Productive error handler

Classification task is the most common in AI because it requires few libraries. I try to write using the resources of an online compiler, without understanding the intricacies of the work.

def rle_decode(mask_rle, shape=(1280, 1918, 1)):
    '''
    mask_rle: run-length as string formated (start length)
    shape: (height,width) of array to return 
    Returns numpy array, 1 - mask, 0 - background
    '''
    img = np.zeros(shape[0]*shape[1], dtype=np.uint8)

    s = mask_rle.split()
    starts, lengths = [np.asarray(x, dtype=int) for x in (s[0:][::2], s[1:][::2])]
    starts -= 1
    ends = starts + lengths    
    for lo, hi in zip(starts, ends):
        img[lo:hi] = 1

    img = img.reshape(shape)
    return img

For example, using the function of decoding masks 0/1, you can rely on their lengths. But to generate batch packets of a neural network, you still need to monitor the current results.

def keras_generator(gen_df, batch_size):
    while True:
        x_batch = []
        y_batch = []

        for i in range(batch_size):
            img_name, mask_rle = gen_df.sample(1).values[0] 
            img = cv2.imread('data/train/{}'.format(img_name))
            mask = rle_decode(mask_rle)

            img = cv2.resize(img, (256, 256)) 
            mask = cv2.resize(mask, (256, 256))

            x_batch += [img] 
            y_batch += [mask]

        x_batch = np.array(x_batch) / 255. 
        y_batch = np.array(y_batch)

        yield x_batch, np.expand_dims(y_batch, -1)

1. I like to put intermediate outputs of the result for eye contact with the code
2. If the result doesn't seem to be satisfactory, I edit the previous functions

im_id = 5
fig, axes = plt.subplots(nrows=1, ncols=2, figsize=(25, 25))
axes[0].imshow(x[im_id]) 
axes[1].imshow(pred[im_id, ..., 0] > 0.5) 

plt.show()

Output of the result = guaranteed contact with the written code. In this case, exception handling is not needed.