trained_model, optimizer, criterion, loss_hist, loss_val_hist = train_classifier.run_train(
n_epochs=n_epochs, lr=learning_rate, batch_size=batch_size)
pre.save_results(loss_hist, loss_val_hist, f'{model_name}')
#get parameters get the paramteres fro the last time forwads was called
#since at the end of traning we check the accuracy of the train data, the pasarmeter have a size of all the train data
detected_patterns = model.get_detected_patterns1()
for idx in range(10):
plt.figure(1, figsize=(20, 10))
for p in range(trained_model.n_patterns1):
pattern = detected_patterns[idx][p].reshape(
detected_patterns.shape[2], detected_patterns.shape[3])
patern_np = pattern.detach().numpy().reshape(8, 8)
plt.subplot(2, 3, 1 + p)
plt.imshow(patern_np, cmap='hot', interpolation='none')
pre.save_plt_as_image(plt, f'patterns1_{idx}')
detected_patterns = model.get_detected_patterns2()
for idx in range(10):
plt.figure(1, figsize=(20, 20))
for p in range(trained_model.n_patterns2):
pattern = detected_patterns[idx][p].reshape(
detected_patterns.shape[2], detected_patterns.shape[3])
patern_np = pattern.detach().numpy().reshape(4, 4)
plt.subplot(4, 4, 1 + p)
plt.imshow(patern_np, cmap='hot', interpolation='none')
pre.save_plt_as_image(plt, f'patterns2_{idx}')
if args.s_model:
m_exporter = ModelExporter('digits')
m_exporter.save_nn_model(trained_model, optimizer, 0, n_classes,
plt.figure(1, figsize=(20, 10))
for idx in range(30):
image = X_test[idx].detach().numpy().reshape(16, 16)
image2 = X_test_encoded[idx].detach().numpy().reshape(
int(math.sqrt(n_features_encoded)),
int(math.sqrt(n_features_encoded)))
image3 = X_test_decoded[idx].detach().numpy().reshape(16, 16)
# Call signature: subplot(nrows, ncols, index, **kwargs)
plt.subplot(10, 9, 1 + idx * 3)
plt.imshow(image, cmap='hot', interpolation='none')
plt.subplot(10, 9, 2 + idx * 3)
plt.imshow(image2, cmap='winter', interpolation='none')
plt.subplot(10, 9, 3 + idx * 3)
plt.imshow(image3, cmap='winter', interpolation='none')
pre.save_plt_as_image(plt, f'encoded_{n_features_encoded}')
#for idx in range(30):
# plt.figure(1, figsize=(10, 5))
# plt.subplot(1, 3, 1)
# X_test_np = X_test[idx].detach().numpy().reshape(16, 16)
# plt.imshow(X_test_np, cmap='hot', interpolation='none')
#
# plt.subplot(1, 3, 2)
# test_num_np = X_test_encoded[idx].detach().numpy().reshape(int(math.sqrt(n_features_encoded)), int(math.sqrt(n_features_encoded)))
# plt.imshow(test_num_np, cmap='hot', interpolation='none')
#
# plt.subplot(1, 3, 3)
# X_test_decoded_np = X_test_decoded[idx].detach().numpy().reshape(16, 16)
# plt.imshow(X_test_decoded_np, cmap='hot', interpolation='none')
#
model = m_importer.load_nn_model(model_name, 0, 10, 100)
image = model.reshape_data(torch.tensor(img_array, device=device, dtype=dtype))
y_test = torch.tensor(int(args.num), device=device, dtype=torch.long)
y_pred = model(image).argmax(1)
accuracy_soft = (y_pred == y_test).float().mean()
print(f'number predicted {y_pred.item()}')
pre = Preprocessing('digits')
detected_patterns = model.get_detected_patterns1()
plt.figure(1, figsize=(20, 10))
for p in range(model.n_patterns1):
pattern = detected_patterns[0][p].reshape(detected_patterns.shape[2], detected_patterns.shape[3])
patern_np = pattern.detach().numpy().reshape(8, 8)
plt.subplot(2, 3, 1 + p)
plt.imshow(patern_np, cmap='hot', interpolation='none')
pre.save_plt_as_image(plt, f'written_number {args.num}_patterns')
detected_patterns = model.get_detected_patterns2()
plt.figure(1, figsize=(20, 20))
for p in range(model.n_patterns2):
pattern = detected_patterns[0][p].reshape(detected_patterns.shape[2], detected_patterns.shape[3])
patern_np = pattern.detach().numpy().reshape(4, 4)
plt.subplot(4, 4, 1 + p)
plt.imshow(patern_np, cmap='hot', interpolation='none')
pre.save_plt_as_image(plt, f'written_number {args.num}_patterns2')
print(t, loss.item())
print(time.time() - start)
y_pred = model(X_train).argmax(1)
accuracy_soft = (y_pred == y_train.long()).float().mean()
print(f'training accuracy: {accuracy_soft}')
print(f'optimal iteration: {acc_val.index(max(acc_val))}')
title = f'train_acc_val_acc_vs_iter_{n_iter}'
plt.figure(1)
plt.plot(acc_train)
plt.plot(acc_val)
plt.title = title
pre.save_plt_as_image(plt, title)
plt.close()
title = f'train_loss_val_loss_vs_iter_{n_iter}'
plt.figure(2)
plt.plot(loss_hist)
plt.plot(loss_val_hist)
plt.title = title
pre.save_plt_as_image(plt, title)
#plt.show()
if args.s_model:
name = f'custom_{n_features}.pt'
pre.save_model(model, name=name)
#torch.save(model, name)