layers.py

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

# References:
#   https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
#   https://github.com/rwightman/pytorch-image-models/tree/master/timm/models/vision_transformer.py

import logging

import torch
from torch import Tensor
from torch import nn
from typing import Callable, List, Any, Tuple, Dict, Union, Optional
from torch.nn.init import trunc_normal_
from torch.nn.utils import weight_norm
import torch.nn.functional as F

logger = logging.getLogger("dinov2")

try:
    from xformers.ops import memory_efficient_attention, unbind, fmha
    from xformers.ops import scaled_index_add, index_select_cat
    XFORMERS_AVAILABLE = True
except ImportError:
    logger.warning("xFormers not available")
    XFORMERS_AVAILABLE = False

def make_2tuple(x):
    if isinstance(x, tuple):
        assert len(x) == 2
        return x

    assert isinstance(x, int)
    return (x, x)


class PatchEmbed(nn.Module):
    """
    2D image to patch embedding: (B,C,H,W) -> (B,N,D)
    Args:
        img_size: Image size.
        patch_size: Patch token size.
        in_chans: Number of input image channels.
        embed_dim: Number of linear projection output channels.
        norm_layer: Normalization layer.
    """

    def __init__(
        self,
        img_size: Union[int, Tuple[int, int]] = 224,
        patch_size: Union[int, Tuple[int, int]] = 16,
        in_chans: int = 3,
        embed_dim: int = 768,
        norm_layer: Optional[Callable] = None,
        flatten_embedding: bool = True,
    ) -> None:
        super().__init__()

        image_HW = make_2tuple(img_size)
        patch_HW = make_2tuple(patch_size)
        patch_grid_size = (
            image_HW[0] // patch_HW[0],
            image_HW[1] // patch_HW[1],
        )

        self.img_size = image_HW
        self.patch_size = patch_HW
        self.patches_resolution = patch_grid_size
        self.num_patches = patch_grid_size[0] * patch_grid_size[1]

        self.in_chans = in_chans
        self.embed_dim = embed_dim

        self.flatten_embedding = flatten_embedding

        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_HW, stride=patch_HW)
        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()

    def forward(self, x: Tensor) -> Tensor:
        _, _, H, W = x.shape
        patch_H, patch_W = self.patch_size

        assert H % patch_H == 0, f"Input image height {H} is not a multiple of patch height {patch_H}"
        assert W % patch_W == 0, f"Input image width {W} is not a multiple of patch width: {patch_W}"

        x = self.proj(x)  # B C H W
        H, W = x.size(2), x.size(3)
        x = x.flatten(2).transpose(1, 2)  # B HW C
        x = self.norm(x)
        if not self.flatten_embedding:
            x = x.reshape(-1, H, W, self.embed_dim)  # B H W C
        return x

    def flops(self) -> float:
        Ho, Wo = self.patches_resolution
        flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
        if self.norm is not None:
            flops += Ho * Wo * self.embed_dim
        return flops

def drop_path(x, drop_prob: float = 0.0, training: bool = False):
    if drop_prob == 0.0 or not training:
        return x
    keep_prob = 1 - drop_prob
    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
    if keep_prob > 0.0:
        random_tensor.div_(keep_prob)
    output = x * random_tensor
    return output

class SwiGLUFFN(nn.Module):
    def __init__(
        self,
        in_features: int,
        hidden_features: Optional[int] = None,
        out_features: Optional[int] = None,
        act_layer: Callable[..., nn.Module] = None,
        drop: float = 0.0,
        bias: bool = True,
    ) -> None:
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.w12 = nn.Linear(in_features, 2 * hidden_features, bias=bias)
        self.w3 = nn.Linear(hidden_features, out_features, bias=bias)

    def forward(self, x: Tensor) -> Tensor:
        x12 = self.w12(x)
        x1, x2 = x12.chunk(2, dim=-1)
        hidden = F.silu(x1) * x2
        return self.w3(hidden)


try:
    from xformers.ops import SwiGLU

    XFORMERS_AVAILABLE = True
except ImportError:
    SwiGLU = SwiGLUFFN
    XFORMERS_AVAILABLE = False

class SwiGLUFFNFused(SwiGLU):
    def __init__(
        self,
        in_features: int,
        hidden_features: Optional[int] = None,
        out_features: Optional[int] = None,
        act_layer: Callable[..., nn.Module] = None,
        drop: float = 0.0,
        bias: bool = True,
    ) -> None:
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        hidden_features = (int(hidden_features * 2 / 3) + 7) // 8 * 8
        super().__init__(
            in_features=in_features,
            hidden_features=hidden_features,
            out_features=out_features,
            bias=bias,
        )

class DropPath(nn.Module):
    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""

    def __init__(self, drop_prob=None):
        super(DropPath, self).__init__()
        self.drop_prob = drop_prob

    def forward(self, x):
        return drop_path(x, self.drop_prob, self.training)

class Mlp(nn.Module):
    def __init__(
        self,
        in_features: int,
        hidden_features: Optional[int] = None,
        out_features: Optional[int] = None,
        act_layer: Callable[..., nn.Module] = nn.GELU,
        drop: float = 0.0,
        bias: bool = True,
    ) -> None:
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias)
        self.drop = nn.Dropout(drop)

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

class LayerScale(nn.Module):
    def __init__(
        self,
        dim: int,
        init_values: Union[float, Tensor] = 1e-5,
        inplace: bool = False,
    ) -> None:
        super().__init__()
        self.inplace = inplace
        self.gamma = nn.Parameter(init_values * torch.ones(dim))

    def forward(self, x: Tensor) -> Tensor:
        return x.mul_(self.gamma) if self.inplace else x * self.gamma

class Attention(nn.Module):
    def __init__(
        self,
        dim: int,
        num_heads: int = 8,
        qkv_bias: bool = False,
        proj_bias: bool = True,
        attn_drop: float = 0.0,
        proj_drop: float = 0.0,
    ) -> None:
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = 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, bias=proj_bias)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x: Tensor) -> Tensor:
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)

        q, k, v = qkv[0] * self.scale, qkv[1], qkv[2]
        attn = q @ k.transpose(-2, -1)

        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

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


class MemEffAttention(Attention):
    def forward(self, x: Tensor, attn_bias=None) -> Tensor:
        if not XFORMERS_AVAILABLE:
            assert attn_bias is None, "xFormers is required for nested tensors usage"
            return super().forward(x)

        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)

        q, k, v = unbind(qkv, 2)

        if attn_bias is not None:
            self_att_op = fmha.MemoryEfficientAttentionFlashAttentionOp
        else:
            self_att_op = None
        x = memory_efficient_attention(q, k, v, attn_bias=attn_bias, op=self_att_op)
        x = x.reshape([B, N, C])

        x = self.proj(x)
        x = self.proj_drop(x)
        return x

class DINOHead(nn.Module):
    def __init__(
        self,
        in_dim,
        out_dim,
        use_bn=False,
        nlayers=3,
        hidden_dim=2048,
        bottleneck_dim=256,
        mlp_bias=True,
    ):
        super().__init__()
        nlayers = max(nlayers, 1)
        self.mlp = _build_mlp(nlayers, in_dim, bottleneck_dim, hidden_dim=hidden_dim, use_bn=use_bn, bias=mlp_bias)
        self.apply(self._init_weights)
        self.last_layer = weight_norm(nn.Linear(bottleneck_dim, out_dim, bias=False))
        self.last_layer.weight_g.data.fill_(1)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=0.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)

    def forward(self, x):
        x = self.mlp(x)
        eps = 1e-6 if x.dtype == torch.float16 else 1e-12
        x = nn.functional.normalize(x, dim=-1, p=2, eps=eps)
        x = self.last_layer(x)
        return x


def _build_mlp(nlayers, in_dim, bottleneck_dim, hidden_dim=None, use_bn=False, bias=True):
    if nlayers == 1:
        return nn.Linear(in_dim, bottleneck_dim, bias=bias)
    else:
        layers = [nn.Linear(in_dim, hidden_dim, bias=bias)]
        if use_bn:
            layers.append(nn.BatchNorm1d(hidden_dim))
        layers.append(nn.GELU())
        for _ in range(nlayers - 2):
            layers.append(nn.Linear(hidden_dim, hidden_dim, bias=bias))
            if use_bn:
                layers.append(nn.BatchNorm1d(hidden_dim))
            layers.append(nn.GELU())
        layers.append(nn.Linear(hidden_dim, bottleneck_dim, bias=bias))
        return nn.Sequential(*layers)

class Block(nn.Module):
    def __init__(
        self,
        dim: int,
        num_heads: int,
        mlp_ratio: float = 4.0,
        qkv_bias: bool = False,
        proj_bias: bool = True,
        ffn_bias: bool = True,
        drop: float = 0.0,
        attn_drop: float = 0.0,
        init_values=None,
        drop_path: float = 0.0,
        act_layer: Callable[..., nn.Module] = nn.GELU,
        norm_layer: Callable[..., nn.Module] = nn.LayerNorm,
        attn_class: Callable[..., nn.Module] = Attention,
        ffn_layer: Callable[..., nn.Module] = Mlp,
    ) -> None:
        super().__init__()
        # print(f"biases: qkv: {qkv_bias}, proj: {proj_bias}, ffn: {ffn_bias}")
        self.norm1 = norm_layer(dim)
        self.attn = attn_class(
            dim,
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            proj_bias=proj_bias,
            attn_drop=attn_drop,
            proj_drop=drop,
        )
        self.ls1 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
        self.drop_path1 = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()

        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = ffn_layer(
            in_features=dim,
            hidden_features=mlp_hidden_dim,
            act_layer=act_layer,
            drop=drop,
            bias=ffn_bias,
        )
        self.ls2 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
        self.drop_path2 = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()

        self.sample_drop_ratio = drop_path

    def forward(self, x: Tensor) -> Tensor:
        def attn_residual_func(x: Tensor) -> Tensor:
            return self.ls1(self.attn(self.norm1(x)))

        def ffn_residual_func(x: Tensor) -> Tensor:
            return self.ls2(self.mlp(self.norm2(x)))

        if self.training and self.sample_drop_ratio > 0.1:
            # the overhead is compensated only for a drop path rate larger than 0.1
            x = drop_add_residual_stochastic_depth(
                x,
                residual_func=attn_residual_func,
                sample_drop_ratio=self.sample_drop_ratio,
            )
            x = drop_add_residual_stochastic_depth(
                x,
                residual_func=ffn_residual_func,
                sample_drop_ratio=self.sample_drop_ratio,
            )
        elif self.training and self.sample_drop_ratio > 0.0:
            x = x + self.drop_path1(attn_residual_func(x))
            x = x + self.drop_path1(ffn_residual_func(x))  # FIXME: drop_path2
        else:
            x = x + attn_residual_func(x)
            x = x + ffn_residual_func(x)
        return x


def drop_add_residual_stochastic_depth(
    x: Tensor,
    residual_func: Callable[[Tensor], Tensor],
    sample_drop_ratio: float = 0.0,
) -> Tensor:
    # 1) extract subset using permutation
    b, n, d = x.shape
    sample_subset_size = max(int(b * (1 - sample_drop_ratio)), 1)
    brange = (torch.randperm(b, device=x.device))[:sample_subset_size]
    x_subset = x[brange]

    # 2) apply residual_func to get residual
    residual = residual_func(x_subset)

    x_flat = x.flatten(1)
    residual = residual.flatten(1)

    residual_scale_factor = b / sample_subset_size

    # 3) add the residual
    x_plus_residual = torch.index_add(x_flat, 0, brange, residual.to(dtype=x.dtype), alpha=residual_scale_factor)
    return x_plus_residual.view_as(x)


def get_branges_scales(x, sample_drop_ratio=0.0):
    b, n, d = x.shape
    sample_subset_size = max(int(b * (1 - sample_drop_ratio)), 1)
    brange = (torch.randperm(b, device=x.device))[:sample_subset_size]
    residual_scale_factor = b / sample_subset_size
    return brange, residual_scale_factor


def add_residual(x, brange, residual, residual_scale_factor, scaling_vector=None):
    if scaling_vector is None:
        x_flat = x.flatten(1)
        residual = residual.flatten(1)
        x_plus_residual = torch.index_add(x_flat, 0, brange, residual.to(dtype=x.dtype), alpha=residual_scale_factor)
    else:
        x_plus_residual = scaled_index_add(
            x, brange, residual.to(dtype=x.dtype), scaling=scaling_vector, alpha=residual_scale_factor
        )
    return x_plus_residual


attn_bias_cache: Dict[Tuple, Any] = {}


def get_attn_bias_and_cat(x_list, branges=None):
    """
    this will perform the index select, cat the tensors, and provide the attn_bias from cache
    """
    batch_sizes = [b.shape[0] for b in branges] if branges is not None else [x.shape[0] for x in x_list]
    all_shapes = tuple((b, x.shape[1]) for b, x in zip(batch_sizes, x_list))
    if all_shapes not in attn_bias_cache.keys():
        seqlens = []
        for b, x in zip(batch_sizes, x_list):
            for _ in range(b):
                seqlens.append(x.shape[1])
        attn_bias = fmha.BlockDiagonalMask.from_seqlens(seqlens)
        attn_bias._batch_sizes = batch_sizes
        attn_bias_cache[all_shapes] = attn_bias

    if branges is not None:
        cat_tensors = index_select_cat([x.flatten(1) for x in x_list], branges).view(1, -1, x_list[0].shape[-1])
    else:
        tensors_bs1 = tuple(x.reshape([1, -1, *x.shape[2:]]) for x in x_list)
        cat_tensors = torch.cat(tensors_bs1, dim=1)

    return attn_bias_cache[all_shapes], cat_tensors


def drop_add_residual_stochastic_depth_list(
    x_list: List[Tensor],
    residual_func: Callable[[Tensor, Any], Tensor],
    sample_drop_ratio: float = 0.0,
    scaling_vector=None,
) -> Tensor:
    # 1) generate random set of indices for dropping samples in the batch
    branges_scales = [get_branges_scales(x, sample_drop_ratio=sample_drop_ratio) for x in x_list]
    branges = [s[0] for s in branges_scales]
    residual_scale_factors = [s[1] for s in branges_scales]

    # 2) get attention bias and index+concat the tensors
    attn_bias, x_cat = get_attn_bias_and_cat(x_list, branges)

    # 3) apply residual_func to get residual, and split the result
    residual_list = attn_bias.split(residual_func(x_cat, attn_bias=attn_bias))  # type: ignore

    outputs = []
    for x, brange, residual, residual_scale_factor in zip(x_list, branges, residual_list, residual_scale_factors):
        outputs.append(add_residual(x, brange, residual, residual_scale_factor, scaling_vector).view_as(x))
    return outputs


class NestedTensorBlock(Block):
    def forward_nested(self, x_list: List[Tensor]) -> List[Tensor]:
        """
        x_list contains a list of tensors to nest together and run
        """
        assert isinstance(self.attn, MemEffAttention)

        if self.training and self.sample_drop_ratio > 0.0:

            def attn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
                return self.attn(self.norm1(x), attn_bias=attn_bias)

            def ffn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
                return self.mlp(self.norm2(x))

            x_list = drop_add_residual_stochastic_depth_list(
                x_list,
                residual_func=attn_residual_func,
                sample_drop_ratio=self.sample_drop_ratio,
                scaling_vector=self.ls1.gamma if isinstance(self.ls1, LayerScale) else None,
            )
            x_list = drop_add_residual_stochastic_depth_list(
                x_list,
                residual_func=ffn_residual_func,
                sample_drop_ratio=self.sample_drop_ratio,
                scaling_vector=self.ls2.gamma if isinstance(self.ls1, LayerScale) else None,
            )
            return x_list
        else:

            def attn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
                return self.ls1(self.attn(self.norm1(x), attn_bias=attn_bias))

            def ffn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
                return self.ls2(self.mlp(self.norm2(x)))

            attn_bias, x = get_attn_bias_and_cat(x_list)
            x = x + attn_residual_func(x, attn_bias=attn_bias)
            x = x + ffn_residual_func(x)
            return attn_bias.split(x)

    def forward(self, x_or_x_list):
        if isinstance(x_or_x_list, Tensor):
            return super().forward(x_or_x_list)
        elif isinstance(x_or_x_list, list):
            assert XFORMERS_AVAILABLE, "Please install xFormers for nested tensors usage"
            return self.forward_nested(x_or_x_list)
        else:
            raise AssertionError