architecture/spectformer.py

import torch
import torch.nn as nn
import torch.nn.functional as F
from functools import partial

from timm.models.layers import DropPath, to_2tuple, trunc_normal_
from timm.models.registry import register_model
from timm.models.vision_transformer import _cfg
import math
import numpy as np
from pytorch_wavelets import DWTForward, DWTInverse # (or import DWT, IDWT)


class SpectralGatingNetwork(nn.Module):
    def __init__(self, dim):
        super().__init__()
        # this weights are valid for h=14 and w=8
        if dim == 64: #96 for large model, 64 for small and base model
            self.h = 56 #H
            self.w = 29 #(W/2)+1            
            self.complex_weight = nn.Parameter(torch.randn(self.h, self.w, dim, 2, dtype=torch.float32) * 0.02)
        if dim ==128:
            self.h = 28 #H
            self.w = 15 #(W/2)+1, this is due to rfft2
            self.complex_weight = nn.Parameter(torch.randn(self.h, self.w, dim, 2, dtype=torch.float32) * 0.02)
        if dim == 96: #96 for large model, 64 for small and base model
            self.h = 56 #H
            self.w = 29 #(W/2)+1            
            self.complex_weight = nn.Parameter(torch.randn(self.h, self.w, dim, 2, dtype=torch.float32) * 0.02)
        if dim ==192:
            self.h = 28 #H
            self.w = 15 #(W/2)+1, this is due to rfft2
            self.complex_weight = nn.Parameter(torch.randn(self.h, self.w, dim, 2, dtype=torch.float32) * 0.02)

    def forward(self, x, H, W):
        # print('wno',x.shape) #CIFAR100 image :[128, 196, 384]
        B, N, C = x.shape 
        # print('wno B, N, C',B, N, C) #CIFAR100 image : 128 196 384
        x = x.view(B, H, W, C)
        # B, H, W, C=x.shape
        x = x.to(torch.float32) 
        # print(x.dtype)
        # Add above for this error, RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same
        x = torch.fft.rfft2(x, dim=(1, 2), norm='ortho')
        # print('wno',x.shape)
        weight = torch.view_as_complex(self.complex_weight)
        # print('weight',weight.shape)
        x = x * weight
        x = torch.fft.irfft2(x, s=(H, W), dim=(1, 2), norm='ortho')
        # print('wno',x.shape)
        x = x.reshape(B, N, C)# permute is not same as reshape or view
        return x
        #return x, weight


def rand_bbox(size, lam, scale=1):
    W = size[1] // scale
    H = size[2] // scale
    cut_rat = np.sqrt(1. - lam)
    cut_w = np.int(W * cut_rat)
    cut_h = np.int(H * cut_rat)

    # uniform
    cx = np.random.randint(W)
    cy = np.random.randint(H)

    bbx1 = np.clip(cx - cut_w // 2, 0, W)
    bby1 = np.clip(cy - cut_h // 2, 0, H)
    bbx2 = np.clip(cx + cut_w // 2, 0, W)
    bby2 = np.clip(cy + cut_h // 2, 0, H)

    return bbx1, bby1, bbx2, bby2

class ClassAttention(nn.Module):
    def __init__(self, dim, num_heads):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.head_dim = head_dim
        self.scale = head_dim**-0.5
        self.kv = nn.Linear(dim, dim * 2)
        self.q = nn.Linear(dim, dim)
        self.proj = nn.Linear(dim, dim)
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.Conv2d):
            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
            fan_out //= m.groups
            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
            if m.bias is not None:
                m.bias.data.zero_()

    def forward(self, x):
        B, N, C = x.shape
        kv = self.kv(x).reshape(B, N, 2, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
        k, v = kv[0], kv[1]
        q = self.q(x[:, :1, :]).reshape(B, self.num_heads, 1, self.head_dim)
        attn = ((q * self.scale) @ k.transpose(-2, -1))
        attn = attn.softmax(dim=-1)
        cls_embed = (attn @ v).transpose(1, 2).reshape(B, 1, self.head_dim * self.num_heads)
        cls_embed = self.proj(cls_embed)
        return cls_embed

class FFN(nn.Module):
    def __init__(self, in_features, hidden_features):
        super().__init__()
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = nn.GELU()
        self.fc2 = nn.Linear(hidden_features, in_features)
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.Conv2d):
            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
            fan_out //= m.groups
            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
            if m.bias is not None:
                m.bias.data.zero_()

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

class ClassBlock(nn.Module):
    def __init__(self, dim, num_heads, mlp_ratio, norm_layer=nn.LayerNorm):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.norm2 = norm_layer(dim)
        self.attn = ClassAttention(dim, num_heads)
        self.mlp = FFN(dim, int(dim * mlp_ratio))
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.Conv2d):
            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
            fan_out //= m.groups
            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
            if m.bias is not None:
                m.bias.data.zero_()

    def forward(self, x):
        cls_embed = x[:, :1]
        cls_embed = cls_embed + self.attn(self.norm1(x))
        cls_embed = cls_embed + self.mlp(self.norm2(cls_embed))
        return torch.cat([cls_embed, x[:, 1:]], dim=1)

class PVT2FFN(nn.Module):
    def __init__(self, in_features, hidden_features):
        super().__init__()
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.dwconv = DWConv(hidden_features)
        self.act = nn.GELU()
        self.fc2 = nn.Linear(hidden_features, in_features)
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.Conv2d):
            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
            fan_out //= m.groups
            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
            if m.bias is not None:
                m.bias.data.zero_()

    def forward(self, x, H, W):
        x = self.fc1(x)
        x = self.dwconv(x, H, W)
        x = self.act(x)
        x = self.fc2(x)
        return x

class Attention(nn.Module):
    def __init__(self, dim, num_heads):
        super().__init__()
        assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."

        self.dim = dim
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = head_dim ** -0.5

        self.q = nn.Linear(dim, dim)
        self.kv = nn.Linear(dim, dim * 2)
        self.proj = nn.Linear(dim, dim)
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.Conv2d):
            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
            fan_out //= m.groups
            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
            if m.bias is not None:
                m.bias.data.zero_()

    def forward(self, x, H, W):
        B, N, C = x.shape
        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
        kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        k, v = kv[0], kv[1]
        attn = (q @ k.transpose(-2, -1)) * self.scale
        attn = attn.softmax(dim=-1)
        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        #return x
        return x, attn

class Block(nn.Module):
    def __init__(self, 
        dim, 
        num_heads, 
        mlp_ratio,
        drop_path=0., 
        norm_layer=nn.LayerNorm, 
        sr_ratio=1, 
        block_type = 'wave'
    ):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.norm2 = norm_layer(dim)

        if block_type == 'std_att':
            self.attn = Attention(dim, num_heads)
        else:
            self.attn = SpectralGatingNetwork(dim)
        self.mlp = PVT2FFN(in_features=dim, hidden_features=int(dim * mlp_ratio))
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.Conv2d):
            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
            fan_out //= m.groups
            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
            if m.bias is not None:
                m.bias.data.zero_()

    # def forward(self, x, H, W): ## !!!!!!!!!!!!!!!!
    #     x = x + self.drop_path(self.attn(self.norm1(x), H, W)) 
    #     x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
    #     return x
    
    
    def forward(self, x, H, W):
        attn_output, attn_weights = self.attn(self.norm1(x), H, W) if isinstance(self.attn, Attention) else (self.attn(self.norm1(x), H, W), None)
        x = x + self.drop_path(attn_output)
        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))

        # Optionally return attention weights for visualization or analysis
        return (x, attn_weights) if attn_weights is not None else x


class DownSamples(nn.Module):
    def __init__(self, in_channels, out_channels):
        super().__init__()
        self.proj = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=2, padding=1)
        self.norm = nn.LayerNorm(out_channels)
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.Conv2d):
            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
            fan_out //= m.groups
            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
            if m.bias is not None:
                m.bias.data.zero_()

    def forward(self, x):
        x = self.proj(x)
        _, _, H, W = x.shape
        x = x.flatten(2).transpose(1, 2)
        x = self.norm(x)
        return x, H, W

class Stem(nn.Module):
    def __init__(self, in_channels, stem_hidden_dim, out_channels):
        super().__init__()
        hidden_dim = stem_hidden_dim
        self.conv = nn.Sequential(
            nn.Conv2d(in_channels, hidden_dim, kernel_size=7, stride=2,
                      padding=3, bias=False),  # 112x112
            nn.BatchNorm2d(hidden_dim),
            nn.ReLU(inplace=True),
            nn.Conv2d(hidden_dim, hidden_dim, kernel_size=3, stride=1,
                      padding=1, bias=False),  # 112x112
            nn.BatchNorm2d(hidden_dim),
            nn.ReLU(inplace=True),
            nn.Conv2d(hidden_dim, hidden_dim, kernel_size=3, stride=1,
                      padding=1, bias=False),  # 112x112
            nn.BatchNorm2d(hidden_dim),
            nn.ReLU(inplace=True),
        )
        self.proj = nn.Conv2d(hidden_dim,
                              out_channels,
                              kernel_size=3,
                              stride=2,
                              padding=1)
        self.norm = nn.LayerNorm(out_channels)

        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.Conv2d):
            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
            fan_out //= m.groups
            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
            if m.bias is not None:
                m.bias.data.zero_()

    def forward(self, x):
        x = self.conv(x)
        x = self.proj(x)
        _, _, H, W = x.shape
        x = x.flatten(2).transpose(1, 2)
        x = self.norm(x)
        return x, H, W

class SpectFormer(nn.Module):
    def __init__(self, 
        in_chans=3, 
        num_classes=1000, 
        stem_hidden_dim = 32,
        embed_dims=[64, 128, 320, 448],
        num_heads=[2, 4, 10, 14], 
        mlp_ratios=[8, 8, 4, 4], 
        drop_path_rate=0., 
        norm_layer=nn.LayerNorm,
        depths=[3, 4, 6, 3], 
        sr_ratios=[4, 2, 1, 1], 
        num_stages=4,
        token_label=False,
        **kwargs
    ):
        super().__init__()
        self.num_classes = num_classes
        self.depths = depths
        self.num_stages = num_stages

        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
        cur = 0

        for i in range(num_stages):
            if i == 0:
                patch_embed = Stem(in_chans, stem_hidden_dim, embed_dims[i])
            else:
                patch_embed = DownSamples(embed_dims[i - 1], embed_dims[i])

            block = nn.ModuleList([Block(
                dim = embed_dims[i], 
                num_heads = num_heads[i], 
                mlp_ratio = mlp_ratios[i], 
                drop_path=dpr[cur + j], 
                norm_layer=norm_layer,
                sr_ratio = sr_ratios[i],
                block_type='wave' if i < 2 else 'std_att')
            for j in range(depths[i])])

            norm = norm_layer(embed_dims[i])
            cur += depths[i]

            setattr(self, f"patch_embed{i + 1}", patch_embed)
            setattr(self, f"block{i + 1}", block)
            setattr(self, f"norm{i + 1}", norm)

        post_layers = ['ca']
        self.post_network = nn.ModuleList([
            ClassBlock(
                dim = embed_dims[-1], 
                num_heads = num_heads[-1], 
                mlp_ratio = mlp_ratios[-1],
                norm_layer=norm_layer)
            for _ in range(len(post_layers))
        ])

        # classification head
        self.head = nn.Linear(embed_dims[-1], num_classes) if num_classes > 0 else nn.Identity()
        ##################################### token_label #####################################
        self.return_dense = token_label
        self.mix_token = token_label
        self.beta = 1.0
        self.pooling_scale = 8
        if self.return_dense:
            self.aux_head = nn.Linear(
                embed_dims[-1],
                num_classes) if num_classes > 0 else nn.Identity()
        ##################################### token_label #####################################

        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.Conv2d):
            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
            fan_out //= m.groups
            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
            if m.bias is not None:
                m.bias.data.zero_()

    def forward_cls(self, x):
        B, N, C = x.shape
        cls_tokens = x.mean(dim=1, keepdim=True)
        x = torch.cat((cls_tokens, x), dim=1)
        for block in self.post_network:
            x = block(x)
        return x

    # def forward_features(self, x):
    #     B = x.shape[0]
    #     for i in range(self.num_stages):
    #         patch_embed = getattr(self, f"patch_embed{i + 1}")
    #         block = getattr(self, f"block{i + 1}")
    #         x, H, W = patch_embed(x)
    #         for blk in block:
    #             x = blk(x, H, W)
    #             tokens = x
            
    #         if i != self.num_stages - 1:
    #             norm = getattr(self, f"norm{i + 1}")
    #             x = norm(x)
    #             x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()

    #     x = self.forward_cls(x)[:, 0]
    #     norm = getattr(self, f"norm{self.num_stages}")
    #     x = norm(x)
    #     return x, tokens
    
    def forward_features(self, x):
        B = x.shape[0]
        attention_maps = []  # Collect attention maps if available
        tokens = None  # Initialize tokens to ensure scope coverage
    
        for i in range(self.num_stages):
            patch_embed = getattr(self, f"patch_embed{i + 1}")
            block = getattr(self, f"block{i + 1}")
            x, H, W = patch_embed(x)
            
            for blk in block:
                outputs = blk(x, H, W)
                if isinstance(outputs, tuple):
                    x, attn_weights = outputs
                    attention_maps.append(attn_weights)  # Store attention maps
                else:
                    x = outputs
            
            tokens = x  # Update tokens with the latest block output
    
            if i != self.num_stages - 1:
                norm = getattr(self, f"norm{i + 1}")
                x = norm(x)
                x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
    
        x = self.forward_cls(x)[:, 0]  # Further processing for classification token
        norm = getattr(self, f"norm{self.num_stages}")
        x = norm(x)
        return x, tokens, attention_maps


    
    # def forward(self, x):
    #     if not self.return_dense:
    #         x, tokens = self.forward_features(x)
    #         x = self.head(x)
    #         return x, tokens
    #     else:
    #         x, H, W = self.forward_embeddings(x)
    #         # mix token, see token labeling for details.
    #         if self.mix_token and self.training:
    #             lam = np.random.beta(self.beta, self.beta)
    #             patch_h, patch_w = x.shape[1] // self.pooling_scale, x.shape[
    #                 2] // self.pooling_scale
    #             bbx1, bby1, bbx2, bby2 = rand_bbox(x.size(), lam, scale=self.pooling_scale)
    #             temp_x = x.clone()
    #             sbbx1,sbby1,sbbx2,sbby2=self.pooling_scale*bbx1,self.pooling_scale*bby1,\
    #                                     self.pooling_scale*bbx2,self.pooling_scale*bby2
    #             temp_x[:, sbbx1:sbbx2, sbby1:sbby2, :] = x.flip(0)[:, sbbx1:sbbx2, sbby1:sbby2, :]
    #             x = temp_x
    #         else:
    #             bbx1, bby1, bbx2, bby2 = 0, 0, 0, 0
            
    #         x = self.forward_tokens(x, H, W)
    #         x_cls = self.head(x[:, 0])
    #         x_aux = self.aux_head(
    #             x[:, 1:]
    #         )  # generate classes in all feature tokens, see token labeling

    #         if not self.training:
    #             return x_cls + 0.5 * x_aux.max(1)[0]

    #         if self.mix_token and self.training:  # reverse "mix token", see token labeling for details.
    #             x_aux = x_aux.reshape(x_aux.shape[0], patch_h, patch_w, x_aux.shape[-1])

    #             temp_x = x_aux.clone()
    #             temp_x[:, bbx1:bbx2, bby1:bby2, :] = x_aux.flip(0)[:, bbx1:bbx2, bby1:bby2, :]
    #             x_aux = temp_x

    #             x_aux = x_aux.reshape(x_aux.shape[0], patch_h * patch_w, x_aux.shape[-1])

    #         return x_cls, x_aux, (bbx1, bby1, bbx2, bby2)


    def forward(self, x):
        attention_maps = []  # Initialize to collect attention maps from all blocks
    
        if not self.return_dense:
            # Retrieve main output, tokens, and attention maps
            x, tokens, new_attention_maps = self.forward_features(x)
            attention_maps.extend(new_attention_maps)  # Collect new attention maps
            x = self.head(x)
            return x, tokens, attention_maps
        else:
            # For dense token labeling and feature manipulation
            x, H, W = self.forward_embeddings(x)
            x, new_attention_maps = self.forward_tokens(x, H, W)  # Adjusted to return attention maps
            attention_maps.extend(new_attention_maps)  # Collect new attention maps
    
            if self.mix_token and self.training:
                lam = np.random.beta(self.beta, self.beta)
                patch_h, patch_w = x.shape[1] // self.pooling_scale, x.shape[2] // self.pooling_scale
                bbx1, bby1, bbx2, bby2 = rand_bbox(x.size(), lam, scale=self.pooling_scale)
                sbbx1, sbby1, sbbx2, sbby2 = self.pooling_scale * bbx1, self.pooling_scale * bby1, self.pooling_scale * bbx2, self.pooling_scale * bby2
                temp_x = x.clone()
                temp_x[:, sbbx1:sbbx2, sbby1:sbby2, :] = x.flip(0)[:, sbbx1:sbbx2, sbby1:sbby2, :]
                x = temp_x
            else:
                bbx1, bby1, bbx2, bby2 = 0, 0, 0, 0  # Default to zero if no mixing
    
            x_cls = self.head(x[:, 0])
            x_aux = self.aux_head(x[:, 1:])  # Class prediction for all feature tokens
    
            if not self.training:
                return x_cls + 0.5 * x_aux.max(1)[0], attention_maps
    
            return x_cls, x_aux, (bbx1, bby1, bbx2, bby2), attention_maps







    def forward_tokens(self, x, H, W):
        B = x.shape[0]
        x = x.view(B, -1, x.size(-1))

        for i in range(self.num_stages):
            if i != 0:
                patch_embed = getattr(self, f"patch_embed{i + 1}")
                x, H, W = patch_embed(x)

            block = getattr(self, f"block{i + 1}")
            for blk in block:
                x = blk(x, H, W)

            if i != self.num_stages - 1:
                norm = getattr(self, f"norm{i + 1}")
                x = norm(x)
                x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()

        x = self.forward_cls(x)
        norm = getattr(self, f"norm{self.num_stages}")
        x = norm(x)    
        return x

    def forward_embeddings(self, x):
        patch_embed = getattr(self, f"patch_embed{0 + 1}")
        x, H, W = patch_embed(x)
        x = x.view(x.size(0), H, W, -1)
        return x, H, W


class DWConv(nn.Module):
    def __init__(self, dim=768):
        super(DWConv, self).__init__()
        self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)

    def forward(self, x, H, W):
        B, N, C = x.shape
        x = x.transpose(1, 2).view(B, C, H, W)
        x = self.dwconv(x)
        x = x.flatten(2).transpose(1, 2)
        return x


@register_model
def spectformer_t_d(pretrained=False, **kwargs):
    model = SpectFormer(
        stem_hidden_dim = 32,
        embed_dims = [64, 128, 160, 400], #64, 128, 320, 448 -----[64, 128, 160, 200]
        num_heads = [2, 4, 10, 16],  #2, 4, 10, 16 ----------[2, 4, 10, 10]
        mlp_ratios = [8, 8, 4, 4],
        norm_layer = partial(nn.LayerNorm, eps=1e-6), 
        depths = [1, 2, 5, 2],   #1, 2, 3, 1 ---------[1, 1, 1, 1]
        sr_ratios = [4, 2, 1, 1], 
        **kwargs)
    model.default_cfg = _cfg()
    return model

@register_model
def spectformer_t_w(pretrained=False, **kwargs):
    model = SpectFormer(
        stem_hidden_dim = 32,
        embed_dims = [64, 128, 320, 96], #64, 128, 320, 448 -----[64, 128, 160, 200]
        num_heads = [2, 4, 10, 16],  #2, 4, 10, 16 ----------[2, 4, 10, 10]
        mlp_ratios = [8, 8, 4, 4],
        norm_layer = partial(nn.LayerNorm, eps=1e-6), 
        depths = [1, 1, 1, 1],   #1, 2, 3, 1 ---------[1, 1, 1, 1]
        sr_ratios = [4, 2, 1, 1], 
        **kwargs)
    model.default_cfg = _cfg()
    return model