modular_vitae_window_noshift.py

# Copyright (c) [2012]-[2021] Shanghai Yitu Technology Co., Ltd.
#
# This source code is licensed under the Clear BSD License
# LICENSE file in the root directory of this file
# All rights reserved.
"""
ViTAE Window NoShift Model - Consolidated modular implementation
"""
from functools import partial
import math
import torch
import torch.nn as nn
import numpy as np

# Use transformers equivalents instead of timm
from transformers.models.swin.modeling_swin import SwinDropPath as DropPath
from transformers import initialization as init
trunc_normal_ = init.trunc_normal_
from timm.models.helpers import load_pretrained

# Simple to_2tuple replacement
def to_2tuple(x):
    """Convert input to 2-tuple if not already a tuple."""
    return x if isinstance(x, tuple) else (x, x)


# ============================================================================
# SELayer
# ============================================================================
class SELayer(nn.Module):
    def __init__(self, channel, reduction=16):
        super(SELayer, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool1d(1)
        self.fc = nn.Sequential(
            nn.Linear(channel, channel // reduction, bias=False),
            nn.ReLU(inplace=True),
            nn.Linear(channel // reduction, channel, bias=False),
            nn.Sigmoid()
        )

    def forward(self, x):  # x: [B, N, C]
        x = torch.transpose(x, 1, 2)  # [B, C, N]
        b, c, _ = x.size()
        y = self.avg_pool(x).view(b, c)
        y = self.fc(y).view(b, c, 1)
        x = x * y.expand_as(x)
        x = torch.transpose(x, 1, 2)  # [B, N, C]
        return x


# ============================================================================
# Swin Components
# ============================================================================
def window_partition(x, window_size):
    """
    Args:
        x: (B, H, W, C)
        window_size (int): window size
    Returns:
        windows: (num_windows*B, window_size, window_size, C)
    """
    B, H, W, C = x.shape
    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
    return windows


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


class WindowAttention(nn.Module):
    r""" Window based multi-head self attention (W-MSA) module with relative position bias.
    It supports both of shifted and non-shifted window.
    """
    def __init__(self, in_dim, out_dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0., relative_pos=False):
        super().__init__()
        self.in_dim = in_dim
        self.dim = out_dim
        self.window_size = window_size  # Wh, Ww
        self.num_heads = num_heads
        head_dim = out_dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5
        self.relative_pos = relative_pos

        # define a parameter table of relative position bias
        self.relative_position_bias_table = nn.Parameter(
            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
        if self.relative_pos:
            print('enable relative pos embedding')

        # get pair-wise relative position index for each token inside the window
        coords_h = torch.arange(self.window_size[0])
        coords_w = torch.arange(self.window_size[1])
        coords = torch.stack(torch.meshgrid([coords_h, coords_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] += self.window_size[0] - 1  # shift to start from 0
        relative_coords[:, :, 1] += self.window_size[1] - 1
        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
        self.register_buffer("relative_position_index", relative_position_index)

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

        init.trunc_normal_(self.relative_position_bias_table, std=0.02)
        self.softmax = nn.Softmax(dim=-1)

    def forward(self, x, mask=None):
        """
        Args:
            x: input features with shape of (num_windows*B, N, C)
            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)
        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)

        q = q * self.scale
        attn = (q @ k.transpose(-2, -1))

        if self.relative_pos:
            relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
                self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
            relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
            attn = attn + relative_position_bias.unsqueeze(0)

        if mask is not None:
            nW = mask.shape[0]
            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
            attn = attn.view(-1, self.num_heads, N, N)
            attn = self.softmax(attn)
        else:
            attn = self.softmax(attn)

        attn = self.attn_drop(attn)

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


class SwinTransformerBlock(nn.Module):
    r""" Swin Transformer Block."""
    def __init__(self, in_dim, out_dim, input_resolution, num_heads, window_size=7, shift_size=0,
                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
                 relative_pos=True, act_layer=nn.GELU, norm_layer=nn.LayerNorm):
        super().__init__()
        self.in_dim = in_dim
        self.dim = out_dim
        self.input_resolution = input_resolution
        self.num_heads = num_heads
        self.window_size = window_size
        self.shift_size = shift_size
        self.mlp_ratio = mlp_ratio
        self.relative_pos = relative_pos
        if min(self.input_resolution) <= self.window_size:
            # if window size is larger than input resolution, we don't partition windows
            self.shift_size = 0
            self.window_size = min(self.input_resolution)
        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"

        self.norm1 = norm_layer(in_dim)
        self.attn = WindowAttention(
            in_dim=in_dim, out_dim=out_dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, relative_pos=relative_pos)

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

        if self.shift_size > 0:
            # calculate attention mask for SW-MSA
            H, W = self.input_resolution
            img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
            h_slices = (slice(0, -self.window_size),
                        slice(-self.window_size, -self.shift_size),
                        slice(-self.shift_size, None))
            w_slices = (slice(0, -self.window_size),
                        slice(-self.window_size, -self.shift_size),
                        slice(-self.shift_size, None))
            cnt = 0
            for h in h_slices:
                for w in w_slices:
                    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_size * self.window_size)
            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)

    def forward(self, x):
        H, W = self.input_resolution
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size"

        shortcut = x
        x = self.norm1(x)
        x = x.view(B, H, W, C)

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

        # 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_size * self.window_size, C)  # nW*B, window_size*window_size, C

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

        # merge windows
        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C

        # reverse cyclic shift
        if self.shift_size > 0:
            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
        else:
            x = shifted_x
        x = x.view(B, H * W, C)

        # FFN
        x = shortcut + self.drop_path(x)
        x = x + self.drop_path(self.mlp(self.norm2(x)))

        return x


# ============================================================================
# MLP and Attention Components
# ============================================================================
class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.hidden_features = hidden_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

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


class Attention(nn.Module):
    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).permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]

        attn = (q @ k.transpose(-2, -1)) * self.scale
        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 AttentionPerformer(nn.Module):
    def __init__(self, dim, num_heads=1, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., kernel_ratio=0.5):
        super().__init__()
        self.head_dim = dim // num_heads
        self.emb = dim
        self.kqv = nn.Linear(dim, 3 * self.emb)
        self.dp = nn.Dropout(proj_drop)
        self.proj = nn.Linear(self.emb, self.emb)
        self.head_cnt = num_heads
        self.norm1 = nn.LayerNorm(dim, eps=1e-6)
        self.epsilon = 1e-8  # for stable in division
        self.drop_path = nn.Identity()

        self.m = int(self.head_dim * kernel_ratio)
        self.w = torch.randn(self.head_cnt, self.m, self.head_dim)
        for i in range(self.head_cnt):
            self.w[i] = nn.Parameter(nn.init.orthogonal_(self.w[i]) * math.sqrt(self.m), requires_grad=False)
        self.w.requires_grad_(False)

    def prm_exp(self, x):
        # part of the function is borrow from https://github.com/lucidrains/performer-pytorch 
        # and Simo Ryu (https://github.com/cloneofsimo)
        # ==== positive random features for gaussian kernels ====
        # x = (B, T, hs)
        # w = (m, hs)
        # return : x : B, T, m
        # SM(x, y) = E_w[exp(w^T x - |x|/2) exp(w^T y - |y|/2)]
        # therefore return exp(w^Tx - |x|/2)/sqrt(m)
        xd = ((x * x).sum(dim=-1, keepdim=True)).repeat(1, 1, 1, self.m) / 2
        wtx = torch.einsum('bhti,hmi->bhtm', x.float(), self.w.to(x.device))

        return torch.exp(wtx - xd) / math.sqrt(self.m)

    def attn(self, x):
        B, N, C = x.shape
        kqv = self.kqv(x).reshape(B, N, 3, self.head_cnt, self.head_dim).permute(2, 0, 3, 1, 4)
        k, q, v = kqv[0], kqv[1], kqv[2] # (B, H, T, hs)
        
        kp, qp = self.prm_exp(k), self.prm_exp(q)  # (B, H, T, m), (B, H, T, m)
        D = torch.einsum('bhti,bhi->bht', qp, kp.sum(dim=2)).unsqueeze(dim=-1)  # (B, H, T, m) * (B, H, m) -> (B, H, T, 1)
        kptv = torch.einsum('bhin,bhim->bhnm', v.float(), kp)  # (B, H, emb, m)
        y = torch.einsum('bhti,bhni->bhtn', qp, kptv) / (D.repeat(1, 1, 1, self.head_dim) + self.epsilon)  # (B, H, T, emb)/Diag

        # skip connection
        y = y.permute(0, 2, 1, 3).reshape(B, N, self.emb)
        y = self.dp(self.proj(y))  # same as token_transformer in T2T layer, use v as skip connection

        return y

    def forward(self, x):
        x = self.attn(x)
        return x


# ============================================================================
# Token Transformer and Performer
# ============================================================================
class TokenTransformerAttention(nn.Module):
    def __init__(self, dim, num_heads=8, in_dim = None, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., gamma=False, init_values=1e-4):
        super().__init__()
        self.num_heads = num_heads
        self.in_dim = in_dim
        head_dim = in_dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5

        self.qkv = nn.Linear(dim, in_dim * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(in_dim, in_dim)
        self.proj_drop = nn.Dropout(proj_drop)
        if gamma:
            self.gamma1 = nn.Parameter(init_values * torch.ones((in_dim)),requires_grad=True)
        else:
            self.gamma1 = 1

    def forward(self, x):
        B, N, C = x.shape

        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.in_dim // self.num_heads).permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]

        # attn：B,H,N,N
        attn = (q @ k.transpose(-2, -1)) * self.scale
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B, N, self.in_dim)
        x = self.proj(x)
        x = self.proj_drop(self.gamma1 * x)
        v = v.permute(0, 2, 1, 3).view(B, N, self.in_dim).contiguous()
        # skip connection
        x = v + x   # because the original x has different size with current x, use v to do skip connection

        return x


class Token_transformer(nn.Module):
    def __init__(self, dim, in_dim, num_heads, mlp_ratio=1., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, gamma=False, init_values=1e-4):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = TokenTransformerAttention(
            dim, in_dim=in_dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, gamma=gamma, init_values=init_values)
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(in_dim)
        self.mlp = Mlp(in_features=in_dim, hidden_features=int(in_dim*mlp_ratio), out_features=in_dim, act_layer=act_layer, drop=drop)

    def forward(self, x):
        x = self.attn(self.norm1(x))
        x = x + self.drop_path(self.mlp(self.norm2(x)))
        return x


class Token_performer(nn.Module):
    """
    Take Performer as T2T Transformer
    """
    def __init__(self, dim, in_dim, head_cnt=1, kernel_ratio=0.5, dp1=0.1, dp2 = 0.1, gamma=False, init_values=1e-4):
        super().__init__()
        self.head_dim = in_dim // head_cnt
        self.emb = in_dim
        self.kqv = nn.Linear(dim, 3 * self.emb)
        self.dp = nn.Dropout(dp1)
        self.proj = nn.Linear(self.emb, self.emb)
        self.head_cnt = head_cnt
        self.norm1 = nn.LayerNorm(dim, eps=1e-6)
        self.norm2 = nn.LayerNorm(self.emb, eps=1e-6)
        self.epsilon = 1e-8  # for stable in division
        self.drop_path = nn.Identity()

        self.mlp = nn.Sequential(
            nn.Linear(self.emb, 1 * self.emb),
            nn.GELU(),
            nn.Linear(1 * self.emb, self.emb),
            nn.Dropout(dp2),
        )

        self.m = int(self.head_dim * kernel_ratio)
        self.w = torch.randn(head_cnt, self.m, self.head_dim)
        for i in range(self.head_cnt):
            self.w[i] = nn.Parameter(nn.init.orthogonal_(self.w[i]) * math.sqrt(self.m), requires_grad=False)
        self.w.requires_grad_(False)

        if gamma:
            self.gamma1 = nn.Parameter(init_values * torch.ones((self.emb)),requires_grad=True)
        else:
            self.gamma1 = 1

    def prm_exp(self, x):
        # part of the function is borrow from https://github.com/lucidrains/performer-pytorch 
        # and Simo Ryu (https://github.com/cloneofsimo)
        # ==== positive random features for gaussian kernels ====
        # x = (B, H, N, hs)
        # w = (H, m, hs)
        # return : x : B, T, m
        # SM(x, y) = E_w[exp(w^T x - |x|/2) exp(w^T y - |y|/2)]
        # therefore return exp(w^Tx - |x|/2)/sqrt(m)

        xd = ((x * x).sum(dim=-1, keepdim=True)).repeat(1, 1, 1, self.m) / 2

        # BHNhs * Hmhs -> BHNm
        wtx = torch.einsum('bhti,hmi->bhtm', x.float(), self.w.to(x.device))

        # BHNm

        return torch.exp(wtx - xd) / math.sqrt(self.m)

    def attn(self, x):
        B, N, C = x.shape
        # B,N,C -> B,N,3C -> B,N,3,H,C/H -> 3,B,H,N,C'
        kqv = self.kqv(x).reshape(B, N, 3, self.head_cnt, self.head_dim).permute(2, 0, 3, 1, 4)
        k, q, v = kqv[0], kqv[1], kqv[2] # (B, H, T, hs)

        kp, qp = self.prm_exp(k), self.prm_exp(q)  # (B, H, T, m), (B, H, T, m)
        D = torch.einsum('bhti,bhi->bht', qp, kp.sum(dim=2)).unsqueeze(dim=-1)  # (B, H, T, m) * (B, H, m) -> (B, H, T, 1)
        kptv = torch.einsum('bhin,bhim->bhnm', v.float(), kp)  # (B, H, emb, m)
        y = torch.einsum('bhti,bhni->bhtn', qp, kptv) / (D.repeat(1, 1, 1, self.head_dim) + self.epsilon)  # (B, H, T, emb)/Diag

        # skip connection

        y = y.permute(0, 2, 1, 3).reshape(B, N, self.emb)
        v = v.permute(0, 2, 1, 3).reshape(B, N, self.emb)

        y = v + self.dp(self.gamma1 * self.proj(y))  # same as token_transformer, use v as skip connection

        return y

    def forward(self, x):
        x = self.attn(self.norm1(x))
        x = x + self.mlp(self.norm2(x))
        return x


# ============================================================================
# NormalCell and ReductionCell
# ============================================================================
class NormalCell(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, class_token=False, group=64, tokens_type='transformer', 
                shift_size=0, window_size=0, gamma=False, init_values=1e-4, SE=False, img_size=224, relative_pos=False):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.class_token = class_token
        self.img_size = img_size
        self.window_size = window_size
        if shift_size > 0 and self.img_size > self.window_size:
            self.shift_size = shift_size
        else:
            self.shift_size = 0
        self.tokens_type = tokens_type
        if tokens_type == 'transformer':
            self.attn = Attention(
                dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
        elif tokens_type == 'performer':
            self.attn = AttentionPerformer(
                dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
        elif tokens_type == 'swin':
            if self.shift_size > 0:
                # calculate attention mask for SW-MSA
                H, W = self.img_size, self.img_size
                img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
                h_slices = (slice(0, -self.window_size),
                            slice(-self.window_size, -self.shift_size),
                            slice(-self.shift_size, None))
                w_slices = (slice(0, -self.window_size),
                            slice(-self.window_size, -self.shift_size),
                            slice(-self.shift_size, None))
                cnt = 0
                for h in h_slices:
                    for w in w_slices:
                        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_size * self.window_size)
                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)
            self.attn = WindowAttention(
                in_dim=dim, out_dim=dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
                qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, relative_pos=relative_pos)

        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.PCM = nn.Sequential(
                            nn.Conv2d(dim, mlp_hidden_dim, 3, 1, 1, 1, group),
                            nn.BatchNorm2d(mlp_hidden_dim),
                            nn.SiLU(inplace=True),
                            nn.Conv2d(mlp_hidden_dim, dim, 3, 1, 1, 1, group),
                            nn.BatchNorm2d(dim),
                            nn.SiLU(inplace=True),
                            nn.Conv2d(dim, dim, 3, 1, 1, 1, group),
                            )
        if gamma:
            self.gamma1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
            self.gamma2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
            self.gamma3 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
        else:
            self.gamma1 = 1
            self.gamma2 = 1
            self.gamma3 = 1
        if SE:
            self.SE = SELayer(dim)
        else:
            self.SE = nn.Identity()

    def forward(self, x):

        b, n, c = x.shape
        shortcut = x
        if self.tokens_type == 'swin':
            H, W = self.img_size, self.img_size
            assert n == self.img_size * self.img_size, "input feature has wrong size"
            x = self.norm1(x)
            x = x.view(b, H, W, c)
            # cyclic shift
            if self.shift_size > 0:
                shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
            else:
                shifted_x = x
            # 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_size * self.window_size, c)  # nW*B, window_size*window_size, C
            # W-MSA/SW-MSA
            attn_windows = self.attn(x_windows, mask=self.attn_mask)  # nW*B, window_size*window_size, C

            # merge windows
            attn_windows = attn_windows.view(-1, self.window_size, self.window_size, c)
            shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C

            # reverse cyclic shift
            if self.shift_size > 0:
                x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
            else:
                x = shifted_x
            x = x.view(b, H * W, c)
        else:
            x = self.gamma1 * self.attn(self.norm1(x))

        if self.class_token:
            n = n - 1
            wh = int(math.sqrt(n))
            convX = self.drop_path(self.gamma2 * self.PCM(shortcut[:, 1:, :].view(b, wh, wh, c).permute(0, 3, 1, 2).contiguous()).permute(0, 2, 3, 1).contiguous().view(b, n, c))
            x = shortcut + self.drop_path(self.gamma1 * x)
            x[:, 1:] = x[:, 1:] + convX
        else:
            wh = int(math.sqrt(n))
            convX = self.drop_path(self.gamma2 * self.PCM(shortcut.view(b, wh, wh, c).permute(0, 3, 1, 2).contiguous()).permute(0, 2, 3, 1).contiguous().view(b, n, c))
            x = shortcut + self.drop_path(self.gamma1 * x) + convX
            # x = x + convX
        x = x + self.drop_path(self.gamma3 * self.mlp(self.norm2(x)))
        x = self.SE(x)
        return x


class PatchEmbedding(nn.Module):
    def __init__(self, inter_channel=32, out_channels=48, img_size=None):
        self.img_size = img_size
        self.inter_channel = inter_channel
        self.out_channel = out_channels
        super().__init__()
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, inter_channel, kernel_size=3, stride=2, padding=1, bias=False),
            nn.BatchNorm2d(inter_channel),
            nn.ReLU(inplace=True)
        )
        self.conv2 = nn.Sequential(
            nn.Conv2d(inter_channel, out_channels, kernel_size=3, stride=2, padding=1, bias=False),
            nn.BatchNorm2d(out_channels),
            nn.ReLU(inplace=True)
        )
        self.conv3 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)

    def forward(self, x):
        x = self.conv3(self.conv2(self.conv1(x)))
        b, c, h, w = x.shape
        x = x.permute(0, 2, 3, 1).reshape(b, h*w, c)
        return x

    def flops(self, ) -> float:
        flops = 0
        flops += 3 * self.inter_channel * self.img_size[0] * self.img_size[1] // 4 * 9
        flops += self.img_size[0] * self.img_size[1] // 4 * self.inter_channel
        flops += self.inter_channel * self.out_channel * self.img_size[0] * self.img_size[1] // 16 * 9
        flops += self.img_size[0] * self.img_size[1] // 16 * self.out_channel
        flops += self.out_channel * self.out_channel * self.img_size[0] * self.img_size[1] // 16
        return flops


class PRM(nn.Module):
    def __init__(self, img_size=224, kernel_size=4, downsample_ratio=4, dilations=[1,6,12], in_chans=3, embed_dim=64, share_weights=False, op='cat'):
        super().__init__()
        self.dilations = dilations
        self.embed_dim = embed_dim
        self.downsample_ratio = downsample_ratio
        self.op = op
        self.kernel_size = kernel_size
        self.stride = downsample_ratio
        self.share_weights = share_weights
        self.outSize = img_size // downsample_ratio

        if share_weights:
            self.convolution = nn.Conv2d(in_channels=in_chans, out_channels=embed_dim, kernel_size=self.kernel_size, \
                stride=self.stride, padding=3*dilations[0]//2, dilation=dilations[0])

        else:
            self.convs = nn.ModuleList()
            for dilation in self.dilations:
                padding = math.ceil(((self.kernel_size-1)*dilation + 1 - self.stride) / 2)
                self.convs.append(nn.Sequential(*[nn.Conv2d(in_channels=in_chans, out_channels=embed_dim, kernel_size=self.kernel_size, \
                    stride=self.stride, padding=padding, dilation=dilation),
                    nn.GELU()]))

        if self.op == 'sum':
            self.out_chans = embed_dim
        elif op == 'cat':
            self.out_chans = embed_dim * len(self.dilations)

    def forward(self, x):
        B, C, W, H = x.shape
        if self.share_weights:
            padding = math.ceil(((self.kernel_size-1)*self.dilations[0] + 1 - self.stride) / 2)
            y = nn.functional.conv2d(x, weight=self.convolution.weight, bias=self.convolution.bias, \
                stride=self.downsample_ratio, padding=padding, dilation=self.dilations[0]).unsqueeze(dim=-1)
            for i in range(1, len(self.dilations)):
                padding = math.ceil(((self.kernel_size-1)*self.dilations[i] + 1 - self.stride) / 2)
                _y = nn.functional.conv2d(x, weight=self.convolution.weight, bias=self.convolution.bias, \
                    stride=self.downsample_ratio, padding=padding, dilation=self.dilations[i]).unsqueeze(dim=-1)
                y = torch.cat((y, _y), dim=-1)
        else:
            y = self.convs[0](x).unsqueeze(dim=-1)
            for i in range(1, len(self.dilations)):
                _y = self.convs[i](x).unsqueeze(dim=-1)
                y = torch.cat((y, _y), dim=-1)
        B, C, W, H, N = y.shape
        if self.op == 'sum':
            y = y.sum(dim=-1).flatten(2).permute(0,2,1).contiguous()
        elif self.op == 'cat':
            y = y.permute(0,4,1,2,3).flatten(3).reshape(B, N*C, W*H).permute(0,2,1).contiguous()
        else:
            raise NotImplementedError('no such operation: {} for multi-levels!'.format(self.op))
        return y, (W, H)


class ReductionCell(nn.Module):
    def __init__(self, img_size=224, in_chans=3, embed_dims=64, token_dims=64, downsample_ratios=4, kernel_size=7,
                 num_heads=1, dilations=[1,2,3,4], share_weights=False, op='cat', tokens_type='performer', group=1,
                 relative_pos=False, drop=0., attn_drop=0., drop_path=0., mlp_ratio=1.0, gamma=False, init_values=1e-4, SE=False, window_size=7):
        super().__init__()

        self.img_size = img_size
        self.window_size = window_size
        self.op = op
        self.dilations = dilations
        self.num_heads = num_heads
        self.embed_dims = embed_dims
        self.token_dims = token_dims
        self.in_chans = in_chans
        self.downsample_ratios = downsample_ratios
        self.kernel_size = kernel_size
        self.outSize = img_size
        self.relative_pos = relative_pos
        PCMStride = []
        residual = downsample_ratios // 2
        for _ in range(3):
            PCMStride.append((residual > 0) + 1)
            residual = residual // 2
        assert residual == 0
        self.pool = None
        self.tokens_type = tokens_type
        if tokens_type == 'pooling':
            PCMStride = [1, 1, 1]
            self.pool = nn.MaxPool2d(downsample_ratios, stride=downsample_ratios, padding=0)
            tokens_type = 'transformer'
            self.outSize = self.outSize // downsample_ratios
            downsample_ratios = 1
        self.PCM = nn.Sequential(
                        nn.Conv2d(in_chans, embed_dims, kernel_size=(3, 3), stride=PCMStride[0], padding=(1, 1), groups=group),  # the 1st convolution
                        nn.BatchNorm2d(embed_dims),
                        nn.SiLU(inplace=True),
                        nn.Conv2d(embed_dims, embed_dims, kernel_size=(3, 3), stride=PCMStride[1], padding=(1, 1), groups=group),  # the 1st convolution
                        nn.BatchNorm2d(embed_dims),
                        nn.SiLU(inplace=True),
                        nn.Conv2d(embed_dims, token_dims, kernel_size=(3, 3), stride=PCMStride[2], padding=(1, 1), groups=group),  # the 1st convolution
                    )
        self.PRM = PRM(img_size=img_size, kernel_size=kernel_size, downsample_ratio=downsample_ratios, dilations=self.dilations,
            in_chans=in_chans, embed_dim=embed_dims, share_weights=share_weights, op=op)
        self.outSize = self.outSize // downsample_ratios

        in_chans = self.PRM.out_chans
        if tokens_type == 'performer':
            # assert num_heads == 1
            self.attn = Token_performer(dim=in_chans, in_dim=token_dims, head_cnt=num_heads, kernel_ratio=0.5, gamma=gamma, init_values=init_values)
        elif tokens_type == 'performer_less':
            self.attn = None
            self.PCM = None
        elif tokens_type == 'transformer':
            self.attn = Token_transformer(dim=in_chans, in_dim=token_dims, num_heads=num_heads, mlp_ratio=mlp_ratio, drop=drop, 
                                          attn_drop=attn_drop, drop_path=drop_path, gamma=gamma, init_values=init_values)
        elif tokens_type == 'swin':
            self.attn = SwinTransformerBlock(in_dim=in_chans, out_dim=token_dims, input_resolution=(self.img_size//self.downsample_ratios, self.img_size//self.downsample_ratios), 
                                            num_heads=num_heads, mlp_ratio=mlp_ratio, drop=drop,
                                            attn_drop=attn_drop, drop_path=drop_path, window_size=window_size, shift_size=0, relative_pos=relative_pos)

        if gamma:
            self.gamma2 = nn.Parameter(init_values * torch.ones((token_dims)),requires_grad=True)
            self.gamma3 = nn.Parameter(init_values * torch.ones((token_dims)),requires_grad=True)
        else:
            self.gamma2 = 1
            self.gamma3 = 1

        if SE:
            self.SE = SELayer(token_dims)
        else:
            self.SE = nn.Identity()

        self.num_patches = (img_size // 2) * (img_size // 2)  # there are 3 sfot split, stride are 4,2,2 seperately

    def forward(self, x):
        if len(x.shape) < 4:
            # B,N,C -> B,H,W,C
            B, N, C  = x.shape
            n = int(np.sqrt(N))
            x = x.view(B, n, n, C).contiguous()
            x = x.permute(0, 3, 1, 2)
        if self.pool is not None:
            x = self.pool(x)
        shortcut = x
        PRM_x, _ = self.PRM(x)
        if self.tokens_type == 'swin':
            pass
            B, N, C = PRM_x.shape
            H, W = self.img_size // self.downsample_ratios, self.img_size // self.downsample_ratios
            b, _, c = PRM_x.shape
            assert N == H*W
            x = self.attn.norm1(PRM_x)
            x = x.view(B, H, W, C)
            x_windows = window_partition(x, self.window_size)
            x_windows = x_windows.view(-1, self.window_size * self.window_size, C)
            attn_windows = self.attn.attn(x_windows, mask=self.attn.attn_mask)  # nW*B, window_size*window_size, C
            attn_windows = attn_windows.view(-1, self.window_size, self.window_size, self.token_dims)
            shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
            x = shifted_x
            x = x.view(B, H * W, self.token_dims)

            convX = self.PCM(shortcut)
            # B,C,H,W->B,H,W,C-> B,H*W,C->B,L,C
            convX = convX.permute(0, 2, 3, 1).view(*x.shape).contiguous()
            x = x + self.attn.drop_path(convX * self.gamma2)
            # x = shortcut + self.attn.drop_path(x)
            # x = x + self.attn.drop_path(self.attn.mlp(self.attn.norm2(x)))
            x = x + self.attn.drop_path(self.gamma3 * self.attn.mlp(self.attn.norm2(x)))
        else:
            if self.attn is None:
                return PRM_x
            convX = self.PCM(shortcut)
            x = self.attn.attn(self.attn.norm1(PRM_x))
            convX = convX.permute(0, 2, 3, 1).view(*x.shape).contiguous()
            x = x + self.attn.drop_path(convX * self.gamma2)
            x = x + self.attn.drop_path(self.gamma3 * self.attn.mlp(self.attn.norm2(x)))
        x = self.SE(x)

        return x


class BasicLayer(nn.Module):
    def __init__(self, img_size=224, in_chans=3, embed_dims=64, token_dims=64, downsample_ratios=4, kernel_size=7, RC_heads=1, NC_heads=6, dilations=[1, 2, 3, 4],
                RC_op='cat', RC_tokens_type='performer', NC_tokens_type='transformer', RC_group=1, NC_group=64, NC_depth=2, dpr=0.1, mlp_ratio=4., qkv_bias=True, 
                qk_scale=None, drop=0, attn_drop=0., norm_layer=nn.LayerNorm, class_token=False, gamma=False, init_values=1e-4, SE=False, window_size=7, relative_pos=False):
        super().__init__()
        self.img_size = img_size
        self.in_chans = in_chans
        self.embed_dims = embed_dims
        self.token_dims = token_dims
        self.downsample_ratios = downsample_ratios
        self.out_size = self.img_size // self.downsample_ratios
        self.RC_kernel_size = kernel_size
        self.RC_heads = RC_heads
        self.NC_heads = NC_heads
        self.dilations = dilations
        self.RC_op = RC_op
        self.RC_tokens_type = RC_tokens_type
        self.RC_group = RC_group
        self.NC_group = NC_group
        self.NC_depth = NC_depth
        self.relative_pos = relative_pos
        if RC_tokens_type == 'stem':
            self.RC = PatchEmbedding(inter_channel=token_dims//2, out_channels=token_dims, img_size=img_size)
        elif downsample_ratios > 1:
            self.RC = ReductionCell(img_size, in_chans, embed_dims, token_dims, downsample_ratios, kernel_size,
                            RC_heads, dilations, op=RC_op, tokens_type=RC_tokens_type, group=RC_group, gamma=gamma, init_values=init_values, SE=SE, relative_pos=relative_pos, window_size=window_size)
        else:
            self.RC = nn.Identity()
        self.NC = nn.ModuleList([
            NormalCell(token_dims, NC_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop, attn_drop=attn_drop,
                       drop_path=dpr[i] if isinstance(dpr, list) else dpr, norm_layer=norm_layer, class_token=class_token, group=NC_group, tokens_type=NC_tokens_type,
                       gamma=gamma, init_values=init_values, SE=SE, img_size=img_size // downsample_ratios, window_size=window_size, shift_size=0, relative_pos=relative_pos)
        for i in range(NC_depth)])

    def forward(self, x):
        x = self.RC(x)
        for nc in self.NC:
            x = nc(x)
        return x


# ============================================================================
# Main Model
# ============================================================================
class ViTAE_Window_NoShift_basic(nn.Module):
    def __init__(self, img_size=224, in_chans=3, stages=4, embed_dims=64, token_dims=64, downsample_ratios=[4, 2, 2, 2], kernel_size=[7, 3, 3, 3], 
                RC_heads=[1, 1, 1, 1], NC_heads=4, dilations=[[1, 2, 3, 4], [1, 2, 3], [1, 2], [1, 2]],
                RC_op='cat', RC_tokens_type=['performer', 'transformer', 'transformer', 'transformer'], NC_tokens_type='transformer',
                RC_group=[1, 1, 1, 1], NC_group=[1, 32, 64, 64], NC_depth=[2, 2, 6, 2], mlp_ratio=4., qkv_bias=True, qk_scale=None, drop_rate=0., 
                attn_drop_rate=0., drop_path_rate=0., norm_layer=partial(nn.LayerNorm, eps=1e-6), num_classes=1000,
                gamma=False, init_values=1e-4, SE=False, window_size=7, relative_pos=False):
        super().__init__()
        self.num_classes = num_classes
        self.stages = stages
        repeatOrNot = (lambda x, y, z=list: x if isinstance(x, z) else [x for _ in range(y)])
        self.embed_dims = repeatOrNot(embed_dims, stages)
        self.tokens_dims = token_dims if isinstance(token_dims, list) else [token_dims * (2 ** i) for i in range(stages)]
        self.downsample_ratios = repeatOrNot(downsample_ratios, stages)
        self.kernel_size = repeatOrNot(kernel_size, stages)
        self.RC_heads = repeatOrNot(RC_heads, stages)
        self.NC_heads = repeatOrNot(NC_heads, stages)
        self.dilaions = repeatOrNot(dilations, stages)
        self.RC_op = repeatOrNot(RC_op, stages)
        self.RC_tokens_type = repeatOrNot(RC_tokens_type, stages)
        self.NC_tokens_type = repeatOrNot(NC_tokens_type, stages)
        self.RC_group = repeatOrNot(RC_group, stages)
        self.NC_group = repeatOrNot(NC_group, stages)
        self.NC_depth = repeatOrNot(NC_depth, stages)
        self.mlp_ratio = repeatOrNot(mlp_ratio, stages)
        self.qkv_bias = repeatOrNot(qkv_bias, stages)
        self.qk_scale = repeatOrNot(qk_scale, stages)
        self.drop = repeatOrNot(drop_rate, stages)
        self.attn_drop = repeatOrNot(attn_drop_rate, stages)
        self.norm_layer = repeatOrNot(norm_layer, stages)
        self.relative_pos = relative_pos

        self.pos_drop = nn.Dropout(p=drop_rate)
        depth = np.sum(self.NC_depth)
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
        Layers = []
        for i in range(stages):
            startDpr = 0 if i==0 else self.NC_depth[i - 1]
            Layers.append(
                BasicLayer(img_size, in_chans, self.embed_dims[i], self.tokens_dims[i], self.downsample_ratios[i],
                self.kernel_size[i], self.RC_heads[i], self.NC_heads[i], self.dilaions[i], self.RC_op[i],
                self.RC_tokens_type[i], self.NC_tokens_type[i], self.RC_group[i], self.NC_group[i], self.NC_depth[i], dpr[startDpr:self.NC_depth[i]+startDpr],
                mlp_ratio=self.mlp_ratio[i], qkv_bias=self.qkv_bias[i], qk_scale=self.qk_scale[i], drop=self.drop[i], attn_drop=self.attn_drop[i],
                norm_layer=self.norm_layer[i], gamma=gamma, init_values=init_values, SE=SE, window_size=window_size, relative_pos=relative_pos)
            )
            img_size = img_size // self.downsample_ratios[i]
            in_chans = self.tokens_dims[i]
        self.layers = nn.ModuleList(Layers)

        # Classifier head
        self.head = nn.Linear(self.tokens_dims[-1], num_classes) if num_classes > 0 else nn.Identity()

        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            init.trunc_normal_(m.weight, std=0.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            init.constant_(m.bias, 0)
            init.constant_(m.weight, 1.0)

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

    def get_classifier(self):
        return self.head

    def reset_classifier(self, num_classes):
        self.num_classes = num_classes
        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()

    def forward_features(self, x):

        for layer in self.layers:
            x = layer(x)

        return torch.mean(x, 1)

    def forward(self, x):
        x = self.forward_features(x)
        x = self.head(x)
        return x
    
    def train(self, mode=True, tag='default'):
        self.training = mode
        if tag == 'default':
            for module in self.modules():
                if module.__class__.__name__ != 'ViTAE_Window_NoShift_basic':
                    module.train(mode)
        elif tag == 'linear':
            for module in self.modules():
                if module.__class__.__name__ != 'ViTAE_Window_NoShift_basic':
                    module.eval()
                    for param in module.parameters():
                        param.requires_grad = False
        elif tag == 'linearLNBN':
            for module in self.modules():
                if module.__class__.__name__ != 'ViTAE_Window_NoShift_basic':
                    if isinstance(module, nn.LayerNorm) or isinstance(module, nn.BatchNorm2d):
                        module.train(mode)
                        for param in module.parameters():
                            param.requires_grad = True
                    else:
                        module.eval()
                        for param in module.parameters():
                            param.requires_grad = False
        self.head.train(mode)
        for param in self.head.parameters():
            param.requires_grad = True
        return self


# ============================================================================
# Model Factory Functions
# ============================================================================
def _cfg(url='', **kwargs):
    return {
        'url': url,
        'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
        'crop_pct': .9, 'interpolation': 'bicubic',
        'mean': (0.485, 0.456, 0.406), 'std': (0.229, 0.224, 0.225),
        'classifier': 'head',
        **kwargs
    }

default_cfgs = {
    'ViTAE_stages3_7': _cfg(),
}

# @register_model
def ViTAE_Window_NoShift_12_basic_stages4_14(pretrained=False, **kwargs): # adopt performer for tokens to token
    # if pretrained:
        # kwargs.setdefault('qk_scale', 256 ** -0.5)
    model = ViTAE_Window_NoShift_basic(RC_tokens_type=['swin', 'swin', 'transformer', 'transformer'], NC_tokens_type=['swin', 'swin', 'transformer', 'transformer'], stages=4, embed_dims=[64, 64, 128, 256], token_dims=[64, 128, 256, 512], downsample_ratios=[4, 2, 2, 2],
                            NC_depth=[2, 2, 8, 2], NC_heads=[1, 2, 4, 8], RC_heads=[1, 1, 2, 4], mlp_ratio=4., NC_group=[1, 32, 64, 128], RC_group=[1, 16, 32, 64], **kwargs)
    model.default_cfg = default_cfgs['ViTAE_stages3_7']
    if pretrained:
        load_pretrained(
            model, num_classes=model.num_classes, in_chans=kwargs.get('in_chans', 3))
    return model