def train_fcnn(params):
print("Training fcnn")
params_dict = file_utils.get_FCNN_hyperparams(params)
trainer = FCNN.Trainer(params_dict)
history = trainer.train()
trainer.save_model()
trainer.save_params(params_dict)
trainer.save_losses(history)
def learn(inputs, error):
fcnn.backprop(net.receptors, net.synapses, net.bias, net.deltas, error)
cnn.backprop(net.filters, net.conv_bias, net.pooled, net.switches, net.conv_error, net.conv_delta, net.deltas[0], inputs)
def execute_net(inputs):
cnn.fwdPass(inputs, net.convolved, net.filters, net.conv_bias, net.pooled, net.switches)
fcnn.fwdPass(net.pooled[net.conv_layers-1], net.receptors, net.synapses, net.bias) #pass the CNN output to FC NNet
num_workers=1)
test_loader = data_utils.DataLoader(test_set,
batch_size=N_test,
shuffle=False,
num_workers=1)
for u in upscale_method:
for lr in learning_rates:
for m in momentums:
for wd in weight_decays:
for cw in class_weights:
torch.manual_seed(1984)
if u == 1:
model = FCNN.FCNN()
elif u == 2:
model = FCNN2.FCNN2()
else:
print("WARNING Unexpected upscale method!")
continue
expid = str(exp).zfill(3)
exp += 1
print("\n\n", expid, "Model LR:", lr, "Momentum:",
m, "WDecay:", wd, "Class Weights:", cw,
"Batch size:", n, "Imsize:", w)
model = train.train(expid, model, train_loader,
test_loader, n, T, w, lr, m,
wd, cw, lr_stepsize, lr_decay)
def main():
# Training settings
nEpochs = 200
lr = 0.001
step = 80
cuda = True
resume = ""
start_epoch = 1
clip = 0.005
threads = 1
momentum = 0.9
weight_decay = 1e-4
# model_path = 'model_HSI_SR_DUALNET2_d2_totalloss_SAVE_s2_250/model_epoch_150.pth'
path_hr = 'D:\data\hr.mat'
path_lr = 'D:\data\lr_3.mat'
global model
if cuda and not torch.cuda.is_available():
raise Exception("No GPU found, please run without --cuda")
seed = random.randint(1, 10000)
print("Random Seed: ", seed)
torch.manual_seed(seed)
if cuda:
torch.cuda.manual_seed(seed)
cudnn.benchmark = True
# print("===> Loading datasets")
# train_set = DatasetFromMat(opt.dataset, sigma)
# train_set = DatasetFromMat7_3(opt.dataset)
print("===> Loading datasets ")
train_data_lr, train_data_hr = DataFromMat(path_lr, path_hr)
_, _, S, H, W = train_data_lr.shape
torch_dataset = data.TensorDataset(train_data_lr, train_data_hr)
training_data_loader = data.DataLoader(dataset=torch_dataset,
num_workers=threads,
batch_size=batchSize,
shuffle=False)
print("===> Building model")
# model = torch.load(model_path)["model"]
model = FCNN.FCNN(1, 1)
# criterion = nn.MSELoss()
# criterion = nn.MSELoss(size_average=False)
print("===> Setting GPU")
if cuda:
# model = to2rch.nn.DataParallel(model).cuda()
model = dataparallel(model, 1)
# set the number of parallel GPUs
# criterion = criterion.cuda()
if resume:
if os.path.isfile(resume):
print("=> loading checkpoint '{}'".format(resume))
checkpoint = torch.load(resume)
start_epoch = checkpoint["epoch"] + 1
model.load_state_dict(checkpoint["model"].state_dict())
else:
print("=> no checkpoint found at '{}'".format(resume))
print("===> Setting Optimizer")
# optimizer = optim.SGD([
# {'params': model.parameters()}
# ], lr=opt.lr, momentum=opt.momentum, weight_decay=opt.weight_decay)
optimizer = optim.Adam([{
'params': model.parameters()
}],
lr=lr,
weight_decay=weight_decay)
print("===> Training")
lossAarry = np.zeros(nEpochs)
losspath = 'losses/'
if not os.path.exists(losspath):
os.makedirs(losspath)
for epoch in range(start_epoch, nEpochs + 1):
start_time = time.time()
lossAarry[epoch - 1] = lossAarry[epoch - 1] + train(
training_data_loader, optimizer, model, epoch, step, lr, cuda)
print("===> Epoch[{}]: Loss={:.7f}, time = {:.4f}".format(
epoch, lossAarry[epoch - 1],
time.time() - start_time))
save_checkpoint(model, epoch)
sio.savemat(losspath + method_name + '_lossArray.mat',
{'lossArray': lossAarry})
