import math import torch import torch.nn as nn import torch.nn.functional as F import torch.utils.checkpoint as checkpoint from timm.models.layers import DropPath, to_2tuple, trunc_normal_, _assert from torchvision.transforms import functional as TF from timm.models.fx_features import register_notrace_function import numpy as np from einops import rearrange from huggingface_hub import PyTorchModelHubMixin from utils import FileClient, imfrombytes, img2tensor, tensor2img class DFE(nn.Module): """ Dual Feature Extraction Args: in_features (int): Number of input channels. out_features (int): Number of output channels. """ def __init__(self, in_features, out_features): super().__init__() self.out_features = out_features self.conv = nn.Sequential(nn.Conv2d(in_features, in_features // 5, 1, 1, 0), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Conv2d(in_features // 5, in_features // 5, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Conv2d(in_features // 5, out_features, 1, 1, 0)) self.linear = nn.Conv2d(in_features, out_features,1,1,0) def forward(self, x, x_size): B, L, C = x.shape H, W = x_size x = x.permute(0, 2, 1).contiguous().view(B, C, H, W) x = self.conv(x) * self.linear(x) x = x.view(B, -1, H*W).permute(0,2,1).contiguous() return x class Mlp(nn.Module): """ MLP-based Feed-Forward Network Args: in_features (int): Number of input channels. hidden_features (int | None): Number of hidden channels. Default: None out_features (int | None): Number of output channels. Default: None act_layer (nn.Module): Activation layer. Default: nn.GELU drop (float): Dropout rate. Default: 0.0 """ 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.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 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 wh, ww = H//window_size, W//window_size 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, (wh, ww) @register_notrace_function # reason: int argument is a Proxy def window_unpartition(windows, num_windows): """ Args: windows: [B*wh*ww, WH, WW, D] num_windows (tuple[int]): The height and width of the window. Returns: x: [B, ph, pw, D] """ x = rearrange(windows, '(p h w) wh ww c -> p (h wh) (w ww) c', h=num_windows[0], w=num_windows[1]) return x.contiguous() def window_reverse(windows, window_size, H, W): """ Args: windows: (num_windows*B, window_size, window_size, C) window_size (tuple): Window size H (int): Height of image W (int): Width of image Returns: x: (B, H, W, C) """ B = int(windows.shape[0] * (window_size[0] * window_size[1]) / (H * W)) x = windows.view(B, H // window_size[0], W // window_size[1], window_size[0], window_size[1], -1) x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) return x class DynamicPosBias(nn.Module): # The implementation builds on Crossformer code https://github.com/cheerss/CrossFormer/blob/main/models/crossformer.py """ Dynamic Relative Position Bias. Args: dim (int): Number of input channels. num_heads (int): Number of heads for spatial self-correlation. residual (bool): If True, use residual strage to connect conv. """ def __init__(self, dim, num_heads, residual): super().__init__() self.residual = residual self.num_heads = num_heads self.pos_dim = dim // 4 self.pos_proj = nn.Linear(2, self.pos_dim) self.pos1 = nn.Sequential( nn.LayerNorm(self.pos_dim), nn.ReLU(inplace=True), nn.Linear(self.pos_dim, self.pos_dim), ) self.pos2 = nn.Sequential( nn.LayerNorm(self.pos_dim), nn.ReLU(inplace=True), nn.Linear(self.pos_dim, self.pos_dim) ) self.pos3 = nn.Sequential( nn.LayerNorm(self.pos_dim), nn.ReLU(inplace=True), nn.Linear(self.pos_dim, self.num_heads) ) def forward(self, biases): if self.residual: pos = self.pos_proj(biases) # 2Gh-1 * 2Gw-1, heads pos = pos + self.pos1(pos) pos = pos + self.pos2(pos) pos = self.pos3(pos) else: pos = self.pos3(self.pos2(self.pos1(self.pos_proj(biases)))) return pos class SCC(nn.Module): """ Spatial-Channel Correlation. Args: dim (int): Number of input channels. base_win_size (tuple[int]): The height and width of the base window. window_size (tuple[int]): The height and width of the window. num_heads (int): Number of heads for spatial self-correlation. value_drop (float, optional): Dropout ratio of value. Default: 0.0 proj_drop (float, optional): Dropout ratio of output. Default: 0.0 """ def __init__(self, dim, base_win_size, window_size, num_heads, value_drop=0., proj_drop=0.): super().__init__() # parameters self.dim = dim self.window_size = window_size self.num_heads = num_heads # feature projection head_dim = dim // (2*num_heads) if dim % (2*num_heads) > 0: head_dim = head_dim + 1 self.attn_dim = head_dim * 2 * num_heads self.qv = DFE(dim, self.attn_dim) self.proj = nn.Linear(self.attn_dim, dim) # dropout self.value_drop = nn.Dropout(value_drop) self.proj_drop = nn.Dropout(proj_drop) # base window size min_h = min(self.window_size[0], base_win_size[0]) min_w = min(self.window_size[1], base_win_size[1]) self.base_win_size = (min_h, min_w) # normalization factor and spatial linear layer for S-SC self.scale = head_dim self.spatial_linear = nn.Linear(self.window_size[0]*self.window_size[1] // (self.base_win_size[0]*self.base_win_size[1]), 1) # NGram window partition without shifting self.ngram_window_partition = NGramWindowPartition(dim, window_size, 2, num_heads, shift_size=0) # define a parameter table of relative position bias self.H_sp, self.W_sp = self.window_size self.pos = DynamicPosBias(self.dim // 4, self.num_heads, residual=False) def spatial_linear_projection(self, x): B, num_h, L, C = x.shape H, W = self.window_size map_H, map_W = self.base_win_size x = x.view(B, num_h, map_H, H//map_H, map_W, W//map_W, C).permute(0,1,2,4,6,3,5).contiguous().view(B, num_h, map_H*map_W, C, -1) x = self.spatial_linear(x).view(B, num_h, map_H*map_W, C) return x def spatial_self_correlation(self, q, v): B, num_head, L, C = q.shape # spatial projection v = self.spatial_linear_projection(v) # compute correlation map corr_map = (q @ v.transpose(-2,-1)) / self.scale # add relative position bias position_bias_h = torch.arange(1 - self.H_sp, self.H_sp, device=v.device) position_bias_w = torch.arange(1 - self.W_sp, self.W_sp, device=v.device) biases = torch.stack(torch.meshgrid([position_bias_h, position_bias_w])) rpe_biases = biases.flatten(1).transpose(0, 1).contiguous().float() pos = self.pos(rpe_biases) # select position bias coords_h = torch.arange(self.H_sp, device=v.device) coords_w = torch.arange(self.W_sp, device=v.device) coords = torch.stack(torch.meshgrid([coords_h, coords_w])) coords_flatten = torch.flatten(coords, 1) relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] relative_coords = relative_coords.permute(1, 2, 0).contiguous() relative_coords[:, :, 0] += self.H_sp - 1 relative_coords[:, :, 1] += self.W_sp - 1 relative_coords[:, :, 0] *= 2 * self.W_sp - 1 relative_position_index = relative_coords.sum(-1) relative_position_bias = pos[relative_position_index.view(-1)].view( self.window_size[0] * self.window_size[1], self.base_win_size[0], self.window_size[0]//self.base_win_size[0], self.base_win_size[1], self.window_size[1]//self.base_win_size[1], -1) # Wh*Ww,Wh*Ww,nH relative_position_bias = relative_position_bias.permute(0,1,3,5,2,4).contiguous().view( self.window_size[0] * self.window_size[1], self.base_win_size[0]*self.base_win_size[1], self.num_heads, -1).mean(-1) relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() corr_map = corr_map + relative_position_bias.unsqueeze(0) # transformation v_drop = self.value_drop(v) x = (corr_map @ v_drop).permute(0,2,1,3).contiguous().view(B, L, -1) return x def channel_self_correlation(self, q, v): B, num_head, L, C = q.shape # apply single head strategy q = q.permute(0,2,1,3).contiguous().view(B, L, num_head*C) v = v.permute(0,2,1,3).contiguous().view(B, L, num_head*C) # compute correlation map corr_map = (q.transpose(-2,-1) @ v) / L # transformation v_drop = self.value_drop(v) x = (corr_map @ v_drop.transpose(-2,-1)).permute(0,2,1).contiguous().view(B, L, -1) return x def forward(self, x): """ Args: x: input features with shape of (B, H, W, C) """ xB,xH,xW,xC = x.shape qv = self.qv(x.view(xB,-1,xC), (xH,xW)).view(xB, xH, xW, xC) # window partition qv = self.ngram_window_partition(qv) qv = qv.view(-1, self.window_size[0]*self.window_size[1], xC) # qv splitting B, L, C = qv.shape qv = qv.view(B, L, 2, self.num_heads, C // (2*self.num_heads)).permute(2,0,3,1,4).contiguous() q, v = qv[0], qv[1] # B, num_heads, L, C//num_heads # spatial self-correlation (S-SC) x_spatial = self.spatial_self_correlation(q, v) x_spatial = x_spatial.view(-1, self.window_size[0], self.window_size[1], C//2) x_spatial = window_reverse(x_spatial, (self.window_size[0],self.window_size[1]), xH, xW) # xB xH xW xC # channel self-correlation (C-SC) x_channel = self.channel_self_correlation(q, v) x_channel = x_channel.view(-1, self.window_size[0], self.window_size[1], C//2) x_channel = window_reverse(x_channel, (self.window_size[0], self.window_size[1]), xH, xW) # xB xH xW xC # spatial-channel information fusion x = torch.cat([x_spatial, x_channel], -1) x = self.proj_drop(self.proj(x)) return x def extra_repr(self) -> str: return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}' class NGramWindowAttention(nn.Module): r""" Window based multi-head self attention (W-MSA) module with relative position bias for NGram attention. It supports both of shifted and non-shifted window. Args: dim (int): Number of input channels. window_size (tuple[int]): The height and width of the window. num_heads (int): Number of attention heads. qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0 proj_drop (float, optional): Dropout ratio of output. Default: 0.0 """ def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.): super().__init__() self.dim = dim self.window_size = window_size # Wh, Ww self.num_heads = num_heads head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 # 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 # 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(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) trunc_normal_(self.relative_position_bias_table, std=.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, C // self.num_heads).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)) 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, C) x = self.proj(x) x = self.proj_drop(x) return x def extra_repr(self) -> str: return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}' class NGramContext(nn.Module): ''' Args: dim (int): Number of input channels. window_size (int or tuple[int]): The height and width of the window. ngram (int): How much windows(or patches) to see. ngram_num_heads (int): padding_mode (str, optional): How to pad. Default: seq_refl_win_pad Options: ['seq_refl_win_pad', 'zero_pad'] ''' def __init__(self, dim, window_size, ngram, ngram_num_heads, padding_mode='seq_refl_win_pad'): super(NGramContext, self).__init__() _assert(padding_mode in ['seq_refl_win_pad', 'zero_pad'], "padding mode should be 'seq_refl_win_pad' or 'zero_pad'!!") self.dim = dim self.window_size = to_2tuple(window_size) self.ngram = ngram self.padding_mode = padding_mode # to alleviate parameter expansion with window sizes self.unigram_embed = nn.Conv2d(2, 1, kernel_size=(self.window_size[0], self.window_size[1]), stride=self.window_size, padding=0, groups=1) self.ngram_attn = NGramWindowAttention(dim=dim//2, num_heads=ngram_num_heads, window_size=(ngram, ngram)) self.avg_pool = nn.AvgPool2d(ngram) self.merge = nn.Conv2d(dim, dim, 1, 1, 0) def seq_refl_win_pad(self, x, back=False): if self.ngram == 1: return x x = TF.pad(x, (0,0,self.ngram-1,self.ngram-1)) if not back else TF.pad(x, (self.ngram-1,self.ngram-1,0,0)) if self.padding_mode == 'zero_pad': return x if not back: (start_h, start_w), (end_h, end_w) = to_2tuple(-2*self.ngram+1), to_2tuple(-self.ngram) # pad lower x[:,:,-(self.ngram-1):,:] = x[:,:,start_h:end_h,:] # pad right x[:,:,:,-(self.ngram-1):] = x[:,:,:,start_w:end_w] else: (start_h, start_w), (end_h, end_w) = to_2tuple(self.ngram), to_2tuple(2*self.ngram-1) # pad upper x[:,:,:self.ngram-1,:] = x[:,:,start_h:end_h,:] # pad left x[:,:,:,:self.ngram-1] = x[:,:,:,start_w:end_w] return x def sliding_window_attention(self, unigram): slide = unigram.unfold(3, self.ngram, 1).unfold(2, self.ngram, 1) slide = rearrange(slide, 'b c h w ww hh -> b (h hh) (w ww) c') # [B, 2(wh+ngram-2), 2(ww+ngram-2), D/2] slide, num_windows = window_partition(slide, self.ngram) # [B*wh*ww, ngram, ngram, D/2], (wh, ww) slide = slide.view(-1, self.ngram*self.ngram, self.dim//2) # [B*wh*ww, ngram*ngram, D/2] context = self.ngram_attn(slide).view(-1, self.ngram, self.ngram, self.dim//2) # [B*wh*ww, ngram, ngram, D/2] context = window_unpartition(context, num_windows) # [B, wh*ngram, ww*ngram, D/2] context = rearrange(context, 'b h w d -> b d h w') # [B, D/2, wh*ngram, ww*ngram] context = self.avg_pool(context) # [B, D/2, wh, ww] return context def forward(self, x): B, ph, pw, D = x.size() x = rearrange(x, 'b ph pw d -> b d ph pw') # [B, D, ph, pw] x = x.contiguous().view(B*(D//2),2,ph,pw) unigram = self.unigram_embed(x).view(B, D//2, ph//self.window_size[0], pw//self.window_size[1]) unigram_forward_pad = self.seq_refl_win_pad(unigram, False) # [B, D/2, wh+ngram-1, ww+ngram-1] unigram_backward_pad = self.seq_refl_win_pad(unigram, True) # [B, D/2, wh+ngram-1, ww+ngram-1] context_forward = self.sliding_window_attention(unigram_forward_pad) # [B, D/2, wh, ww] context_backward = self.sliding_window_attention(unigram_backward_pad) # [B, D/2, wh, ww] context_bidirect = torch.cat([context_forward, context_backward], dim=1) # [B, D, wh, ww] context_bidirect = self.merge(context_bidirect) # [B, D, wh, ww] context_bidirect = rearrange(context_bidirect, 'b d h w -> b h w d') # [B, wh, ww, D] return context_bidirect.unsqueeze(-2).unsqueeze(-2).contiguous() # [B, wh, ww, 1, 1, D] class NGramWindowPartition(nn.Module): """ Args: dim (int): Number of input channels. window_size (int): The height and width of the window. ngram (int): How much windows to see as context. ngram_num_heads (int): shift_size (int, optional): Shift size for SW-MSA. Default: 0 """ def __init__(self, dim, window_size, ngram, ngram_num_heads, shift_size=0): super(NGramWindowPartition, self).__init__() self.window_size = window_size[0] self.ngram = ngram self.shift_size = shift_size self.ngram_context = NGramContext(dim, window_size, ngram, ngram_num_heads, padding_mode='seq_refl_win_pad') def forward(self, x): B, ph, pw, D = x.size() wh, ww = ph//self.window_size, pw//self.window_size # number of windows (height, width) _assert(0 not in [wh, ww], "feature map size should be larger than window size!") context = self.ngram_context(x) # [B, wh, ww, 1, 1, D] windows = rearrange(x, 'b (h wh) (w ww) c -> b h w wh ww c', wh=self.window_size, ww=self.window_size).contiguous() # [B, wh, ww, WH, WW, D]. semi window partitioning windows+=context # [B, wh, ww, WH, WW, D]. inject context # Cyclic Shift if self.shift_size>0: x = rearrange(windows, 'b h w wh ww c -> b (h wh) (w ww) c').contiguous() # [B, ph, pw, D]. re-patchfying shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)) # [B, ph, pw, D]. cyclic shift windows = rearrange(shifted_x, 'b (h wh) (w ww) c -> b h w wh ww c', wh=self.window_size, ww=self.window_size).contiguous() # [B, wh, ww, WH, WW, D]. re-semi window partitioning windows = rearrange(windows, 'b h w wh ww c -> (b h w) wh ww c').contiguous() # [B*wh*ww, WH, WW, D]. window partitioning return windows class HierarchicalTransformerBlock(nn.Module): """ Hierarchical Transformer Block. Args: dim (int): Number of input channels. input_resolution (tuple[int]): Input resulotion. num_heads (int): Number of heads for spatial self-correlation. base_win_size (tuple[int]): The height and width of the base window. window_size (tuple[int]): The height and width of the window. mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. drop (float, optional): Dropout rate. Default: 0.0 value_drop (float, optional): Dropout ratio of value. Default: 0.0 drop_path (float, optional): Stochastic depth rate. Default: 0.0 act_layer (nn.Module, optional): Activation layer. Default: nn.GELU norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm """ def __init__(self, dim, input_resolution, num_heads, base_win_size, window_size, mlp_ratio=4., drop=0., value_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm): super().__init__() self.dim = dim self.input_resolution = input_resolution self.num_heads = num_heads self.window_size = window_size self.mlp_ratio = mlp_ratio # check window size if (window_size[0] > base_win_size[0]) and (window_size[1] > base_win_size[1]): assert window_size[0] % base_win_size[0] == 0, "please ensure the window size is smaller than or divisible by the base window size" assert window_size[1] % base_win_size[1] == 0, "please ensure the window size is smaller than or divisible by the base window size" self.norm1 = norm_layer(dim) self.correlation = SCC( dim, base_win_size=base_win_size, window_size=self.window_size, num_heads=num_heads, value_drop=value_drop, proj_drop=drop) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) def check_image_size(self, x, win_size): x = x.permute(0,3,1,2).contiguous() _, _, h, w = x.size() mod_pad_h = (win_size[0] - h % win_size[0]) % win_size[0] mod_pad_w = (win_size[1] - w % win_size[1]) % win_size[1] if mod_pad_h >= h or mod_pad_w >= w: pad_h, pad_w = h-1, w-1 x = F.pad(x, (0, pad_w, 0, pad_h), 'reflect') else: pad_h, pad_w = 0, 0 mod_pad_h = mod_pad_h - pad_h mod_pad_w = mod_pad_w - pad_w x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), 'reflect') x = x.permute(0,2,3,1).contiguous() return x def forward(self, x, x_size, win_size): H, W = x_size B, L, C = x.shape shortcut = x x = x.view(B, H, W, C) # padding x = self.check_image_size(x, (win_size[0]*2, win_size[1]*2)) _, H_pad, W_pad, _ = x.shape # shape after padding x = self.correlation(x) # unpad x = x[:, :H, :W, :].contiguous() # norm x = x.view(B, H * W, C) x = self.norm1(x) # FFN x = shortcut + self.drop_path(x) x = x + self.drop_path(self.norm2(self.mlp(x))) return x def extra_repr(self) -> str: return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \ f"window_size={self.window_size}, mlp_ratio={self.mlp_ratio}" class PatchMerging(nn.Module): """ Patch Merging Layer. Args: input_resolution (tuple[int]): Resolution of input feature. dim (int): Number of input channels. norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm """ def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm): super().__init__() self.input_resolution = input_resolution self.dim = dim self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False) self.norm = norm_layer(4 * dim) def forward(self, x): """ x: B, H*W, C """ H, W = self.input_resolution B, L, C = x.shape assert L == H * W, "input feature has wrong size" assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even." x = x.view(B, H, W, C) x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C x = self.norm(x) x = self.reduction(x) return x def extra_repr(self) -> str: return f"input_resolution={self.input_resolution}, dim={self.dim}" class BasicLayer(nn.Module): """ A basic Hierarchical Transformer layer for one stage. Args: dim (int): Number of input channels. input_resolution (tuple[int]): Input resolution. depth (int): Number of blocks. num_heads (int): Number of heads for spatial self-correlation. base_win_size (tuple[int]): The height and width of the base window. mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. drop (float, optional): Dropout rate. Default: 0.0 value_drop (float, optional): Dropout ratio of value. Default: 0.0 drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. hier_win_ratios (list): hierarchical window ratios for a transformer block. Default: [0.5,1,2,4,6,8]. """ def __init__(self, dim, input_resolution, depth, num_heads, base_win_size, mlp_ratio=4., drop=0., value_drop=0.,drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False, hier_win_ratios=[0.5,1,2,4,6,8]): super().__init__() self.dim = dim self.input_resolution = input_resolution self.depth = depth self.use_checkpoint = use_checkpoint self.win_hs = [int(base_win_size[0] * ratio) for ratio in hier_win_ratios] self.win_ws = [int(base_win_size[1] * ratio) for ratio in hier_win_ratios] # build blocks self.blocks = nn.ModuleList([ HierarchicalTransformerBlock(dim=dim, input_resolution=input_resolution, num_heads=num_heads, base_win_size=base_win_size, window_size=(self.win_hs[i], self.win_ws[i]), mlp_ratio=mlp_ratio, drop=drop, value_drop=value_drop, drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path, norm_layer=norm_layer) for i in range(depth)]) # patch merging layer if downsample is not None: self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer) else: self.downsample = None def forward(self, x, x_size): i = 0 for blk in self.blocks: if self.use_checkpoint: x = checkpoint.checkpoint(blk, x, x_size, (self.win_hs[i], self.win_ws[i])) else: x = blk(x, x_size, (self.win_hs[i], self.win_ws[i])) i = i + 1 if self.downsample is not None: x = self.downsample(x) return x def extra_repr(self) -> str: return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}" class RHTB(nn.Module): """Residual Hierarchical Transformer Block (RHTB). Args: dim (int): Number of input channels. input_resolution (tuple[int]): Input resolution. depth (int): Number of blocks. num_heads (int): Number of heads for spatial self-correlation. base_win_size (tuple[int]): The height and width of the base window. mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. drop (float, optional): Dropout rate. Default: 0.0 value_drop (float, optional): Dropout ratio of value. Default: 0.0 drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. img_size: Input image size. patch_size: Patch size. resi_connection: The convolutional block before residual connection. hier_win_ratios (list): hierarchical window ratios for a transformer block. Default: [0.5,1,2,4,6,8]. """ def __init__(self, dim, input_resolution, depth, num_heads, base_win_size, mlp_ratio=4., drop=0., value_drop=0., drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False, img_size=224, patch_size=4, resi_connection='1conv', hier_win_ratios=[0.5,1,2,4,6,8]): super(RHTB, self).__init__() self.dim = dim self.input_resolution = input_resolution self.residual_group = BasicLayer(dim=dim, input_resolution=input_resolution, depth=depth, num_heads=num_heads, base_win_size=base_win_size, mlp_ratio=mlp_ratio, drop=drop, value_drop=value_drop, drop_path=drop_path, norm_layer=norm_layer, downsample=downsample, use_checkpoint=use_checkpoint, hier_win_ratios=hier_win_ratios) if resi_connection == '1conv': self.conv = nn.Conv2d(dim, dim, 3, 1, 1) elif resi_connection == '3conv': # to save parameters and memory self.conv = nn.Sequential(nn.Conv2d(dim, dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Conv2d(dim // 4, dim // 4, 1, 1, 0), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Conv2d(dim // 4, dim, 3, 1, 1)) self.patch_embed = PatchEmbed( img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim, norm_layer=None) self.patch_unembed = PatchUnEmbed( img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim, norm_layer=None) def forward(self, x, x_size): return self.patch_embed(self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))) + x class PatchEmbed(nn.Module): r""" Image to Patch Embedding Args: img_size (int): Image size. Default: 224. patch_size (int): Patch token size. Default: 4. in_chans (int): Number of input image channels. Default: 3. embed_dim (int): Number of linear projection output channels. Default: 96. norm_layer (nn.Module, optional): Normalization layer. Default: None """ def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None): super().__init__() img_size = to_2tuple(img_size) patch_size = to_2tuple(patch_size) patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]] self.img_size = img_size self.patch_size = patch_size self.patches_resolution = patches_resolution self.num_patches = patches_resolution[0] * patches_resolution[1] self.in_chans = in_chans self.embed_dim = embed_dim if norm_layer is not None: self.norm = norm_layer(embed_dim) else: self.norm = None def forward(self, x): x = x.flatten(2).transpose(1, 2) # B Ph*Pw C if self.norm is not None: x = self.norm(x) return x class PatchUnEmbed(nn.Module): r""" Image to Patch Unembedding Args: img_size (int): Image size. Default: 224. patch_size (int): Patch token size. Default: 4. in_chans (int): Number of input image channels. Default: 3. embed_dim (int): Number of linear projection output channels. Default: 96. norm_layer (nn.Module, optional): Normalization layer. Default: None """ def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None): super().__init__() img_size = to_2tuple(img_size) patch_size = to_2tuple(patch_size) patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]] self.img_size = img_size self.patch_size = patch_size self.patches_resolution = patches_resolution self.num_patches = patches_resolution[0] * patches_resolution[1] self.in_chans = in_chans self.embed_dim = embed_dim def forward(self, x, x_size): B, HW, C = x.shape x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1]) # B Ph*Pw C return x class Upsample(nn.Sequential): """Upsample module. Args: scale (int): Scale factor. Supported scales: 2^n and 3. num_feat (int): Channel number of intermediate features. """ def __init__(self, scale, num_feat): m = [] if (scale & (scale - 1)) == 0: # scale = 2^n for _ in range(int(math.log(scale, 2))): m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1)) m.append(nn.PixelShuffle(2)) elif scale == 3: m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1)) m.append(nn.PixelShuffle(3)) else: raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.') super(Upsample, self).__init__(*m) class UpsampleOneStep(nn.Sequential): """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle) Used in lightweight SR to save parameters. Args: scale (int): Scale factor. Supported scales: 2^n and 3. num_feat (int): Channel number of intermediate features. """ def __init__(self, scale, num_feat, num_out_ch, input_resolution=None): self.num_feat = num_feat self.input_resolution = input_resolution m = [] m.append(nn.Conv2d(num_feat, (scale ** 2) * num_out_ch, 3, 1, 1)) m.append(nn.PixelShuffle(scale)) super(UpsampleOneStep, self).__init__(*m) class HiT_SNG(nn.Module, PyTorchModelHubMixin): """ HiT-SNG network. Args: img_size (int | tuple(int)): Input image size. Default 64 patch_size (int | tuple(int)): Patch size. Default: 1 in_chans (int): Number of input image channels. Default: 3 embed_dim (int): Patch embedding dimension. Default: 96 depths (tuple(int)): Depth of each Transformer block. num_heads (tuple(int)): Number of heads for spatial self-correlation in different layers. base_win_size (tuple[int]): The height and width of the base window. mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4 drop_rate (float): Dropout rate. Default: 0 value_drop_rate (float): Dropout ratio of value. Default: 0.0 drop_path_rate (float): Stochastic depth rate. Default: 0.1 norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm. ape (bool): If True, add absolute position embedding to the patch embedding. Default: False patch_norm (bool): If True, add normalization after patch embedding. Default: True use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False upscale (int): Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction img_range (float): Image range. 1. or 255. upsampler (str): The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None resi_connection (str): The convolutional block before residual connection. '1conv'/'3conv' hier_win_ratios (list): hierarchical window ratios for a transformer block. Default: [0.5,1,2,4,6,8]. """ def __init__(self, img_size=64, patch_size=1, in_chans=3, embed_dim=60, depths=[6, 6, 6, 6], num_heads=[6, 6, 6, 6], base_win_size=[8,8], mlp_ratio=2., drop_rate=0., value_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, ape=False, patch_norm=True, use_checkpoint=False, upscale=4, img_range=1., upsampler='pixelshuffledirect', resi_connection='1conv', hier_win_ratios=[0.5,1,2,4,6,8], **kwargs): super(HiT_SNG, self).__init__() num_in_ch = in_chans num_out_ch = in_chans num_feat = 64 self.img_range = img_range if in_chans == 3: rgb_mean = (0.4488, 0.4371, 0.4040) self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1) else: self.mean = torch.zeros(1, 1, 1, 1) self.upscale = upscale self.upsampler = upsampler self.base_win_size = base_win_size ##################################################################################################### ################################### 1, shallow feature extraction ################################### self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1) ##################################################################################################### ################################### 2, deep feature extraction ###################################### self.num_layers = len(depths) self.embed_dim = embed_dim self.ape = ape self.patch_norm = patch_norm self.num_features = embed_dim self.mlp_ratio = mlp_ratio # split image into non-overlapping patches self.patch_embed = PatchEmbed( img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim, norm_layer=norm_layer if self.patch_norm else None) num_patches = self.patch_embed.num_patches patches_resolution = self.patch_embed.patches_resolution self.patches_resolution = patches_resolution # merge non-overlapping patches into image self.patch_unembed = PatchUnEmbed( img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim, norm_layer=norm_layer if self.patch_norm else None) # absolute position embedding if self.ape: self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim)) trunc_normal_(self.absolute_pos_embed, std=.02) self.pos_drop = nn.Dropout(p=drop_rate) # stochastic depth dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule # build Residual Hierarchical Transformer blocks (RHTB) self.layers = nn.ModuleList() for i_layer in range(self.num_layers): layer = RHTB(dim=embed_dim, input_resolution=(patches_resolution[0], patches_resolution[1]), depth=depths[i_layer], num_heads=num_heads[i_layer], base_win_size=base_win_size, mlp_ratio=self.mlp_ratio, drop=drop_rate, value_drop=value_drop_rate, drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])], # no impact on SR results norm_layer=norm_layer, downsample=None, use_checkpoint=use_checkpoint, img_size=img_size, patch_size=patch_size, resi_connection=resi_connection, hier_win_ratios=hier_win_ratios ) self.layers.append(layer) self.norm = norm_layer(self.num_features) # build the last conv layer in deep feature extraction if resi_connection == '1conv': self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1) elif resi_connection == '3conv': # to save parameters and memory self.conv_after_body = nn.Sequential(nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1)) ##################################################################################################### ################################ 3, high quality image reconstruction ################################ if self.upsampler == 'pixelshuffle': # for classical SR self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True)) self.upsample = Upsample(upscale, num_feat) self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) elif self.upsampler == 'pixelshuffledirect': # for lightweight SR (to save parameters) self.upsample = UpsampleOneStep(upscale, embed_dim, num_out_ch, (patches_resolution[0], patches_resolution[1])) elif self.upsampler == 'nearest+conv': # for real-world SR (less artifacts) assert self.upscale == 4, 'only support x4 now.' self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True)) self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1) self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1) self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1) self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True) else: # for image denoising and JPEG compression artifact reduction self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1) 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) @torch.jit.ignore def no_weight_decay(self): return {'absolute_pos_embed'} @torch.jit.ignore def no_weight_decay_keywords(self): return {'relative_position_bias_table'} def forward_features(self, x): x_size = (x.shape[2], x.shape[3]) x = self.patch_embed(x) if self.ape: x = x + self.absolute_pos_embed x = self.pos_drop(x) for layer in self.layers: x = layer(x, x_size) x = self.norm(x) # B L C x = self.patch_unembed(x, x_size) return x def infer_image(self, image_path, cuda=True): io_backend_opt = {'type':'disk'} self.file_client = FileClient(io_backend_opt.pop('type'), **io_backend_opt) # load lq image lq_path = image_path img_bytes = self.file_client.get(lq_path, 'lq') img_lq = imfrombytes(img_bytes, float32=True) # BGR to RGB, HWC to CHW, numpy to tensor x = img2tensor(img_lq, bgr2rgb=True, float32=True)[None,...] if cuda: x= x.cuda() out = self(x) if cuda: out = out.cpu() out = tensor2img(out) return out def forward(self, x): H, W = x.shape[2:] self.mean = self.mean.type_as(x) x = (x - self.mean) * self.img_range if self.upsampler == 'pixelshuffle': # for classical SR x = self.conv_first(x) x = self.conv_after_body(self.forward_features(x)) + x x = self.conv_before_upsample(x) x = self.conv_last(self.upsample(x)) elif self.upsampler == 'pixelshuffledirect': # for lightweight SR x = self.conv_first(x) x = self.conv_after_body(self.forward_features(x)) + x x = self.upsample(x) elif self.upsampler == 'nearest+conv': # for real-world SR x = self.conv_first(x) x = self.conv_after_body(self.forward_features(x)) + x x = self.conv_before_upsample(x) x = self.lrelu(self.conv_up1(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest'))) x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest'))) x = self.conv_last(self.lrelu(self.conv_hr(x))) else: # for image denoising and JPEG compression artifact reduction x_first = self.conv_first(x) res = self.conv_after_body(self.forward_features(x_first)) + x_first x = x + self.conv_last(res) x = x / self.img_range + self.mean return x[:, :, :H*self.upscale, :W*self.upscale] if __name__ == '__main__': upscale = 4 base_win_size = [8, 8] height = (1024 // upscale // base_win_size[0] + 1) * base_win_size[0] width = (720 // upscale // base_win_size[1] + 1) * base_win_size[1] ## HiT-SIR model = HiT_SNG(upscale=4, img_size=(height, width), base_win_size=base_win_size, img_range=1., depths=[6, 6, 6, 6], embed_dim=60, num_heads=[6, 6, 6, 6], mlp_ratio=2, upsampler='pixelshuffledirect') params_num = sum(p.numel() for p in model.parameters() if p.requires_grad) print("params: ", params_num)