input_size = train_data.shape[2]
n_filters = 64
hidden_size = 64
EPOCH = 100
lr = 1e-2
model = CNN_LSTM(input_size, n_filters, hidden_size, time_step).to(DEVICE)
if not TEST:
print('training model...')
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
loss_func = nn.CrossEntropyLoss()
x = torch.from_numpy(train_data).float()
y = torch.from_numpy(train_label).long()
for epoch in range(EPOCH):
batches = gen_batch(x, y, batch_size)
loss_sum = 0
for var_x, var_y in batches:
pred = model(var_x)
loss = loss_func(pred, var_y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_sum += loss.item() * pred.shape[0]
print('epoch %d loss: %f' % (epoch, loss_sum / x.shape[0]))
if (epoch + 1) % 50 == 0:
torch.save(model.state_dict(), tmp_save_path % (epoch + 1))
torch.save(model.state_dict(), model_save_path)
else:
print('loading model...')
model.load_state_dict(torch.load(model_save_path))
print(train_data.shape, test_data.shape, sum(train_label), sum(test_label))
batch_size = 512
EPOCH = 500
lr = 0.001
model_save_path = './models/nn.model'
TEST = True
x = torch.from_numpy(train_data).float()
y = torch.from_numpy(train_label).long()
model = NN().to(DEVICE)
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
loss_func = torch.nn.CrossEntropyLoss()
if not TEST:
for epoch in range(EPOCH):
batches = gen_batch(x, y, batch_size)
loss_sum = 0
for var_x, var_y in batches:
pred = model(var_x)
loss = loss_func(pred, var_y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_sum += loss.item()
print('epoch %d loss: %f' % (epoch, loss_sum))
torch.save(model.state_dict(), model_save_path)
else:
model.load_state_dict(torch.load(model_save_path))
var_test_x = torch.from_numpy(test_data).float().to(DEVICE)
var_test_y = torch.from_numpy(test_label).long().to(DEVICE)