基于mmaction2的Video-Swin-Transformer源码

网络结构图：

位置：Video-Swin-Transformer/mmaction/models/backbones/swin_transformer.py

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint
import numpy as np
from timm.models.layers import DropPath, trunc_normal_

from mmcv.runner import load_checkpoint
from mmaction.utils import get_root_logger
from ..builder import BACKBONES

from functools import reduce, lru_cache
from operator import mul
from einops import rearrange

多层感知器：

class Mlp(nn.Module):
    """ Multilayer perceptron."""

    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x

窗口划分、逆转、尺寸：

def window_partition(x, window_size):
    """
    Args:
        x: (B, D, H, W, C)
        window_size (tuple[int]): window size

    Returns:
        windows: (B*num_windows, window_size*window_size, C)
    """
    B, D, H, W, C = x.shape
    x = x.view(B, D // window_size[0], window_size[0], H // window_size[1], window_size[1], W // window_size[2], window_size[2], C)
    windows = x.permute(0, 1, 3, 5, 2, 4, 6, 7).contiguous().view(-1, reduce(mul, window_size), C)
    return windows


def window_reverse(windows, window_size, B, D, H, W):
    """
    Args:
        windows: (B*num_windows, window_size, window_size, C)
        window_size (tuple[int]): Window size
        H (int): Height of image
        W (int): Width of image

    Returns:
        x: (B, D, H, W, C)
    """
    x = windows.view(B, D // window_size[0], H // window_size[1], W // window_size[2], window_size[0], window_size[1], window_size[2], -1)
    x = x.permute(0, 1, 4, 2, 5, 3, 6, 7).contiguous().view(B, D, H, W, -1)
    return x




def get_window_size(x_size, window_size, shift_size=None):
    use_window_size = list(window_size)
    if shift_size is not None:
        use_shift_size = list(shift_size)
    for i in range(len(x_size)):
        if x_size[i] <= window_size[i]:
            use_window_size[i] = x_size[i]
            if shift_size is not None:
                use_shift_size[i] = 0

    if shift_size is None:
        return tuple(use_window_size)
    else:
        return tuple(use_window_size), tuple(use_shift_size)

W-MSA：具有相对位置偏差的基于窗口的多头自注意力模块

class WindowAttention3D(nn.Module):
    """ Window based multi-head self attention (W-MSA) module with relative position bias.
    It supports both of shifted and non-shifted window.
    Args:
        dim (int): Number of input channels.
        window_size (tuple[int]): The temporal length, height and width of the window.
        num_heads (int): Number of attention heads.
        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
    """

    def __init__(self, dim, window_size, num_heads, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):

        super().__init__()
        self.dim = dim
        self.window_size = window_size  # Wd, Wh, Ww
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5

        # define a parameter table of relative position bias
        self.relative_position_bias_table = nn.Parameter(
            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1) * (2 * window_size[2] - 1), num_heads))  # 2*Wd-1 * 2*Wh-1 * 2*Ww-1, nH

        # get pair-wise relative position index for each token inside the window
        coords_d = torch.arange(self.window_size[0])
        coords_h = torch.arange(self.window_size[1])
        coords_w = torch.arange(self.window_size[2])
        coords = torch.stack(torch.meshgrid(coords_d, coords_h, coords_w))  # 3, Wd, Wh, Ww
        coords_flatten = torch.flatten(coords, 1)  # 3, Wd*Wh*Ww
        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 3, Wd*Wh*Ww, Wd*Wh*Ww
        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wd*Wh*Ww, Wd*Wh*Ww, 3
        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
        relative_coords[:, :, 1] += self.window_size[1] - 1
        relative_coords[:, :, 2] += self.window_size[2] - 1

        relative_coords[:, :, 0] *= (2 * self.window_size[1] - 1) * (2 * self.window_size[2] - 1)
        relative_coords[:, :, 1] *= (2 * self.window_size[2] - 1)
        relative_position_index = relative_coords.sum(-1)  # Wd*Wh*Ww, Wd*Wh*Ww
        self.register_buffer("relative_position_index", relative_position_index)

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

        trunc_norm