def eval(modis_utils_obj, data_type, idx, metric):
if metric is None:
metric = mse_with_mask
target_path = modis_utils_obj._data_files[data_type]['target']
target = get_target_test(target_path, idx)
mask_path = modis_utils_obj._data_files[data_type]['mask']
mask = get_target_test(mask_path, idx)
predict = modis_utils_obj.get_inference(data_type, idx)
return metric(target, predict, mask)
def predict_and_visualize_by_data_file_one_output_and_gridding(
data_file_path, target_file_path, model, which=0, result_dir=None):
input_seq = get_data_test(data_file_path, which)
if real_time_steps is not None:
time_steps = real_time_steps
input_seq = input_seq[input_seq.shape[0] - real_time_steps:, :, :]
else:
time_steps = input_seq.shape[0]
groundtruth = get_target_test(target_file_path, which)
if mask_file_path is not None:
mask_groundtruth = get_target_test(mask_file_path, which)
mask_groundtruth[mask_groundtruth == -1] = 0
groundtruth = scale_normalized_data(groundtruth, groundtruth_range)
plt.figure(figsize=(10, 10))
offset_x = input_seq.shape[1] % crop_size
offset_y = input_seq.shape[2] % crop_size
input_seq = input_seq[:, offset_x//2:-(offset_x - offset_x//2), \
offset_y//2:-(offset_y - offset_y//2)]
groundtruth = groundtruth[offset_x//2:-(offset_x - offset_x//2), \
offset_y//2:-(offset_y - offset_y//2)]
pred_img = np.zeros_like(groundtruth)
for i in range(input_seq.shape[1] // crop_size):
for j in range(input_seq.shape[2] // crop_size):
pred = model.predict(input_seq[np.newaxis, :, \
i*crop_size:(i+1)*crop_size, \
j*crop_size:(j+1)*crop_size, np.newaxis])
pred_img[i*crop_size:(i+1)*crop_size, \
j*crop_size:(j+1)*crop_size] = pred[1][0,:,:,0]
G = gridspec.GridSpec(2, time_steps)
for i, img in enumerate(input_seq):
axe = plt.subplot(G[0, i])
axe.imshow(img)
ax_groundtruth = plt.subplot(G[1, :time_steps // 2])
ax_groundtruth.imshow(groundtruth)
ax_groundtruth.set_title('groundtruth')
ax_pred = plt.subplot(G[1, time_steps // 2:2 * (time_steps // 2)])
ax_pred.imshow(pred_img)
ax_pred.set_title('predict')
if result_dir is not None:
if not os.path.exists(result_dir):
os.makedirs(result_dir)
plt.savefig(os.path.join(result_dir, '{}.png'.format(which)))
return (groundtruth, pred_img)
def predict_and_visualize_by_data_file_one_output(data_file_path,
target_file_path,
pred_img,
which=0,
result_dir=None,
model=None):
example = get_data_test(data_file_path, which)
time_steps = example.shape[0]
groundtruth = get_target_test(target_file_path, which)
plt.figure(figsize=(10, 10))
if model is not None and pred is None:
pred = model.predict(example[np.newaxis, :, :, :, np.newaxis])
G = gridspec.GridSpec(2, time_steps)
pred_img
for i, img in enumerate(example):
axe = plt.subplot(G[0, i])
axe.imshow(img)
ax_groundtruth = plt.subplot(G[1, :time_steps // 2])
ax_groundtruth.imshow(groundtruth)
ax_groundtruth.set_title('groundtruth')
ax_pred = plt.subplot(G[1, time_steps // 2:2 * (time_steps // 2)])
ax_pred.imshow(pred_img)
ax_pred.set_title('predict')
if result_dir is not None:
if not os.path.exists(result_dir):
os.makedirs(result_dir)
plt.savefig(os.path.join(result_dir, '{}.png'.format(which)))
return (groundtruth, pred_img)
def get_water_cloud_mask(self, data_type, idx):
mask_file_path = self._data_files[data_type]['mask']
return get_target_test(mask_file_path, idx)
def get_groundtruth(self, data_type, idx):
target_file_path = self._data_files[data_type]['target']
return get_target_test(target_file_path, idx)
