def transform(self, observation: TensorType) -> np.ndarray:
"""Downsamples images from (210, 160, 3) by the configured factor."""
self.check_shape(observation)
scaled = observation[25:-25, :, :]
if self._dim < 84:
scaled = resize(scaled, height=84, width=84)
# OpenAI: Resize by half, then down to 42x42 (essentially mipmapping).
# If we resize directly we lose pixels that, when mapped to 42x42,
# aren't close enough to the pixel boundary.
scaled = resize(scaled, height=self._dim, width=self._dim)
if self._grayscale:
scaled = scaled.mean(2)
scaled = scaled.astype(np.float32)
# Rescale needed for maintaining 1 channel
scaled = np.reshape(scaled, [self._dim, self._dim, 1])
if self._zero_mean:
scaled = (scaled - 128) / 128
else:
scaled *= 1.0 / 255.0
return scaled
def observation(self, frame):
frame = rgb2gray(frame)
frame = resize(frame, height=self.height, width=self.width)
return frame[:, :, None]