PixelRNNs, dataset logic

pixel_rnn_dataset_logic.py

def pad_nulls(arraylike, seq_len, input_size):
  '''
  Pad nulls to the end of the arraylike, so that the length of the arraylike
  is divisible by the batch_size.
  '''
  if len(arraylike)==0: 
    arraylike = np.zeros( (seq_len, input_size), dtype=np.float32)

  remainder = len(arraylike) % seq_len
  if remainder != 0:
      nulls = np.zeros((seq_len - remainder, input_size), dtype=np.float32)
      arraylike = np.concatenate((nulls, arraylike), axis=0)

  return arraylike
        
X, Y = list(), list()
seq_len = image_flatten.shape[0]
input_size = image_flatten.shape[-1]

for idx, row in enumerate(image_flatten):
  x = image_flatten[0 : idx+1, :]
  y = image_flatten[idx+1, :] if idx+1<len(image_flatten) else np.zeros((input_size), dtype=np.float32)
  x, y = pad_nulls(x, seq_len, input_size), y
  x, y = np.array(x), np.array(y)

  X.append(x); Y.append(y)

X_np, Y_np = np.array(X), np.array(Y)