detection/backbone/xcit.py [26:447]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
class PositionalEncodingFourier(nn.Module):
    """
    Positional encoding relying on a fourier kernel matching the one used in the
    "Attention is all of Need" paper. The implementation builds on DeTR code
    https://github.com/facebookresearch/detr/blob/master/models/position_encoding.py
    """

    def __init__(self, hidden_dim=32, dim=768, temperature=10000):
        super().__init__()
        self.token_projection = nn.Conv2d(hidden_dim * 2, dim, kernel_size=1)
        self.scale = 2 * math.pi
        self.temperature = temperature
        self.hidden_dim = hidden_dim
        self.dim = dim

    def forward(self, B, H, W):
        mask = torch.zeros(B, H, W).bool().to(self.token_projection.weight.device)
        not_mask = ~mask
        y_embed = not_mask.cumsum(1, dtype=torch.float32)
        x_embed = not_mask.cumsum(2, dtype=torch.float32)
        eps = 1e-6
        y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
        x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale

        dim_t = torch.arange(self.hidden_dim, dtype=torch.float32, device=mask.device)
        dim_t = self.temperature ** (2 * (dim_t // 2) / self.hidden_dim)

        pos_x = x_embed[:, :, :, None] / dim_t
        pos_y = y_embed[:, :, :, None] / dim_t
        pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(),
                             pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
        pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(),
                             pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
        pos = self.token_projection(pos)
        return pos


def conv3x3(in_planes, out_planes, stride=1):
    """3x3 convolution with padding"""
    return torch.nn.Sequential(
        nn.Conv2d(
            in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False
        ),
        nn.SyncBatchNorm(out_planes)
    )


class ConvPatchEmbed(nn.Module):
    """ Image to Patch Embedding using multiple convolutional layers
    """

    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
        self.img_size = img_size
        self.patch_size = patch_size
        self.num_patches = num_patches

        if patch_size[0] == 16:
            self.proj = torch.nn.Sequential(
                conv3x3(3, embed_dim // 8, 2),
                nn.GELU(),
                conv3x3(embed_dim // 8, embed_dim // 4, 2),
                nn.GELU(),
                conv3x3(embed_dim // 4, embed_dim // 2, 2),
                nn.GELU(),
                conv3x3(embed_dim // 2, embed_dim, 2),
            )
        elif patch_size[0] == 8:
            self.proj = torch.nn.Sequential(
                conv3x3(3, embed_dim // 4, 2),
                nn.GELU(),
                conv3x3(embed_dim // 4, embed_dim // 2, 2),
                nn.GELU(),
                conv3x3(embed_dim // 2, embed_dim, 2),
            )
        else:
            raise("For convolutional projection, patch size has to be in [8, 16]")

    def forward(self, x, padding_size=None):
        B, C, H, W = x.shape
        x = self.proj(x)
        Hp, Wp = x.shape[2], x.shape[3]
        x = x.flatten(2).transpose(1, 2)

        return x, (Hp, Wp)


class LPI(nn.Module):
    """
    Local Patch Interaction module that allows explicit communication between tokens in 3x3 windows
    to augment the implicit communcation performed by the block diagonal scatter attention.
    Implemented using 2 layers of separable 3x3 convolutions with GeLU and BatchNorm2d
    """

    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU,
                 drop=0., kernel_size=3):
        super().__init__()
        out_features = out_features or in_features

        padding = kernel_size // 2

        self.conv1 = torch.nn.Conv2d(in_features, out_features, kernel_size=kernel_size,
                                     padding=padding, groups=out_features)
        self.act = act_layer()
        self.bn = nn.SyncBatchNorm(in_features)
        self.conv2 = torch.nn.Conv2d(in_features, out_features, kernel_size=kernel_size,
                                     padding=padding, groups=out_features)

    def forward(self, x, H, W):
        B, N, C = x.shape
        x = x.permute(0, 2, 1).reshape(B, C, H, W)
        x = self.conv1(x)
        x = self.act(x)
        x = self.bn(x)
        x = self.conv2(x)
        x = x.reshape(B, C, N).permute(0, 2, 1)

        return x


class ClassAttention(nn.Module):
    """Class Attention Layer as in CaiT https://arxiv.org/abs/2103.17239
    """

    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5

        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)

    def forward(self, x):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
        qkv = qkv.permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)

        qc = q[:, :, 0:1]   # CLS token
        attn_cls = (qc * k).sum(dim=-1) * self.scale
        attn_cls = attn_cls.softmax(dim=-1)
        attn_cls = self.attn_drop(attn_cls)

        cls_tkn = (attn_cls.unsqueeze(2) @ v).transpose(1, 2).reshape(B, 1, C)
        cls_tkn = self.proj(cls_tkn)
        x = torch.cat([self.proj_drop(cls_tkn), x[:, 1:]], dim=1)
        return x


class ClassAttentionBlock(nn.Module):
    """Class Attention Layer as in CaiT https://arxiv.org/abs/2103.17239
    """

    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0.,
                 attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, eta=None,
                 tokens_norm=False):
        super().__init__()
        self.norm1 = norm_layer(dim)

        self.attn = ClassAttention(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop,
            proj_drop=drop
        )

        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer,
                       drop=drop)

        if eta is not None:     # LayerScale Initialization (no layerscale when None)
            self.gamma1 = nn.Parameter(eta * torch.ones(dim), requires_grad=True)
            self.gamma2 = nn.Parameter(eta * torch.ones(dim), requires_grad=True)
        else:
            self.gamma1, self.gamma2 = 1.0, 1.0

        # FIXME: A hack for models pre-trained with layernorm over all the tokens not just the CLS
        self.tokens_norm = tokens_norm

    def forward(self, x, H, W, mask=None):
        x = x + self.drop_path(self.gamma1 * self.attn(self.norm1(x)))
        if self.tokens_norm:
            x = self.norm2(x)
        else:
            x[:, 0:1] = self.norm2(x[:, 0:1])

        x_res = x
        cls_token = x[:, 0:1]
        cls_token = self.gamma2 * self.mlp(cls_token)
        x = torch.cat([cls_token, x[:, 1:]], dim=1)
        x = x_res + self.drop_path(x)
        return x


class XCA(nn.Module):
    """ Cross-Covariance Attention (XCA) operation where the channels are updated using a weighted
     sum. The weights are obtained from the (softmax normalized) Cross-covariance
    matrix (Q^T K \\in d_h \\times d_h)
    """

    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
        super().__init__()
        self.num_heads = num_heads
        self.temperature = nn.Parameter(torch.ones(num_heads, 1, 1))

        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)

    def forward(self, x):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
        qkv = qkv.permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)

        q = q.transpose(-2, -1)
        k = k.transpose(-2, -1)
        v = v.transpose(-2, -1)

        q = torch.nn.functional.normalize(q, dim=-1)
        k = torch.nn.functional.normalize(k, dim=-1)

        attn = (q @ k.transpose(-2, -1)) * self.temperature
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        x = (attn @ v).permute(0, 3, 1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'temperature'}


class XCABlock(nn.Module):
    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0.,
                 attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm,
                 num_tokens=196, eta=None):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = XCA(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop,
            proj_drop=drop
        )
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)

        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer,
                       drop=drop)

        self.norm3 = norm_layer(dim)
        self.local_mp = LPI(in_features=dim, act_layer=act_layer)

        self.gamma1 = nn.Parameter(eta * torch.ones(dim), requires_grad=True)
        self.gamma2 = nn.Parameter(eta * torch.ones(dim), requires_grad=True)
        self.gamma3 = nn.Parameter(eta * torch.ones(dim), requires_grad=True)

    def forward(self, x, H, W):
        x = x + self.drop_path(self.gamma1 * self.attn(self.norm1(x)))
        x = x + self.drop_path(self.gamma3 * self.local_mp(self.norm3(x), H, W))
        x = x + self.drop_path(self.gamma2 * self.mlp(self.norm2(x)))
        return x


@BACKBONES.register_module()
class XCiT(nn.Module):
    """
    Based on timm and DeiT code bases
    https://github.com/rwightman/pytorch-image-models/tree/master/timm
    https://github.com/facebookresearch/deit/
    """

    def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=80, embed_dim=768,
                 depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None,
                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=None,
                 cls_attn_layers=2, use_pos=True, eta=None, tokens_norm=False,
                 out_indices=[3, 5, 7, 11]):
        """
        Args:
            img_size (int, tuple): input image size
            patch_size (int, tuple): patch size
            in_chans (int): number of input channels
            num_classes (int): number of classes for classification head
            embed_dim (int): embedding dimension
            depth (int): depth of transformer
            num_heads (int): number of attention heads
            mlp_ratio (int): ratio of mlp hidden dim to embedding dim
            qkv_bias (bool): enable bias for qkv if True
            qk_scale (float): override default qk scale of head_dim ** -0.5 if set
            drop_rate (float): dropout rate
            attn_drop_rate (float): attention dropout rate
            drop_path_rate (float): stochastic depth rate
            norm_layer: (nn.Module): normalization layer
            cls_attn_layers: (int) Depth of Class attention layers
            use_pos: (bool) whether to use positional encoding
            eta: (float) layerscale initialization value
            tokens_norm: (bool) Whether to normalize all tokens or just the cls_token in the CA
            out_indices: (list) Indices of layers from which FPN features are extracted
        """
        super().__init__()
        self.num_classes = num_classes
        self.num_features = self.embed_dim = embed_dim
        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)

        self.patch_embed = ConvPatchEmbed(img_size=img_size, embed_dim=embed_dim,
                                          patch_size=patch_size)

        num_patches = self.patch_embed.num_patches

        self.pos_drop = nn.Dropout(p=drop_rate)

        dpr = [drop_path_rate for i in range(depth)]
        self.blocks = nn.ModuleList([
            XCABlock(
                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias,
                qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i],
                norm_layer=norm_layer, num_tokens=num_patches, eta=eta)
            for i in range(depth)])

        self.pos_embeder = PositionalEncodingFourier(dim=embed_dim)
        self.use_pos = use_pos

        self.out_indices = out_indices

        if patch_size == 16:
            self.fpn1 = nn.Sequential(
                nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2),
                nn.SyncBatchNorm(embed_dim),
                nn.GELU(),
                nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2),
            )

            self.fpn2 = nn.Sequential(
                nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2),
            )

            self.fpn3 = nn.Identity()

            self.fpn4 = nn.MaxPool2d(kernel_size=2, stride=2)
        elif patch_size == 8:
            self.fpn1 = nn.Sequential(
                nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2),
            )

            self.fpn2 = nn.Identity()

            self.fpn3 = nn.Sequential(
                nn.MaxPool2d(kernel_size=2, stride=2),
            )

            self.fpn4 = nn.Sequential(
                nn.MaxPool2d(kernel_size=4, stride=4),
            )

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'pos_embed', 'cls_token', 'dist_token'}

    def init_weights(self, pretrained=None):
        """Initialize the weights in backbone.

        Args:
            pretrained (str, optional): Path to pre-trained weights.
                Defaults to None.
        """

        def _init_weights(m):
            if isinstance(m, nn.Linear):
                trunc_normal_(m.weight, std=.02)
                if isinstance(m, nn.Linear) and m.bias is not None:
                    nn.init.constant_(m.bias, 0)
            elif isinstance(m, nn.LayerNorm):
                nn.init.constant_(m.bias, 0)
                nn.init.constant_(m.weight, 1.0)

        if isinstance(pretrained, str):
            self.apply(_init_weights)
            logger = get_root_logger()
            load_checkpoint(self, pretrained, strict=False, logger=logger)
        elif pretrained is None:
            self.apply(_init_weights)
        else:
            raise TypeError('pretrained must be a str or None')

    def forward_features(self, x):
        B, C, H, W = x.shape
        x, (Hp, Wp) = self.patch_embed(x)

        pos_encoding = self.pos_embeder(B, Hp, Wp).reshape(B, -1, x.shape[1]).permute(0, 2, 1)

        if self.use_pos:
            x = x + pos_encoding

        x = self.pos_drop(x)

        features = []
        for i, blk in enumerate(self.blocks):
            x = blk(x, Hp, Wp)
            if i in self.out_indices:
                xp = x.permute(0, 2, 1).reshape(B, -1, Hp, Wp)
                features.append(xp)

        ops = [self.fpn1, self.fpn2, self.fpn3, self.fpn4]
        for i in range(len(features)):
            features[i] = ops[i](features[i])

        return tuple(features)

    def forward(self, x):
        x = self.forward_features(x)
        return x
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



semantic_segmentation/backbone/xcit.py [26:447]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
class PositionalEncodingFourier(nn.Module):
    """
    Positional encoding relying on a fourier kernel matching the one used in the
    "Attention is all of Need" paper. The implementation builds on DeTR code
    https://github.com/facebookresearch/detr/blob/master/models/position_encoding.py
    """

    def __init__(self, hidden_dim=32, dim=768, temperature=10000):
        super().__init__()
        self.token_projection = nn.Conv2d(hidden_dim * 2, dim, kernel_size=1)
        self.scale = 2 * math.pi
        self.temperature = temperature
        self.hidden_dim = hidden_dim
        self.dim = dim

    def forward(self, B, H, W):
        mask = torch.zeros(B, H, W).bool().to(self.token_projection.weight.device)
        not_mask = ~mask
        y_embed = not_mask.cumsum(1, dtype=torch.float32)
        x_embed = not_mask.cumsum(2, dtype=torch.float32)
        eps = 1e-6
        y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
        x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale

        dim_t = torch.arange(self.hidden_dim, dtype=torch.float32, device=mask.device)
        dim_t = self.temperature ** (2 * (dim_t // 2) / self.hidden_dim)

        pos_x = x_embed[:, :, :, None] / dim_t
        pos_y = y_embed[:, :, :, None] / dim_t
        pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(),
                             pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
        pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(),
                             pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
        pos = self.token_projection(pos)
        return pos


def conv3x3(in_planes, out_planes, stride=1):
    """3x3 convolution with padding"""
    return torch.nn.Sequential(
        nn.Conv2d(
            in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False
        ),
        nn.SyncBatchNorm(out_planes)
    )


class ConvPatchEmbed(nn.Module):
    """ Image to Patch Embedding using multiple convolutional layers
    """

    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
        self.img_size = img_size
        self.patch_size = patch_size
        self.num_patches = num_patches

        if patch_size[0] == 16:
            self.proj = torch.nn.Sequential(
                conv3x3(3, embed_dim // 8, 2),
                nn.GELU(),
                conv3x3(embed_dim // 8, embed_dim // 4, 2),
                nn.GELU(),
                conv3x3(embed_dim // 4, embed_dim // 2, 2),
                nn.GELU(),
                conv3x3(embed_dim // 2, embed_dim, 2),
            )
        elif patch_size[0] == 8:
            self.proj = torch.nn.Sequential(
                conv3x3(3, embed_dim // 4, 2),
                nn.GELU(),
                conv3x3(embed_dim // 4, embed_dim // 2, 2),
                nn.GELU(),
                conv3x3(embed_dim // 2, embed_dim, 2),
            )
        else:
            raise("For convolutional projection, patch size has to be in [8, 16]")

    def forward(self, x, padding_size=None):
        B, C, H, W = x.shape
        x = self.proj(x)
        Hp, Wp = x.shape[2], x.shape[3]
        x = x.flatten(2).transpose(1, 2)

        return x, (Hp, Wp)


class LPI(nn.Module):
    """
    Local Patch Interaction module that allows explicit communication between tokens in 3x3 windows
    to augment the implicit communcation performed by the block diagonal scatter attention.
    Implemented using 2 layers of separable 3x3 convolutions with GeLU and BatchNorm2d
    """

    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU,
                 drop=0., kernel_size=3):
        super().__init__()
        out_features = out_features or in_features

        padding = kernel_size // 2

        self.conv1 = torch.nn.Conv2d(in_features, out_features, kernel_size=kernel_size,
                                     padding=padding, groups=out_features)
        self.act = act_layer()
        self.bn = nn.SyncBatchNorm(in_features)
        self.conv2 = torch.nn.Conv2d(in_features, out_features, kernel_size=kernel_size,
                                     padding=padding, groups=out_features)

    def forward(self, x, H, W):
        B, N, C = x.shape
        x = x.permute(0, 2, 1).reshape(B, C, H, W)
        x = self.conv1(x)
        x = self.act(x)
        x = self.bn(x)
        x = self.conv2(x)
        x = x.reshape(B, C, N).permute(0, 2, 1)

        return x


class ClassAttention(nn.Module):
    """Class Attention Layer as in CaiT https://arxiv.org/abs/2103.17239
    """

    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5

        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)

    def forward(self, x):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
        qkv = qkv.permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)

        qc = q[:, :, 0:1]   # CLS token
        attn_cls = (qc * k).sum(dim=-1) * self.scale
        attn_cls = attn_cls.softmax(dim=-1)
        attn_cls = self.attn_drop(attn_cls)

        cls_tkn = (attn_cls.unsqueeze(2) @ v).transpose(1, 2).reshape(B, 1, C)
        cls_tkn = self.proj(cls_tkn)
        x = torch.cat([self.proj_drop(cls_tkn), x[:, 1:]], dim=1)
        return x


class ClassAttentionBlock(nn.Module):
    """Class Attention Layer as in CaiT https://arxiv.org/abs/2103.17239
    """

    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0.,
                 attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, eta=None,
                 tokens_norm=False):
        super().__init__()
        self.norm1 = norm_layer(dim)

        self.attn = ClassAttention(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop,
            proj_drop=drop
        )

        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer,
                       drop=drop)

        if eta is not None:     # LayerScale Initialization (no layerscale when None)
            self.gamma1 = nn.Parameter(eta * torch.ones(dim), requires_grad=True)
            self.gamma2 = nn.Parameter(eta * torch.ones(dim), requires_grad=True)
        else:
            self.gamma1, self.gamma2 = 1.0, 1.0

        # FIXME: A hack for models pre-trained with layernorm over all the tokens not just the CLS
        self.tokens_norm = tokens_norm

    def forward(self, x, H, W, mask=None):
        x = x + self.drop_path(self.gamma1 * self.attn(self.norm1(x)))
        if self.tokens_norm:
            x = self.norm2(x)
        else:
            x[:, 0:1] = self.norm2(x[:, 0:1])

        x_res = x
        cls_token = x[:, 0:1]
        cls_token = self.gamma2 * self.mlp(cls_token)
        x = torch.cat([cls_token, x[:, 1:]], dim=1)
        x = x_res + self.drop_path(x)
        return x


class XCA(nn.Module):
    """ Cross-Covariance Attention (XCA) operation where the channels are updated using a weighted
     sum. The weights are obtained from the (softmax normalized) Cross-covariance
    matrix (Q^T K \\in d_h \\times d_h)
    """

    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
        super().__init__()
        self.num_heads = num_heads
        self.temperature = nn.Parameter(torch.ones(num_heads, 1, 1))

        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)

    def forward(self, x):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
        qkv = qkv.permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)

        q = q.transpose(-2, -1)
        k = k.transpose(-2, -1)
        v = v.transpose(-2, -1)

        q = torch.nn.functional.normalize(q, dim=-1)
        k = torch.nn.functional.normalize(k, dim=-1)

        attn = (q @ k.transpose(-2, -1)) * self.temperature
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        x = (attn @ v).permute(0, 3, 1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'temperature'}


class XCABlock(nn.Module):
    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0.,
                 attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm,
                 num_tokens=196, eta=None):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = XCA(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop,
            proj_drop=drop
        )
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)

        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer,
                       drop=drop)

        self.norm3 = norm_layer(dim)
        self.local_mp = LPI(in_features=dim, act_layer=act_layer)

        self.gamma1 = nn.Parameter(eta * torch.ones(dim), requires_grad=True)
        self.gamma2 = nn.Parameter(eta * torch.ones(dim), requires_grad=True)
        self.gamma3 = nn.Parameter(eta * torch.ones(dim), requires_grad=True)

    def forward(self, x, H, W):
        x = x + self.drop_path(self.gamma1 * self.attn(self.norm1(x)))
        x = x + self.drop_path(self.gamma3 * self.local_mp(self.norm3(x), H, W))
        x = x + self.drop_path(self.gamma2 * self.mlp(self.norm2(x)))
        return x


@BACKBONES.register_module()
class XCiT(nn.Module):
    """
    Based on timm and DeiT code bases
    https://github.com/rwightman/pytorch-image-models/tree/master/timm
    https://github.com/facebookresearch/deit/
    """

    def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=80, embed_dim=768,
                 depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None,
                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=None,
                 cls_attn_layers=2, use_pos=True, eta=None, tokens_norm=False,
                 out_indices=[3, 5, 7, 11]):
        """
        Args:
            img_size (int, tuple): input image size
            patch_size (int, tuple): patch size
            in_chans (int): number of input channels
            num_classes (int): number of classes for classification head
            embed_dim (int): embedding dimension
            depth (int): depth of transformer
            num_heads (int): number of attention heads
            mlp_ratio (int): ratio of mlp hidden dim to embedding dim
            qkv_bias (bool): enable bias for qkv if True
            qk_scale (float): override default qk scale of head_dim ** -0.5 if set
            drop_rate (float): dropout rate
            attn_drop_rate (float): attention dropout rate
            drop_path_rate (float): stochastic depth rate
            norm_layer: (nn.Module): normalization layer
            cls_attn_layers: (int) Depth of Class attention layers
            use_pos: (bool) whether to use positional encoding
            eta: (float) layerscale initialization value
            tokens_norm: (bool) Whether to normalize all tokens or just the cls_token in the CA
            out_indices: (list) Indices of layers from which FPN features are extracted
        """
        super().__init__()
        self.num_classes = num_classes
        self.num_features = self.embed_dim = embed_dim
        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)

        self.patch_embed = ConvPatchEmbed(img_size=img_size, embed_dim=embed_dim,
                                          patch_size=patch_size)

        num_patches = self.patch_embed.num_patches

        self.pos_drop = nn.Dropout(p=drop_rate)

        dpr = [drop_path_rate for i in range(depth)]
        self.blocks = nn.ModuleList([
            XCABlock(
                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias,
                qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i],
                norm_layer=norm_layer, num_tokens=num_patches, eta=eta)
            for i in range(depth)])

        self.pos_embeder = PositionalEncodingFourier(dim=embed_dim)
        self.use_pos = use_pos

        self.out_indices = out_indices

        if patch_size == 16:
            self.fpn1 = nn.Sequential(
                nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2),
                nn.SyncBatchNorm(embed_dim),
                nn.GELU(),
                nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2),
            )

            self.fpn2 = nn.Sequential(
                nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2),
            )

            self.fpn3 = nn.Identity()

            self.fpn4 = nn.MaxPool2d(kernel_size=2, stride=2)
        elif patch_size == 8:
            self.fpn1 = nn.Sequential(
                nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2),
            )

            self.fpn2 = nn.Identity()

            self.fpn3 = nn.Sequential(
                nn.MaxPool2d(kernel_size=2, stride=2),
            )

            self.fpn4 = nn.Sequential(
                nn.MaxPool2d(kernel_size=4, stride=4),
            )

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'pos_embed', 'cls_token', 'dist_token'}

    def init_weights(self, pretrained=None):
        """Initialize the weights in backbone.

        Args:
            pretrained (str, optional): Path to pre-trained weights.
                Defaults to None.
        """

        def _init_weights(m):
            if isinstance(m, nn.Linear):
                trunc_normal_(m.weight, std=.02)
                if isinstance(m, nn.Linear) and m.bias is not None:
                    nn.init.constant_(m.bias, 0)
            elif isinstance(m, nn.LayerNorm):
                nn.init.constant_(m.bias, 0)
                nn.init.constant_(m.weight, 1.0)

        if isinstance(pretrained, str):
            self.apply(_init_weights)
            logger = get_root_logger()
            load_checkpoint(self, pretrained, strict=False, logger=logger)
        elif pretrained is None:
            self.apply(_init_weights)
        else:
            raise TypeError('pretrained must be a str or None')

    def forward_features(self, x):
        B, C, H, W = x.shape
        x, (Hp, Wp) = self.patch_embed(x)

        pos_encoding = self.pos_embeder(B, Hp, Wp).reshape(B, -1, x.shape[1]).permute(0, 2, 1)

        if self.use_pos:
            x = x + pos_encoding

        x = self.pos_drop(x)

        features = []
        for i, blk in enumerate(self.blocks):
            x = blk(x, Hp, Wp)
            if i in self.out_indices:
                xp = x.permute(0, 2, 1).reshape(B, -1, Hp, Wp)
                features.append(xp)

        ops = [self.fpn1, self.fpn2, self.fpn3, self.fpn4]
        for i in range(len(features)):
            features[i] = ops[i](features[i])

        return tuple(features)

    def forward(self, x):
        x = self.forward_features(x)
        return x
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -