def visualize_test_time(self, content_dict, visualize_indices,
verbose_folder):
# save classifier preds
sel_sigmoids = self.sigmoids[visualize_indices]
sigmoid_images = visualize_barplot_array(sel_sigmoids)
row_name = 'sigmoid_images'
content_dict[row_name] = save_imgs_direct(verbose_folder, row_name,
sigmoid_images)
sel_softmax_dists = self.softmax_differences[visualize_indices]
sigmoid_images = visualize_barplot_array(sel_softmax_dists)
row_name = 'softmax_of_differences'
content_dict[row_name] = save_imgs_direct(verbose_folder, row_name,
sigmoid_images)
def log_pair_predictions(self, img_pair, softmax_prediction, label, name,
step, phase):
image_pairs = img_pair.data.cpu().numpy().squeeze()
softmax_prediction = softmax_prediction.data.cpu().numpy().squeeze()
first_row = image_pairs[:, 0]
first_row = first_row[:self._n_logged_samples]
first_row = np.concatenate(unstack(first_row, 0), 2)
second_row = image_pairs[:, 1]
second_row = second_row[:self._n_logged_samples]
second_row = np.concatenate(unstack(second_row, 0), 2)
pred_score_images = visualize_barplot_array(
softmax_prediction[:self._n_logged_samples])
pred_score_images = [
np.transpose((img.astype(np.float32)) / 255., [2, 0, 1])
for img in pred_score_images
]
pred_row = np.concatenate(pred_score_images, axis=2)
label_row = get_text_row(label,
self._n_logged_samples,
shape=(30, pred_score_images[0].shape[1]))
full_image = (np.concatenate(
[first_row, second_row, pred_row, label_row], 1) + 1.) / 2.0
self._summ_writer.add_image('{}_{}'.format(name, phase), full_image,
step)
def visualize_test_time(self, content_dict, visualize_indices,
verbose_folder):
# save classifier preds
sel_softmax = self.out_softmax[visualize_indices]
sigmoid_images = visualize_barplot_array(sel_softmax)
row_name = 'softmax'
content_dict[row_name] = save_imgs_direct(verbose_folder, row_name,
sigmoid_images)
def vis_dist_over_traj(self, inputs, step):
images = inputs.demo_seq_images
cols = []
n_ex = 10
for t in range(images.shape[1]):
outputs = self.forward({
'current_img': images[:, t],
'goal_img': images[:, -1]
})
sigmoid = []
for i in range(len(outputs)):
sigmoid.append(
outputs[i].out_sigmoid.data.cpu().numpy().squeeze())
sigmoids = np.stack(sigmoid, axis=1)
sig_img_t = visualize_barplot_array(sigmoids[:n_ex])
img_column_t = []
images_npy = images.data.cpu().numpy().squeeze()
for b in range(n_ex):
img_column_t.append(ptrch2uint8(images_npy[b, t]))
img_column_t.append(np.transpose(sig_img_t[b], [2, 0, 1]))
img_column_t = np.concatenate(img_column_t, axis=1)
cols.append(img_column_t)
image = np.concatenate(cols, axis=2)
# cv2.imwrite('/nfs/kun1/users/febert/data/vmpc_exp/dist_over_traj.png', np.transpose(image, [1,2,0]))
self._logger.log_image(image, 'dist_over_traj', step, phase='test')
self._logger.log_video(np.stack(cols, axis=0),
'dist_over_traj_gif',
step,
phase='test')
print('logged dist over traj')