# log time
start_time = time.time()
for epoch in range(num_epoch):
for data in dataloader:
inp = data[1]
target = data[2]
output = model(inp)
loss = criterion(output, target)
optimizer.zero_grad()
loss.backward()
optimizer.step()
# log
acc = models.cal_accuracy(dev_dataset, model)
print('epoch [{}/{}], loss:{:.4f}, accuracy:{:.4f}'.format(
epoch + 1, num_epoch, loss.data.item(), acc))
log_df.loc[epoch] = [epoch + 1, loss.data.item(), acc]
# time log
print("--- %s seconds ---" % (time.time() - start_time))
# save log df
if is_consider_residual:
log_df.to_pickle(f'{log_path}/conv_hybrid_loss_acc_log.pkl')
else:
log_df.to_pickle(f'{log_path}/conv_hybrid_no_res_loss_acc_log.pkl')
def print_acc(model, dataset, print_note=''):
acc = models.cal_accuracy(dataset, model)
def print_acc(model, dataset, print_note=''):
acc = models.cal_accuracy(dataset, model)
print(print_note, 'accuracy: ', acc)
# file paths
model_path = 'networks/model'
# prepare test set
test_dataset, channel_len, _ = load_labeled_dataset()
# load model
autoencoder = models.sdae(dimensions=[channel_len, 10, 10, 20, 3])
model = models.sdae_lr(autoencoder)
model.load_state_dict(
torch.load(f'{model_path}/ae_on_{data_class}.pth', map_location='cpu'))
model.eval()
# view output
acc = models.cal_accuracy(test_dataset, model)
print('accuracy: ', acc)
features = models.encode(test_dataset, model)
features = features.numpy()
features_with_label = {0: [], 1: [], 2: []}
for i in range(len(test_dataset)):
label = test_dataset.grdtruth[i].item()
features_with_label[label].append(features[i])
fig = plt.figure()
ax = Axes3D(fig)
colors = ['r', 'g', 'b']
for i in range(3):
class_f = np.array(features_with_label[i])
