Function resample_patch_embed

src/misc/weight_modify.py:13–84

Resample the weights of the patch embedding kernel to target resolution. We resample the patch embedding kernel by approximately inverting the effect of patch resizing. Code based on: https://github.com/google-research/big_vision/blob/b00544b81f8694488d5f36295aeb7972f3755ffe/big_v

(
        patch_embed,
        new_size: List[int],
        interpolation: str = 'bicubic',
        antialias: bool = True,
        verbose: bool = False,
)

def resample_patch_embed(
        patch_embed,
        new_size: List[int],
        interpolation: str = 'bicubic',
        antialias: bool = True,
        verbose: bool = False,
):
    """Resample the weights of the patch embedding kernel to target resolution.
    We resample the patch embedding kernel by approximately inverting the effect
    of patch resizing.

    Code based on:
      https://github.com/google-research/big_vision/blob/b00544b81f8694488d5f36295aeb7972f3755ffe/big_vision/models/proj/flexi/vit.py

    With this resizing, we can for example load a B/8 filter into a B/16 model
    and, on 2x larger input image, the result will match.

    Args:
        patch_embed: original parameter to be resized.
        new_size (tuple(int, int): target shape (height, width)-only.
        interpolation (str): interpolation for resize
        antialias (bool): use anti-aliasing filter in resize
        verbose (bool): log operation
    Returns:
        Resized patch embedding kernel.
    """
    import numpy as np
    try:
        import functorch
        vmap = functorch.vmap
    except ImportError:
        if hasattr(torch, 'vmap'):
            vmap = torch.vmap
        else:
            assert False, "functorch or a version of torch with vmap is required for FlexiViT resizing."

    assert len(patch_embed.shape) == 4, "Four dimensions expected"
    assert len(new_size) == 2, "New shape should only be hw"
    old_size = patch_embed.shape[-2:]
    if tuple(old_size) == tuple(new_size):
        return patch_embed

    if verbose:
        _logger.info(f"Resize patch embedding {patch_embed.shape} to {new_size}, w/ {interpolation} interpolation.")

    def resize(x_np, _new_size):
        x_tf = torch.Tensor(x_np)[None, None, ...]
        x_upsampled = F.interpolate(
            x_tf, size=_new_size, mode=interpolation, antialias=antialias)[0, 0, ...].numpy()
        return x_upsampled

    def get_resize_mat(_old_size, _new_size):
        mat = []
        for i in range(np.prod(_old_size)):
            basis_vec = np.zeros(_old_size)
            basis_vec[np.unravel_index(i, _old_size)] = 1.
            mat.append(resize(basis_vec, _new_size).reshape(-1))
        return np.stack(mat).T