How can I transform the render ? about minigrid HOT 6 CLOSED

Hi @DamienAllonsius. I'm not quite clear on what you want to do. I'm assuming you want to transform the observations the agent gets by processing them as RGB images?

One thing you should be aware of is that the image that you see when running manual_control.py, that is the human-viewable input, is not the same as what the agent sees. The agent gets a partially observable view of the environment that is encoded as a (7,7,3) array, which is not an image, but rather a symbolic encoding of the grid cells in front of the agent. What you see in manual_control.py is a rendering of the whole grid as an RGB image, which is produced by a call to env.render('human').

The easiest way to transform what the agent sees is using a wrapper, see the code in wrappers.py for some examples. There is a wrapper that transforms the input the agent gets into an RGB image, which may be a good starting point for what you want to do.

from minigrid.

Hi @maximecb Thanks a lot for your answer !
No, I really want to transform the rendering, not the observation.
I already give this observation to my agent

env.render(mode='rgb_array', highlight=False)

But then I transform the matrix by applying some downsampling with cv2, and I would like to see the result of it. Is that possible ? Thanks a lot Maxime
(sinon je suis français, on peut aussi communiquer par mail en français si tu préfères)
Damien

from minigrid.

You would have to modify the renderer class or create your own window using PyQT or Pyglet. The slightly annoying part is that you have to convert the rgba array into a QPixmap object, basically. Then the Window class here has a setPixmap method that you can use: https://github.com/maximecb/gym-minigrid/blob/master/gym_minigrid/rendering.py#L55

It's a bit annoying but definitely doable: https://stackoverflow.com/questions/34232632/convert-python-opencv-image-numpy-array-to-pyqt-qpixmap-image

If you do it, maybe you could add a method to the Window class to set what is being displayed from a numpy rgba array and create a pull request? :)

from minigrid.

Thanks for your answer. Ok I'm on it.

from minigrid.

Following the stackoverflow advices, I made a new function in Renderer class

    def drawArray(self, rgb_array):
        from PyQt5.QtGui import QPixmap

        height, width, _ = rgb_array.shape
        bytes_per_line = 3 * width

        qt_img = QImage(rgb_array, width, height, bytes_per_line, QImage.Format_RGB888)  # .rgbSwapped()
        pixmap = QPixmap(qt_img)

        if self.window.closed:
            self.window = None
        else:
            self.window.setPixmap(pixmap)
            self.app.processEvents()

But I get a segmentation fault when I run it. Any Idea ?

from minigrid.

I don't think that you can just pass the numpy array directly. You will need to get a pointer to the data it contains using ctypes. Something like:

import ctypes

rgb_array = np.ascontiguousarray(rgb_array)
ptr = rgb_array.ctypes.data_as(ctypes.POINTER(ctypes.c_char))
qt_img = QImage(ptr, ...)

The ascontiguousarray call is to make sure the memory is all in one contiguous block.

References:
https://docs.scipy.org/doc/numpy/reference/generated/numpy.ndarray.ctypes.html
https://docs.python.org/3/library/ctypes.html

from minigrid.