model_rgb = model_rgb.cuda()
model_depth = model_depth.cuda()
model_fusion = model_fusion.cuda()
if args.phase == 'train':
# Trainer: class, defined in trainer.py
optimizer_rgb = optim.SGD(model_rgb.parameters(),
lr=cfg['lr'],
momentum=cfg['momentum'],
weight_decay=cfg['weight_decay'])
optimizer_depth = optim.SGD(model_depth.parameters(),
lr=cfg['lr'],
momentum=cfg['momentum'],
weight_decay=cfg['weight_decay'])
optimizer_fusion = optim.SGD(model_fusion.parameters(),
lr=cfg['lr'],
momentum=cfg['momentum'],
weight_decay=cfg['weight_decay'])
training = Trainer(cuda=cuda,
model_rgb=model_rgb,
model_depth=model_depth,
model_fusion=model_fusion,
optimizer_rgb=optimizer_rgb,
optimizer_depth=optimizer_depth,
optimizer_fusion=optimizer_fusion,
train_loader=train_loader,
max_iter=cfg['max_iteration'],
snapshot=cfg['spshot'],
outpath=args.snapshot_root,
pilot_val = np.array(pilot_val)
pilot_val = pilot_val.reshape(num_val, -1)
pilot_val = de_complex(pilot_val)
pilot_val = add_noise(pilot_val, 20)
#merging the pilot with the sub6G signal
#total_train = np.concatenate((train_x,pilot_train),axis = 1)
#total_val = np.concatenate((val_x,pilot_val),axis = 1)
total_train = pilot_train
total_val = pilot_val
##### The training phase! ######
num_sample = train_x.shape[0]
fcn = Fusion()
optimizer = optim.Adam(fcn.parameters(), lr=lr)
loss_fun = nn.CrossEntropyLoss()
fcn.train()
for epoch in range(epoches):
print('epoch', epoch)
adjust_lr(optimizer, epoch)
# have shuffled already in the data preprocessing process
for i in range(int(num_sample / batch_size) + 1):
if ((i + 1) * batch_size <= num_sample):
X = train_x[i * batch_size:(i + 1) * batch_size, :]
P = pilot_train[i * batch_size:(i + 1) * batch_size, :]
Y = train_labels[i * batch_size:(i + 1) * batch_size]
else:
