# max_scale=5,
# noise_level=5,
# interp='bicubic',
# sub_pixel_flow=True)
inputs, targets = gen.generate_batch(1000)
# 0 prediction
error3 = np.mean(np.square(targets - 0), axis=1)
# CNN
cnn = CNN(split=False, normalize=True)
prediction = cnn.predict('checkpoints/normal/step81000.ckpt', inputs)
error = np.mean(np.square(prediction - targets), axis=1)
# CNN-split
cnn2 = CNN(split=True, normalize=True, fully_connected=500)
prediction = cnn2.predict('checkpoints/split/step44000.ckpt', inputs)
error4 = np.mean(np.square(prediction - targets), axis=1)
# FAST+LK
fastlk = FastLK(40, True)
# print(probs)
# index = np.argmax(index)
# word.append("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"[index])
# print(word)
# =============================================================================
# repeat test
for i in range(patches.shape[0]):
t = patches[i]
t_unscaled = tf.cast((t + .5) * 255.0, tf.int32)
#repeat_img = np.repeat(t_unscaled,2, axis=1)
repeat_img = tf.keras.backend.repeat_elements(t_unscaled, 2, axis=1)
print(repeat_img.shape)
show_img(repeat_img, scaled=False)
repeat_img = tf.cast(repeat_img, tf.float32)
t_scaled = repeat_img / 255.0 - .5
t_scaled = tf.expand_dims(t_scaled, axis=0)
index = cnn.predict(t_scaled)
pred_prob = int(np.max(index) * 100)
probs.append(pred_prob)
index = np.argmax(index)
pred = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"[
index]
word.append(pred)
# =============================================================================
# print(probs)
# print(word)
# =============================================================================
