How do pixel unshuffle in ggml ?

[delldu]

Dear friends, We have following pytorch code in esrgan network, hope to used in ggml, would you like show us how to do in ggml ? As you know, ggml tensor default maximum dims is 4, but here view and permute is 6. We need change GGML_MAX_DIMS to 6 ?

def pixel_unshuffle(x, scale: int):
    """Pixel unshuffle."""
    b, c, hh, hw = x.size()
    out_channel = int(c * (scale ** 2))
    assert hh % scale == 0 and hw % scale == 0
    h = hh // scale
    w = hw // scale
    x_view = x.view(b, c, h, scale, w, scale)
    x_view = x_view.permute(0, 1, 3, 5, 2, 4)

    return x_view.reshape(b, out_channel, h, w)

[ggerganov]

Maybe as a sequence of views and 4D permutes:

x_view = x.view(b*c, h, scale, w*scale)
x_view = x_view.permute(0, 2, 1, 3)

x_view = x.view(b*c*scale, h, w, scale)
x_view = x_view.permute(0, 3, 1, 2)

return x_view.reshape(b, out_channel, h, w)

(not sure if this is correct so verify)

If it is correct, it would be something like this in ggml:

x = ggml_reshape_4d(ctx, x, w*scale, scale, h, b*c);
x = ggml_permute   (ctx, x, 0, 2, 1, 3);
x = ggml_cont      (ctx, x);

x = ggml_reshape_4d(ctx, x, scale, w, h, b*c*scale);
x = ggml_permute   (ctx, x, 2, 0, 1, 3);
x = ggml_cont      (ctx, x);

x = ggml_reshape_4d(ctx, x, w, h, out_channel, b);

[delldu]

Thanks@ggerganov , we tested, the answer is not correct.

[farris]

Thanks@ggerganov , we tested, the answer is not correct.

How did you work around this?