def save_model():
if use_cuda:
Decoder.cpu()
torch.save(Decoder.state_dict(), './Decoder.pkl')
if use_cuda:
Decoder.cuda()
def load_weights():
Decoder.load_state_dict(torch.load('./Decoder.pkl'))
load_weights()
while True:
Decoder.train()
train_batch = []
ground_truth = []
for i in range(batch_size):
f = np.random.uniform(0, 1, 9)
train_batch.append(f)
ground_truth.append(draw(f))
train_batch = torch.tensor(train_batch).float()
ground_truth = torch.tensor(ground_truth).float()
if use_cuda:
Decoder = Decoder.cuda()
train_batch = train_batch.cuda()
ground_truth = ground_truth.cuda()
gen = Decoder(train_batch)
optimizer.zero_grad()
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=3, gamma=0.9)
loss_func = nn.NLLLoss2d(weight=torch.FloatTensor([1,8000]).cuda())
data_loder = DataLoader()
zeros = np.zeros((512,512))
for epoch in range(epoch0, epoch0+EPOCH):
scheduler.step()
train_step = 0
test_step = 0
train_loss = 0
test_loss = 0
train_accu = 0
test_accu = 0
for train, test in data_loder.gen(BATCH_SIZE, get_stored=True):
net = net.train()
for X, y in train:
X = torch.FloatTensor(X[:,np.newaxis,:,:])
y = torch.LongTensor(y)
X = Variable(X).cuda()
y = Variable(y).cuda()
output = net(X)
output = F.log_softmax(output,dim=1)
loss = loss_func(output, y)
train_loss += loss.data[0]
optimizer.zero_grad()
loss.backward()
optimizer.step()
train_step += 1