src/masks/utils.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

import torch


def apply_masks(x, masks, concat=True):
    """
    :param x: tensor of shape [B (batch-size), N (num-patches), D (feature-dim)]
    :param masks: list of tensors of shape [B, K] containing indices of K patches in [N] to keep
    """
    all_x = []
    for m in masks:
        mask_keep = m.unsqueeze(-1).repeat(1, 1, x.size(-1))
        all_x += [torch.gather(x, dim=1, index=mask_keep)]
    if not concat:
        return all_x

    return torch.cat(all_x, dim=0)


def _list_of_index_tensors_to_bool_mask(idxs: list[torch.Tensor],
                                        B: int,
                                        N: int,
                                        device: torch.device) -> torch.BoolTensor:
    """
    idxs  : list of 1×Ni or Ni×1 tensors of indices (one per sample but any length)
    B, N  : wanted mask shape  [B, N]
    """
    mask = torch.zeros(B, N, dtype=torch.bool, device=device)
    for i, t in enumerate(idxs):
        if i >= B:           # safety guard
            break
        mask[i, t.view(-1).clamp_(0, N - 1)] = True
    return mask