os.makedirs(path_to_results)
n_epoch = 40
batch_size = 32
n_cv_runs = 10
for cv_run in tqdm(range(n_cv_runs)):
data = np.load(path_to_data + 'train_test_data_{}.npz'.format(cv_run))
train_images = data['train_images']
train_labels = data['train_labels']
test_images = data['test_images']
test_labels = data['test_labels']
cnn = cnn_adjust_lr(n_classes=train_labels.shape[1], input_shape=train_images[0].shape, lr=1e-4)
train_accuracy = np.zeros((n_epoch,), dtype=np.float64)
test_accuracy = np.zeros((n_epoch,), dtype=np.float64)
for epoch in range(n_epoch):
print('\tepoch %d...' % epoch)
cnn.fit(train_images, train_labels, epochs=1, shuffle=True, batch_size=batch_size, verbose=0)
train_prediction = cnn.predict(train_images)
train_accuracy[epoch] = np.mean(np.argmax(train_prediction, axis=1) == np.argmax(train_labels, axis=1))
print('\ttrain accuracy {}'.format(train_accuracy[epoch]))
test_prediction = cnn.predict(test_images)
test_accuracy[epoch] = np.mean(np.argmax(test_prediction, axis=1) == np.argmax(test_labels, axis=1))
'results_{}_run.npz'.format(run))
test_accuracy = cur_results['test_accuracy'][-1]
if test_accuracy > best_accuracy:
best_accuracy = test_accuracy
best_model_id = run
train_data = np.load(path_to_train_data +
'train_test_data_{}.npz'.format(best_model_id))
image_mean = train_data['image_mean']
image_std = train_data['image_std']
patch_width, patch_height, patch_channels = train_data['train_images'][0].shape
model = cnn_adjust_lr(n_classes=train_data['train_labels'].shape[1],
input_shape=(patch_width, patch_height, patch_channels),
lr=1e-4)
model.load_weights(path_to_train_results +
'trained_model/weights_{}_run.ckpt'.format(best_model_id))
writer = GdalIO()
for file in os.listdir(path_to_test_images):
image = np.transpose(np.array(
gdal.Open(path_to_test_images + file).ReadAsArray()),
axes=(1, 2, 0)).astype(np.float64)
writer.parse_meta_with_gdal(path_to_test_images + file)
pred_boxes, pred_probs = apply_for_sliding_window(
model,
image,