quant_swin_transformer.py

""" Swin Transformer
A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`
    - https://arxiv.org/pdf/2103.14030

Code/weights from https://github.com/microsoft/Swin-Transformer, original copyright/license info below

S3 (AutoFormerV2, https://arxiv.org/abs/2111.14725) Swin weights from
    - https://github.com/microsoft/Cream/tree/main/AutoFormerV2

Modifications and additions for timm hacked together by / Copyright 2021, Ross Wightman
"""
# --------------------------------------------------------
# Swin Transformer
# Copyright (c) 2021 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Written by Ze Liu
# --------------------------------------------------------
import logging
import math
from typing import Callable, List, Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F

from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from classifier import ClassifierHead
from timm.layers import PatchEmbed, Mlp, DropPath, to_2tuple, to_ntuple, trunc_normal_, \
    _assert, use_fused_attn, resize_rel_pos_bias_table, resample_patch_embed
from timm.models._builder import build_model_with_cfg
from timm.models._features_fx import register_notrace_function
from timm.models._manipulate import checkpoint_seq, named_apply
from timm.models._registry import generate_default_cfgs, register_model, register_model_deprecations
from timm.models.vision_transformer import get_init_weights_vit
from timm.layers.format import Format, nchw_to
from Quant import *
from _quan_base import *

__all__ = ['SwinTransformer']  # model_registry will add each entrypoint fn to this

_logger = logging.getLogger(__name__)

_int_or_tuple_2_t = Union[int, Tuple[int, int]]

# 把x按照windowsize分割成 [B * num_windows, window_size, window_size, C]
def window_partition(
        x: torch.Tensor,
        window_size: Tuple[int, int],
) -> torch.Tensor:
    """
    Partition into non-overlapping windows with padding if needed.
    Args:
        x (tensor): input tokens with [B, H, W, C].
        window_size (int): window size.

    Returns:
        windows: windows after partition with [B * num_windows, window_size, window_size, C].
        (Hp, Wp): padded height and width before partition
    """
    B, H, W, C = x.shape
    x = x.view(B, H // window_size[0], window_size[0], W // window_size[1], window_size[1], C)
    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size[0], window_size[1], C)
    #       (B, H // window_size[0], W // window_size[1], window_size[0], window_size[1], C)
    #     ->(B * num_windows, window_size, window_size, C)
    return windows


@register_notrace_function  # reason: int argument is a Proxy
# 重新组合成原始图像的形状（B, H, W, C）？ 还是特征图
def window_reverse(windows, window_size: Tuple[int, int], H: int, W: int):
    """
    Args:
        windows: (num_windows*B, window_size, window_size, C)
        window_size (int): Window size
        H (int): Height of image
        W (int): Width of image

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

# 计算每个窗口中每个像素之间的相对位置索引
def get_relative_position_index(win_h: int, win_w: int):
    # get pair-wise relative position index for each token inside the window
    coords = torch.stack(torch.meshgrid([torch.arange(win_h), torch.arange(win_w)]))  # 2, Wh, Ww
    coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
    relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
    relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
    relative_coords[:, :, 0] += win_h - 1  # shift to start from 0
    relative_coords[:, :, 1] += win_w - 1
    relative_coords[:, :, 0] *= 2 * win_w - 1
    return relative_coords.sum(-1)  # Wh*Ww, Wh*Ww

# 量化MLP
class Q_Mlp(nn.Module):
    """ MLP as used in Vision Transformer, MLP-Mixer and related networks
    """
    def __init__(self, nbits, 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
        drop_probs = to_2tuple(drop)

        self.fc1 = LinearQ(in_features, hidden_features, nbits_w=nbits, mode=Qmodes.kernel_wise)
        self.act = act_layer()
        self.drop1 = nn.Dropout(drop_probs[0])
        self.fc2 = LinearQ(hidden_features, out_features, nbits_w=nbits, mode=Qmodes.kernel_wise)
        self.drop2 = nn.Dropout(drop_probs[1])

    def forward(self, x):
        x = self.fc1(x)
        # print(torch.max(x), torch.min(x))
        x = self.act(x)
        
        x = torch.clip(x, -10., 10.)
        # print(torch.clip(x, -10., 10.))
        x = self.drop1(x)
        x = self.fc2(x)
        x = self.drop2(x)
        return x

class Q_WindowAttention(nn.Module):
    """ Window based multi-head self attention (W-MSA) module with relative position bias.
    It supports shifted and non-shifted windows.
    """
    fused_attn: torch.jit.Final[bool]

    def __init__(
            self, 
            nbits:int,
            dim: int,
            num_heads: int,
            head_dim: Optional[int] = None,
            window_size: _int_or_tuple_2_t = 7,
            quantize_attn=True,
            qkv_bias: bool = True,
            attn_drop: float = 0.,
            proj_drop: float = 0.,
    ):
        """
        Args:
            dim: Number of input channels.
            num_heads: Number of attention heads.
            head_dim: Number of channels per head (dim // num_heads if not set)
            window_size: The height and width of the window.
            qkv_bias:  If True, add a learnable bias to query, key, value.
            attn_drop: Dropout ratio of attention weight.
            proj_drop: Dropout ratio of output.
        """
        super().__init__()
        self.dim = dim
        self.window_size = to_2tuple(window_size)  # Wh, Ww
        win_h, win_w = self.window_size
        self.window_area = win_h * win_w
        self.num_heads = num_heads
        head_dim = head_dim or dim // num_heads
        attn_dim = head_dim * num_heads
        self.scale = head_dim ** -0.5
        # self.fused_attn = use_fused_attn(experimental=True)  # NOTE not tested for prime-time yet
        self.fused_attn = False
        self.quantize_attn = quantize_attn
        self.norm_q = nn.LayerNorm(head_dim)
        self.norm_k = nn.LayerNorm(head_dim)

        # define a parameter table of relative position bias, shape: 2*Wh-1 * 2*Ww-1, nH
        # 定义了一个相对位置偏置表
        # 对于一个点来说 其他点和他的相对位置会有(2 * win_h - 1) * (2 * win_w - 1)的情况
        self.relative_position_bias_table = nn.Parameter(torch.zeros((2 * win_h - 1) * (2 * win_w - 1), num_heads))

        # get pair-wise relative position index for each token inside the window
        # 计算每个像素与像素之间的相对距离 get_relative_position_index函数返回的是 ( w*h , w*h )
        self.register_buffer("relative_position_index", get_relative_position_index(win_h, win_w), persistent=False)
        # 将一个张量注册为模型的一个缓冲区 它在模型训练过程中不会被优化器更新，而是被视为固定的常量。
        # 这样可以避免在训练过程中重复计算相同的数据，从而提高训练效率。
        # 看不懂

        if self.quantize_attn:

            self.qkv = LinearQ(dim, dim * 3, bias=qkv_bias, nbits_w=nbits, mode=Qmodes.kernel_wise)

            self.attn_drop = nn.Dropout(attn_drop)

            self.proj = LinearQ(dim, dim, nbits_w=nbits, mode=Qmodes.kernel_wise)
            self.q_act = ActQ(nbits_a=nbits, in_features=self.num_heads)
            self.k_act = ActQ(nbits_a=nbits, in_features=self.num_heads)
            self.v_act = ActQ(nbits_a=nbits, in_features=self.num_heads)
            self.attn_act = ActQ(nbits_a=nbits, in_features=self.num_heads)
        else:
            self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
            self.attn_drop = nn.Dropout(attn_drop)
            self.proj = nn.Linear(dim, dim)
            self.q_act = ActQ(nbits_a=nbits, in_features=self.num_heads)
            self.k_act = ActQ(nbits_a=nbits, in_features=self.num_heads)
            self.v_act = ActQ(nbits_a=nbits, in_features=self.num_heads)
            self.attn_act = ActQ(nbits_a=nbits, in_features=self.num_heads)


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


        trunc_normal_(self.relative_position_bias_table, std=.02)
        self.softmax = nn.Softmax(dim=-1)

    def _get_rel_pos_bias(self) -> torch.Tensor:
        relative_position_bias = self.relative_position_bias_table[                   # table是可学习的    用relative_position_index来进行索引
            self.relative_position_index.view(-1)].view(self.window_area, self.window_area, -1)  # Wh*Ww,Wh*Ww,nH
        # relative_position_bias_table = (2*window_size-1*2*window_size-1, numHeads)
        # 索引得到(window_size*window_size, window_size*window_size, numHeads)
        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
        return relative_position_bias.unsqueeze(0)
        # (1, nH, self.window_area, self.window_area)

    def forward(self, x, mask: Optional[torch.Tensor] = None):
        """
        Args:
            x: input features with shape of (num_windows*B, N, C)
            N是window_size * window_size！！！！  
            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
        """
        B_, N, C = x.shape
        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
        # (num_windows*B, N, 3*C) ->(B,N,3,heads,num_windows*c/heads)  ->  (3,B,heads,N,num_windows*c/heads)
        # 3, numWindows*B, num_heads, window_size*window_size, c//num_heads
        q, k, v = qkv.unbind(0)

        # 暂时先不用这个
        if self.fused_attn:
            attn_mask = self._get_rel_pos_bias()    #拿到位置编码的偏置↑
            if mask is not None:
                num_win = mask.shape[0]
                mask = mask.view(1, num_win, 1, N, N).expand(B_ // num_win, -1, self.num_heads, -1, -1)
                attn_mask = attn_mask + mask.reshape(-1, self.num_heads, N, N)
            x = torch.nn.functional.scaled_dot_product_attention(
                q, k, v,
                attn_mask=attn_mask,
                dropout_p=self.attn_drop.p if self.training else 0.,
            )
        else:
            # q = q * self.scale
            # attn = q @ k.transpose(-2, -1)    #(numWindows*B, num_heads, window_size*window_size, window_size*window_size)
            # attn = attn + self._get_rel_pos_bias()   #这里加上位置编码的偏置了
            # if mask is not None:
            #     num_win = mask.shape[0]
            #     attn = attn.view(-1, num_win, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
            #     attn = attn.view(-1, self.num_heads, N, N)
            # attn = self.softmax(attn)
            # attn = self.attn_drop(attn)
            q = self.norm_q(q)
            k = self.norm_k(k)

            q = self.q_act(q)
            k = self.k_act(k)
            v = self.v_act(v)

            attn1 = (q @ k.transpose(-2, -1)) * self.scale
            attn = attn1.softmax(dim=-1)
            attn = self.attn_drop(attn)
            attn = self.attn_act(attn)
            x = attn @ v
        # x = (numWindows*B, num_heads, window_size*window_size, window_size*window_size)
        x = x.transpose(1, 2).reshape(B_, N, -1)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x,attn1


class Q_SwinTransformerBlock(nn.Module):
    """ Swin Transformer Block.
    """

    def __init__(
            self,
            nbits,
            dim: int,
            input_resolution: _int_or_tuple_2_t,
            num_heads: int = 4,
            head_dim: Optional[int] = None,
            quantize_attn=True,
            window_size: _int_or_tuple_2_t = 7,
            shift_size: int = 0,
            mlp_ratio: float = 4.,
            qkv_bias: bool = True,
            proj_drop: float = 0.,
            attn_drop: float = 0.,
            drop_path: float = 0.,
            act_layer: Callable = nn.GELU,
            norm_layer: Callable = nn.LayerNorm,
    ):
        """
        Args:
            dim: Number of input channels.
            input_resolution: Input resolution.
            window_size: Window size.
            num_heads: Number of attention heads.
            head_dim: Enforce the number of channels per head
            shift_size: Shift size for SW-MSA.
            mlp_ratio: Ratio of mlp hidden dim to embedding dim.
            qkv_bias: If True, add a learnable bias to query, key, value.
            proj_drop: Dropout rate.
            attn_drop: Attention dropout rate.
            drop_path: Stochastic depth rate.
            act_layer: Activation layer.
            norm_layer: Normalization layer.
        """
        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution  #特征图的尺寸
        ws, ss = self._calc_window_shift(window_size, shift_size) # ?
        self.window_size: Tuple[int, int] = ws
        self.shift_size: Tuple[int, int] = ss
        self.window_area = self.window_size[0] * self.window_size[1]
        self.mlp_ratio = mlp_ratio
        self.quantize_attn = quantize_attn

        self.norm1 = norm_layer(dim)
        self.attn = Q_WindowAttention(
            nbits,
            dim,
            num_heads=num_heads,
            head_dim=head_dim,
            window_size=to_2tuple(self.window_size),
            quantize_attn = self.quantize_attn,
            qkv_bias=qkv_bias,
            attn_drop=attn_drop,
            proj_drop=proj_drop,
        )# 多一个相对位置编码的注意力计算
        self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()

        self.norm2 = norm_layer(dim)
        # self.mlp = Mlp(
        #     in_features=dim,
        #     hidden_features=int(dim * mlp_ratio),
        #     act_layer=act_layer,
        #     drop=proj_drop,
        # )
        self.mlp = Q_Mlp(
            nbits=nbits, 
            in_features=dim, 
            hidden_features=int(dim * mlp_ratio), 
            act_layer=act_layer, 
            drop=proj_drop)
        self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()

        if any(self.shift_size):
            # calculate attention mask for SW-MSA
            H, W = self.input_resolution 
            H = math.ceil(H / self.window_size[0]) * self.window_size[0] # math.ceil向上取整
            W = math.ceil(W / self.window_size[1]) * self.window_size[1]
            img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
            cnt = 0
            for h in (
                    slice(0, -self.window_size[0]),
                    slice(-self.window_size[0], -self.shift_size[0]),
                    slice(-self.shift_size[0], None)):
                for w in (
                        slice(0, -self.window_size[1]),
                        slice(-self.window_size[1], -self.shift_size[1]),
                        slice(-self.shift_size[1], None)):
                    img_mask[:, h, w, :] = cnt
                    cnt += 1
            mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
            mask_windows = mask_windows.view(-1, self.window_area)
            attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
            attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
        else:
            attn_mask = None

        self.register_buffer("attn_mask", attn_mask, persistent=False)

    def _calc_window_shift(self, target_window_size, target_shift_size) -> Tuple[Tuple[int, int], Tuple[int, int]]:
        target_window_size = to_2tuple(target_window_size)
        target_shift_size = to_2tuple(target_shift_size)
        window_size = [r if r <= w else w for r, w in zip(self.input_resolution, target_window_size)]
        shift_size = [0 if r <= w else s for r, w, s in zip(self.input_resolution, window_size, target_shift_size)]
        return tuple(window_size), tuple(shift_size)

    def _attn(self, x):
        B, H, W, C = x.shape

        # cyclic shift
        has_shift = any(self.shift_size)
        if has_shift:
            shifted_x = torch.roll(x, shifts=(-self.shift_size[0], -self.shift_size[1]), dims=(1, 2))
        else:
            shifted_x = x

        # pad for resolution not divisible by window size
        pad_h = (self.window_size[0] - H % self.window_size[0]) % self.window_size[0]
        pad_w = (self.window_size[1] - W % self.window_size[1]) % self.window_size[1]
        shifted_x = torch.nn.functional.pad(shifted_x, (0, 0, 0, pad_w, 0, pad_h))
        Hp, Wp = H + pad_h, W + pad_w

        # partition windows
        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
        x_windows = x_windows.view(-1, self.window_area, C)  # nW*B, window_size*window_size, C

        # W-MSA/SW-MSA
        attn_windows = self.attn(x_windows, mask=self.attn_mask)[0]  # nW*B, window_size*window_size, C

        # merge windows
        attn_windows = attn_windows.view(-1, self.window_size[0], self.window_size[1], C)
        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C
        shifted_x = shifted_x[:, :H, :W, :].contiguous()

        # reverse cyclic shift
        if has_shift:
            x = torch.roll(shifted_x, shifts=self.shift_size, dims=(1, 2))
        else:
            x = shifted_x
        return x

    def forward(self, x):
        B, H, W, C = x.shape
        x = x + self.drop_path1(self._attn(self.norm1(x)))
        x = x.reshape(B, -1, C)
        x = x + self.drop_path2(self.mlp(self.norm2(x)))
        x = x.reshape(B, H, W, C)
        return x


class Q_PatchMerging(nn.Module):
    """ Patch Merging Layer.
    """
    # 把 (W,H,C)->(W/2,H/2,4C)最后的维度降维 保持和卷积一样
    def __init__(
            self,
            nbits: int,
            dim: int,
            out_dim: Optional[int] = None,
            norm_layer: Callable = nn.LayerNorm,
    ):
        """
        Args:
            dim: Number of input channels.
            out_dim: Number of output channels (or 2 * dim if None)
            norm_layer: Normalization layer.
        """
        super().__init__()
        self.dim = dim
        self.out_dim = out_dim or 2 * dim
        self.norm = norm_layer(4 * dim)
        # self.reduction = nn.Linear(4 * dim, self.out_dim, bias=False)
        self.reduction = LinearQ(4 * dim, self.out_dim, nbits_w=nbits, mode=Qmodes.kernel_wise)

    def forward(self, x):
        B, H, W, C = x.shape
        _assert(H % 2 == 0, f"x height ({H}) is not even.")
        _assert(W % 2 == 0, f"x width ({W}) is not even.")
        x = x.reshape(B, H // 2, 2, W // 2, 2, C).permute(0, 1, 3, 4, 2, 5).flatten(3)
        x = self.norm(x)
        x = self.reduction(x)
        return x


class Q_SwinTransformerStage(nn.Module):
    """ A basic Swin Transformer layer for one stage.
    """

    def __init__(
            self,
            nbits: int,
            dim: int,
            out_dim: int,
            input_resolution: Tuple[int, int],
            depth: int,
            downsample: bool = True,
            num_heads: int = 4,
            head_dim: Optional[int] = None,
            window_size: _int_or_tuple_2_t = 7,
            mlp_ratio: float = 4.,
            qkv_bias: bool = True,
            proj_drop: float = 0.,
            attn_drop: float = 0.,
            drop_path: Union[List[float], float] = 0.,
            norm_layer: Callable = nn.LayerNorm,
    ):
        """
        Args:
            dim: Number of input channels.
            input_resolution: Input resolution.
            depth: Number of blocks.
            downsample: Downsample layer at the end of the layer.
            num_heads: Number of attention heads.
            head_dim: Channels per head (dim // num_heads if not set)
            window_size: Local window size.
            mlp_ratio: Ratio of mlp hidden dim to embedding dim.
            qkv_bias: If True, add a learnable bias to query, key, value.
            proj_drop: Projection dropout rate.
            attn_drop: Attention dropout rate.
            drop_path: Stochastic depth rate.
            norm_layer: Normalization layer.
        """
        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.output_resolution = tuple(i // 2 for i in input_resolution) if downsample else input_resolution
        self.depth = depth
        self.grad_checkpointing = False
        window_size = to_2tuple(window_size)
        shift_size = tuple([w // 2 for w in window_size])

        # patch merging layer
        if downsample:
            self.downsample = Q_PatchMerging(
                nbits=nbits,
                dim=dim,
                out_dim=out_dim,
                norm_layer=norm_layer,
            )
        else:
            assert dim == out_dim
            self.downsample = nn.Identity()

        # build blocks
        self.blocks = nn.Sequential(*[
            Q_SwinTransformerBlock(
                nbits = nbits,
                dim=out_dim,
                input_resolution=self.output_resolution,
                num_heads=num_heads,
                head_dim=head_dim,
                window_size=window_size,
                shift_size=0 if (i % 2 == 0) else shift_size,
                mlp_ratio=mlp_ratio,
                qkv_bias=qkv_bias,
                proj_drop=proj_drop,
                attn_drop=attn_drop,
                drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
                norm_layer=norm_layer,
            )
            for i in range(depth)])

    def forward(self, x):
        x = self.downsample(x)

        if self.grad_checkpointing and not torch.jit.is_scripting():
            x = checkpoint_seq(self.blocks, x)
        else:
            x = self.blocks(x)
        return x

class Q_PatchEmbed(nn.Module):
    """ 2D Image to Patch Embedding
    """
    output_fmt: Format
    dynamic_img_pad: torch.jit.Final[bool]

    def __init__(
            self,
            nbits: int ,
            img_size: Optional[int] = 224,
            patch_size: int = 16,
            in_chans: int = 3,
            embed_dim: int = 768,
            norm_layer: Optional[Callable] = None,
            flatten: bool = True,
            output_fmt: Optional[str] = None,
            bias: bool = True,
            strict_img_size: bool = True,
            dynamic_img_pad: bool = False,
    ):
        super().__init__()
        self.patch_size = to_2tuple(patch_size)
        if img_size is not None:
            self.img_size = to_2tuple(img_size)
            self.grid_size = tuple([s // p for s, p in zip(self.img_size, self.patch_size)])
            self.num_patches = self.grid_size[0] * self.grid_size[1]
        else:
            self.img_size = None
            self.grid_size = None
            self.num_patches = None

        if output_fmt is not None:
            self.flatten = False
            self.output_fmt = Format(output_fmt)
        else:
            # flatten spatial dim and transpose to channels last, kept for bwd compat
            self.flatten = flatten
            self.output_fmt = Format.NCHW
        self.strict_img_size = strict_img_size
        self.dynamic_img_pad = dynamic_img_pad

        self.proj = Conv2dQ(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias)
        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()

    def forward(self, x):
        B, C, H, W = x.shape
        if self.img_size is not None:
            if self.strict_img_size:
                _assert(H == self.img_size[0], f"Input height ({H}) doesn't match model ({self.img_size[0]}).")
                _assert(W == self.img_size[1], f"Input width ({W}) doesn't match model ({self.img_size[1]}).")
            elif not self.dynamic_img_pad:
                _assert(
                    H % self.patch_size[0] == 0,
                    f"Input height ({H}) should be divisible by patch size ({self.patch_size[0]})."
                )
                _assert(
                    W % self.patch_size[1] == 0,
                    f"Input width ({W}) should be divisible by patch size ({self.patch_size[1]})."
                )
        if self.dynamic_img_pad:
            pad_h = (self.patch_size[0] - H % self.patch_size[0]) % self.patch_size[0]
            pad_w = (self.patch_size[1] - W % self.patch_size[1]) % self.patch_size[1]
            x = F.pad(x, (0, pad_w, 0, pad_h))
        x = self.proj(x)
        if self.flatten:
            x = x.flatten(2).transpose(1, 2)  # NCHW -> NLC
        elif self.output_fmt != Format.NCHW:
            x = nchw_to(x, self.output_fmt)
        x = self.norm(x)
        return x

class SwinTransformer(nn.Module):
    """ Swin Transformer

    A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
          https://arxiv.org/pdf/2103.14030
    """

    def __init__(
            self,
            nbits = 4,
            img_size: _int_or_tuple_2_t = 224,
            patch_size: int = 4,
            in_chans: int = 3,
            num_classes: int = 1000,
            global_pool: str = 'avg',
            embed_dim: int = 96,
            depths: Tuple[int, ...] = (2, 2, 6, 2),
            num_heads: Tuple[int, ...] = (3, 6, 12, 24),
            head_dim: Optional[int] = None,
            window_size: _int_or_tuple_2_t = 7,
            mlp_ratio: float = 4.,
            qkv_bias: bool = True,
            drop_rate: float = 0.,
            proj_drop_rate: float = 0.,
            attn_drop_rate: float = 0.,
            drop_path_rate: float = 0.1,
            embed_layer: Callable = Q_PatchEmbed,
            norm_layer: Union[str, Callable] = nn.LayerNorm,
            weight_init: str = '',
            **kwargs,
    ):
        """
        Args:
            img_size: Input image size.
            patch_size: Patch size.
            in_chans: Number of input image channels.
            num_classes: Number of classes for classification head.
            embed_dim: Patch embedding dimension.
            depths: Depth of each Swin Transformer layer.
            num_heads: Number of attention heads in different layers.
            head_dim: Dimension of self-attention heads.
            window_size: Window size.
            mlp_ratio: Ratio of mlp hidden dim to embedding dim.
            qkv_bias: If True, add a learnable bias to query, key, value.
            drop_rate: Dropout rate.
            attn_drop_rate (float): Attention dropout rate.
            drop_path_rate (float): Stochastic depth rate.
            embed_layer: Patch embedding layer.
            norm_layer (nn.Module): Normalization layer.
        """
        super().__init__()
        assert global_pool in ('', 'avg')
        self.num_classes = num_classes
        self.global_pool = global_pool
        self.output_fmt = 'NHWC'

        self.num_layers = len(depths)
        self.embed_dim = embed_dim
        self.num_features = int(embed_dim * 2 ** (self.num_layers - 1))
        self.feature_info = []

        if not isinstance(embed_dim, (tuple, list)):
            embed_dim = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]

        # split image into non-overlapping patches
        self.patch_embed = embed_layer(
            nbits=nbits,
            img_size=img_size,
            patch_size=patch_size,
            in_chans=in_chans,
            embed_dim=embed_dim[0],
            norm_layer=norm_layer,
            output_fmt='NHWC',
        )
        self.patch_grid = self.patch_embed.grid_size

        # build layers
        head_dim = to_ntuple(self.num_layers)(head_dim)
        if not isinstance(window_size, (list, tuple)):
            window_size = to_ntuple(self.num_layers)(window_size)
        elif len(window_size) == 2:
            window_size = (window_size,) * self.num_layers
        assert len(window_size) == self.num_layers
        mlp_ratio = to_ntuple(self.num_layers)(mlp_ratio)
        dpr = [x.tolist() for x in torch.linspace(0, drop_path_rate, sum(depths)).split(depths)]
        layers = []
        in_dim = embed_dim[0]
        scale = 1
        for i in range(self.num_layers):
            out_dim = embed_dim[i]
            layers += [Q_SwinTransformerStage(
                nbits=nbits,
                dim=in_dim,
                out_dim=out_dim,
                input_resolution=(
                    self.patch_grid[0] // scale,
                    self.patch_grid[1] // scale
                ),
                depth=depths[i],
                downsample=i > 0,
                num_heads=num_heads[i],
                head_dim=head_dim[i],
                window_size=window_size[i],
                mlp_ratio=mlp_ratio[i],
                qkv_bias=qkv_bias,
                proj_drop=proj_drop_rate,
                attn_drop=attn_drop_rate,
                drop_path=dpr[i],
                norm_layer=norm_layer,
            )]
            in_dim = out_dim
            if i > 0:
                scale *= 2
            self.feature_info += [dict(num_chs=out_dim, reduction=4 * scale, module=f'layers.{i}')]
        self.layers = nn.Sequential(*layers)

        self.norm = norm_layer(self.num_features)
        self.head = ClassifierHead(
            nbits,
            self.num_features,
            num_classes,
            pool_type=global_pool,
            drop_rate=drop_rate,
            input_fmt=self.output_fmt,
        )
        if weight_init != 'skip':
            self.init_weights(weight_init)

    @torch.jit.ignore
    def init_weights(self, mode=''):
        assert mode in ('jax', 'jax_nlhb', 'moco', '')
        head_bias = -math.log(self.num_classes) if 'nlhb' in mode else 0.
        named_apply(get_init_weights_vit(mode, head_bias=head_bias), self)

    @torch.jit.ignore
    def no_weight_decay(self):
        nwd = set()
        for n, _ in self.named_parameters():
            if 'relative_position_bias_table' in n:
                nwd.add(n)
        return nwd

    @torch.jit.ignore
    def group_matcher(self, coarse=False):
        return dict(
            stem=r'^patch_embed',  # stem and embed
            blocks=r'^layers\.(\d+)' if coarse else [
                (r'^layers\.(\d+).downsample', (0,)),
                (r'^layers\.(\d+)\.\w+\.(\d+)', None),
                (r'^norm', (99999,)),
            ]
        )

    @torch.jit.ignore
    def set_grad_checkpointing(self, enable=True):
        for l in self.layers:
            l.grad_checkpointing = enable

    @torch.jit.ignore
    def get_classifier(self):
        return self.head.fc

    def reset_classifier(self, num_classes, global_pool=None):
        self.num_classes = num_classes
        self.head.reset(num_classes, pool_type=global_pool)

    def forward_features(self, x):
        x = self.patch_embed(x)
        x = self.layers(x)
        x = self.norm(x)
        return x

    def forward_head(self, x, pre_logits: bool = False):
        return self.head(x, pre_logits=True) if pre_logits else self.head(x)

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


def checkpoint_filter_fn(state_dict, model):
    """ convert patch embedding weight from manual patchify + linear proj to conv"""
    old_weights = True
    if 'head.fc.weight' in state_dict:
        old_weights = False
    import re
    out_dict = {}
    state_dict = state_dict.get('model', state_dict)
    state_dict = state_dict.get('state_dict', state_dict)
    for k, v in state_dict.items():
        if any([n in k for n in ('relative_position_index', 'attn_mask')]):
            continue  # skip buffers that should not be persistent

        if 'patch_embed.proj.weight' in k:
            _, _, H, W = model.patch_embed.proj.weight.shape
            if v.shape[-2] != H or v.shape[-1] != W:
                v = resample_patch_embed(
                    v,
                    (H, W),
                    interpolation='bicubic',
                    antialias=True,
                    verbose=True,
                )

        if k.endswith('relative_position_bias_table'):
            m = model.get_submodule(k[:-29])
            if v.shape != m.relative_position_bias_table.shape or m.window_size[0] != m.window_size[1]:
                v = resize_rel_pos_bias_table(
                    v,
                    new_window_size=m.window_size,
                    new_bias_shape=m.relative_position_bias_table.shape,
                )

        if old_weights:
            k = re.sub(r'layers.(\d+).downsample', lambda x: f'layers.{int(x.group(1)) + 1}.downsample', k)
            k = k.replace('head.', 'head.fc.')

        out_dict[k] = v
    return out_dict


def _create_swin_transformer(variant, pretrained=False, **kwargs):
    default_out_indices = tuple(i for i, _ in enumerate(kwargs.get('depths', (1, 1, 3, 1))))
    out_indices = kwargs.pop('out_indices', default_out_indices)

    model = build_model_with_cfg(
        SwinTransformer, variant, pretrained,
        pretrained_filter_fn=checkpoint_filter_fn,
        feature_cfg=dict(flatten_sequential=True, out_indices=out_indices),
        **kwargs)

    return model


def _cfg(url='', **kwargs):
    return {
        'url': url,
        'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': (7, 7),
        'crop_pct': .9, 'interpolation': 'bicubic', 'fixed_input_size': True,
        'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
        'first_conv': 'patch_embed.proj', 'classifier': 'head.fc',
        'license': 'mit', **kwargs
    }


# default_cfgs = generate_default_cfgs({
#     'swin_small_patch4_window7_224.ms_in22k_ft_in1k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.8/swin_small_patch4_window7_224_22kto1k_finetune.pth', ),
#     'swin_base_patch4_window7_224.ms_in22k_ft_in1k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window7_224_22kto1k.pth',),
#     'swin_base_patch4_window12_384.ms_in22k_ft_in1k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window12_384_22kto1k.pth',
#         input_size=(3, 384, 384), pool_size=(12, 12), crop_pct=1.0),
#     'swin_large_patch4_window7_224.ms_in22k_ft_in1k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_large_patch4_window7_224_22kto1k.pth',),
#     'swin_large_patch4_window12_384.ms_in22k_ft_in1k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_large_patch4_window12_384_22kto1k.pth',
#         input_size=(3, 384, 384), pool_size=(12, 12), crop_pct=1.0),

#     'swin_tiny_patch4_window7_224.ms_in1k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_tiny_patch4_window7_224.pth',),
#     'swin_small_patch4_window7_224.ms_in1k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_small_patch4_window7_224.pth',),
#     'swin_base_patch4_window7_224.ms_in1k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window7_224.pth',),
#     'swin_base_patch4_window12_384.ms_in1k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window12_384.pth',
#         input_size=(3, 384, 384), pool_size=(12, 12), crop_pct=1.0),

#     # tiny 22k pretrain is worse than 1k, so moved after (untagged priority is based on order)
#     'swin_tiny_patch4_window7_224.ms_in22k_ft_in1k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.8/swin_tiny_patch4_window7_224_22kto1k_finetune.pth',),

#     'swin_tiny_patch4_window7_224.ms_in22k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.8/swin_tiny_patch4_window7_224_22k.pth',
#         num_classes=21841),
#     'swin_small_patch4_window7_224.ms_in22k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.8/swin_small_patch4_window7_224_22k.pth',
#         num_classes=21841),
#     'swin_base_patch4_window7_224.ms_in22k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window7_224_22k.pth',
#         num_classes=21841),
#     'swin_base_patch4_window12_384.ms_in22k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window12_384_22k.pth',
#         input_size=(3, 384, 384), pool_size=(12, 12), crop_pct=1.0, num_classes=21841),
#     'swin_large_patch4_window7_224.ms_in22k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_large_patch4_window7_224_22k.pth',
#         num_classes=21841),
#     'swin_large_patch4_window12_384.ms_in22k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_large_patch4_window12_384_22k.pth',
#         input_size=(3, 384, 384), pool_size=(12, 12), crop_pct=1.0, num_classes=21841),

#     'swin_s3_tiny_224.ms_in1k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/s3_t-1d53f6a8.pth'),
#     'swin_s3_small_224.ms_in1k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/s3_s-3bb4c69d.pth'),
#     'swin_s3_base_224.ms_in1k': _cfg(
#         hf_hub_id='timm/',
#         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/s3_b-a1e95db4.pth'),
# })


# @register_model
# def swin_tiny_patch4_window7_224(pretrained=False, **kwargs) -> SwinTransformer:
#     """ Swin-T @ 224x224, trained ImageNet-1k
#     """
#     model_args = dict(patch_size=4, window_size=7, embed_dim=96, depths=(2, 2, 6, 2), num_heads=(3, 6, 12, 24))
#     return _create_swin_transformer(
#         'swin_tiny_patch4_window7_224', pretrained=pretrained, **dict(model_args, **kwargs))


# @register_model
# def swin_small_patch4_window7_224(pretrained=False, **kwargs) -> SwinTransformer:
#     """ Swin-S @ 224x224
#     """
#     model_args = dict(patch_size=4, window_size=7, embed_dim=96, depths=(2, 2, 18, 2), num_heads=(3, 6, 12, 24))
#     return _create_swin_transformer(
#         'swin_small_patch4_window7_224', pretrained=pretrained, **dict(model_args, **kwargs))


# @register_model
# def swin_base_patch4_window7_224(pretrained=False, **kwargs) -> SwinTransformer:
#     """ Swin-B @ 224x224
#     """
#     model_args = dict(patch_size=4, window_size=7, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32))
#     return _create_swin_transformer(
#         'swin_base_patch4_window7_224', pretrained=pretrained, **dict(model_args, **kwargs))


# @register_model
# def swin_base_patch4_window12_384(pretrained=False, **kwargs) -> SwinTransformer:
#     """ Swin-B @ 384x384
#     """
#     model_args = dict(patch_size=4, window_size=12, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32))
#     return _create_swin_transformer(
#         'swin_base_patch4_window12_384', pretrained=pretrained, **dict(model_args, **kwargs))


# @register_model
# def swin_large_patch4_window7_224(pretrained=False, **kwargs) -> SwinTransformer:
#     """ Swin-L @ 224x224
#     """
#     model_args = dict(patch_size=4, window_size=7, embed_dim=192, depths=(2, 2, 18, 2), num_heads=(6, 12, 24, 48))
#     return _create_swin_transformer(
#         'swin_large_patch4_window7_224', pretrained=pretrained, **dict(model_args, **kwargs))


# @register_model
# def swin_large_patch4_window12_384(pretrained=False, **kwargs) -> SwinTransformer:
#     """ Swin-L @ 384x384
#     """
#     model_args = dict(patch_size=4, window_size=12, embed_dim=192, depths=(2, 2, 18, 2), num_heads=(6, 12, 24, 48))
#     return _create_swin_transformer(
#         'swin_large_patch4_window12_384', pretrained=pretrained, **dict(model_args, **kwargs))


# @register_model
# def swin_s3_tiny_224(pretrained=False, **kwargs) -> SwinTransformer:
#     """ Swin-S3-T @ 224x224, https://arxiv.org/abs/2111.14725
#     """
#     model_args = dict(
#         patch_size=4, window_size=(7, 7, 14, 7), embed_dim=96, depths=(2, 2, 6, 2), num_heads=(3, 6, 12, 24))
#     return _create_swin_transformer('swin_s3_tiny_224', pretrained=pretrained, **dict(model_args, **kwargs))


# @register_model
# def swin_s3_small_224(pretrained=False, **kwargs) -> SwinTransformer:
#     """ Swin-S3-S @ 224x224, https://arxiv.org/abs/2111.14725
#     """
#     model_args = dict(
#         patch_size=4, window_size=(14, 14, 14, 7), embed_dim=96, depths=(2, 2, 18, 2), num_heads=(3, 6, 12, 24))
#     return _create_swin_transformer('swin_s3_small_224', pretrained=pretrained, **dict(model_args, **kwargs))


# @register_model
# def swin_s3_base_224(pretrained=False, **kwargs) -> SwinTransformer:
#     """ Swin-S3-B @ 224x224, https://arxiv.org/abs/2111.14725
#     """
#     model_args = dict(
#         patch_size=4, window_size=(7, 7, 14, 7), embed_dim=96, depths=(2, 2, 30, 2), num_heads=(3, 6, 12, 24))
#     return _create_swin_transformer('swin_s3_base_224', pretrained=pretrained, **dict(model_args, **kwargs))


# register_model_deprecations(__name__, {
#     'swin_base_patch4_window7_224_in22k': 'swin_base_patch4_window7_224.ms_in22k',
#     'swin_base_patch4_window12_384_in22k': 'swin_base_patch4_window12_384.ms_in22k',
#     'swin_large_patch4_window7_224_in22k': 'swin_large_patch4_window7_224.ms_in22k',
#     'swin_large_patch4_window12_384_in22k': 'swin_large_patch4_window12_384.ms_in22k',
# })