batch_size=BATCH_SIZE,
num_workers=4,
shuffle=True)
val_dataloader = DataLoader(val_set,
batch_size=BATCH_SIZE,
num_workers=4,
shuffle=False)
# -------------------------- Model loading ------------------------------
model = BaseModel(IF_PRETRAINED=True)
model.to(device)
if MODEL_LOAD_PATH is not None:
model.load_state_dict(torch.load(MODEL_LOAD_PATH)['model'])
# -------------------------- Optimizer loading --------------------------
optimizer = optim.Adam(model.parameters(), lr=LR)
lr_schduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.2,
patience=5)
if MODEL_LOAD_PATH is not None:
optimizer.load_state_dict(torch.load(MODEL_LOAD_PATH)['optimizer'])
# ------------------------- Loss loading --------------------------------
camera_distance = 2.732
elevation = 0
azimuth = 0
renderer = sr.SoftRenderer(image_size=224,
sigma_val=1e-4,
aggr_func_rgb='hard',
camera_mode='look_at',
train_metrics = init_metrics()
test_metrics = init_metrics()
print('--- SETTINGS ---')
print('Number of sentiments to classify:', args.num_sentiments)
print('Learning rate:', INITIAL_LR)
print('Num of epochs per fold:', NUM_EPOCHS)
print('Use gaze features:', args.use_gaze)
print('\n> Starting 10-fold CV.')
for k, (train_loader, test_loader) in enumerate(dataset.split_cross_val(10)):
# initialize model and optimizer every fold
model = BaseModel(lstm_units,
dataset.max_sentence_length, args.num_sentiments,
initial_word_embedding.clone(), args.use_gaze)
optimizer = SGD(model.parameters(),
lr=INITIAL_LR,
momentum=0.95,
nesterov=True)
# optimizer_scheduler = lr_scheduler.StepLR(
# optimizer, step_size=halve_lr_every_passes, gamma=0.5)
if USE_CUDA:
model = model.cuda()
for e in range(NUM_EPOCHS):
train_loss, train_results = iterate(train_loader)
# save the training metrics of last epoch
for metric, value in train_results.items():
train_metrics[metric].append(value)
if args.train:
# prepare dataloader
train_loader = DataLoader(dataset=CustomData('train', dir_path=args.data_dir),
batch_size=args.batch_size,
num_workers=args.num_workers,
shuffle=True)
val_loader = DataLoader(dataset=CustomData('val', dir_path=args.data_dir),
batch_size=args.batch_size,
num_workers=args.num_workers,
shuffle=False)
# prepare optimizer
optimizer = None
if args.optim == 'sgd':
optimizer = torch.optim.SGD([
{'params': network.parameters()}
], lr=args.lr, weight_decay=1e-4, momentum=0.9)
elif args.optim == 'adam':
optimizer = torch.optim.Adam([
{'params': network.parameters()}
], lr=args.lr, weight_decay=1e-5)
scheduler = None
if args.scheduler == 'poly':
func = lambda epoch: (1 - epoch / args.n_epochs)**0.9 # poly
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, func)
# prepare criterion(s)
criterion_cls = nn.CrossEntropyLoss()
# start training