
+ "-val_loss:{val_loss:.2f}-val_acc:{val_acc:.2f}.pkl"
model.compile(loss=criterion,
opt=optimizer,
weight_decay=0.1,
decay_step=10)
model.fit_loader(dset_loaders,
dset_sizes,
batch_size=batch_size,
num_epochs=num_epochs,
model_checkpoint=model_checkpoint,
env=env_name)
else:
run_evaluate(model, 'data/annotations_test.json', 'test')
val_dataset = ImprovedVRDDataset('test', args)
val_loader = DataLoader(val_dataset,
batch_size=args.batch_size,
num_workers=args.num_workers)
dset_loaders = {'train': train_loader, 'val': val_loader}
dset_sizes = {'train': len(train_loader), 'val': len(val_dataset)}
criterion = getattr(nn, args.loss)()
optimizer = getattr(optim, args.optimizer)(net.parameters(), args.lr)
scheduler = ReduceLROnPlateau(optimizer,
'min',
patience=2,
verbose=True)
model.compile(criterion, optimizer, scheduler=scheduler)
folder_checkpoint = "model_checkpoints/" + args.env_name.replace(
' ', '_') + '/'
if not os.path.exists(folder_checkpoint):
os.mkdir(folder_checkpoint)
model_checkpoint = folder_checkpoint + "weights.{epoch:02d}-train_loss:{train_loss:.2f}" \
+ "-train_acc:{train_acc:.2f}-val_loss:{val_loss:.2f}-val_acc:{val_acc:.2f}.pkl"
monitor_outputs = [0, 1, 2] if args.exp_num in [
6, 11, 12, 14, 16, 17, 18, 19, 20, 22
] else None
model.fit_loader(dset_loaders,
def _compile(self, inn, outt):
model = Model(inputs=inn, outputs=outt)
model.compile(optimizer='RMSprop', loss=tnorm_loss, metrics=[bin_acc])
return model
