def run_convbase_on_images(src_folder, dst_file, image_shape, force_all=False):
"""
Run the imported convolutional base on the cropped images.
Save the result in a file.
Parameters
----------
src_folder (str):
The folder of the images which have been cropped.
dst_file (str):
The destination file of images which have been processes in the convolutional base.
force_all (bool), default False.
If false, process only images which are present in the src_folder, but not the dst_folder.
If True, delete everything in the dst_folder and process everything in the src_folder.
Returns
-------
None
"""
if force_all:
xd = XData()
else:
try:
with open(dst_file, "rb") as f:
xd = pickle.load(f)
except FileNotFoundError:
xd = XData()
src_images = os.listdir(src_folder)
dst_images = xd.get_stored_filenames()
missing_in_dst = list(set(src_images).difference(set(dst_images)))
convbase = get_VGG16_convbase(image_shape)
for i, filename in enumerate(missing_in_dst):
print(f"{i + 1}/{len(missing_in_dst)}: Processing {filename}")
src_path = os.path.join(src_folder, filename)
im = cv2.imread(src_path)
im = normalize_image(im)
single_image_batch = np.array([im])
after_convbase_single = convbase.predict(single_image_batch)
after_convbase_single = after_convbase_single.flatten()
xd.add_image(filename, after_convbase_single, assert_dimensions=1)
if i % 100 == 0:
print("Saving to avoid losing work if interrupted...")
save_data(xd, dst_file)
save_data(xd, dst_file)
return xd
x_train = x[n_train:]
z_train = z[n_train:]
# ---- Test and viualize ---- #
x_old = x_test[:hist, ...].copy()
frames = []
obs_frames = []
state_frames = []
pf_frames = []
direct_frames = []
for t in range(0 + hist, n_test - 1):
z_new = z_test[t].copy()
z_new_test = z_test[t].copy()
x_new = x_test[t].copy()
x_hat_pf = pf.step(z_new)
cv2.imshow("real", normalize_image(x_new))
cv2.imshow("pf", normalize_image(x_hat_pf))
if cv2.waitKey(1) & 0xFF == ord('q'):
break
obs_frames.append(add_border(normalize_image(z_new)))
state_frames.append(add_border(normalize_image(x_new)))
pf_frames.append(add_border(normalize_image(x_hat_pf)))
cv2.destroyAllWindows()
# ---- Saves multiple samples as an image ---- #
idxs = np.arange(15, 55, 4, dtype=np.int16)
obs_img = np.concatenate(tuple(np.array(obs_frames)[idxs]), axis=1)
state_img = np.concatenate(tuple(np.array(state_frames)[idxs]), axis=1)
pf_img = np.concatenate(tuple(np.array(pf_frames)[idxs]), axis=1)
full_img = np.concatenate((obs_img, state_img, df_img),
axis=0).astype(np.uint8)
direct_frames = []
for t in range(0+hist,n_test-1):
z_new = z_test[t].copy()
z_new_test = z_test[t].copy()
x_new = x_test[t].copy()
x_hat_pf = pf.step(z_new)
plt.subplot(1,2,1)
plt.imshow(x_hat_pf)
plt.subplot(1,2,2)
plt.imshow(x_new)
plt.show()
# cv2.imshow("real",normalize_image(x_new))
# cv2.imshow("pf",normalize_image(x_hat_pf))
# if cv2.waitKey(1) & 0xFF == ord('q'):
# break
obs_frames.append(add_border(normalize_image(z_new)))
state_frames.append(add_border(normalize_image(x_new)))
pf_frames.append(add_border(normalize_image(x_hat_pf)))
# cv2.destroyAllWindows()
# ---- Saves multiple samples as an image ---- #
idxs = np.arange(15,55,4, dtype = np.int16)
obs_img = np.concatenate(tuple(np.array(obs_frames)[idxs]),axis=1)
state_img = np.concatenate(tuple(np.array(state_frames)[idxs]),axis=1)
pf_img = np.concatenate(tuple(np.array(pf_frames)[idxs]),axis=1)
full_img = np.concatenate(( obs_img,state_img, df_img), axis = 0).astype(np.uint8)
matplotlib.image.imsave('samples_pf.png', full_img, cmap='gray')
# ---- Saves a video ---- #
outputdata = np.array(frames).astype(np.uint8)
skvideo.io.vwrite("samples.mp4", frames)
frames = []
obs_frames = []
state_frames = []
df_frames = []
direct_frames = []
for t in range(0 + hist, n_test - 1):
z_new = z_test[t].copy()
z_new_test = z_test[t].copy()
x_new = x_test[t].copy()
x_hat_df = df.predict_mean(x_old, z_new)
x_hat_df = x_hat_df[:, :, 0]
x_hat_df_like = df.estimate(z_new_test)
x_hat_df_like = x_hat_df_like[0, :, :, 0]
x_old[:-1, :, :] = x_old[1:, :, :]
x_old[-1, :, :] = x_hat_df
obs_frames.append(add_border(normalize_image(z_new)))
state_frames.append(add_border(normalize_image(x_new)))
df_frames.append(add_border(normalize_image(x_hat_df)))
direct_frames.append(add_border(normalize_image(x_hat_df_like)))
frame1 = np.concatenate((normalize_image(x_new), normalize_image(z_new)),
axis=1)
frame2 = np.concatenate(
(normalize_image(x_hat_df), normalize_image(x_hat_df_like)), axis=1)
frame = np.concatenate((frame1, frame2), axis=0)
frames.append(frame)
# ---- Saves multiple samples as an image ---- #
idxs = np.arange(15, 55, 4, dtype=np.int16)
obs_img = np.concatenate(tuple(np.array(obs_frames)[idxs]), axis=1)
state_img = np.concatenate(tuple(np.array(state_frames)[idxs]), axis=1)
df_img = np.concatenate(tuple(np.array(df_frames)[idxs]), axis=1)