import torch import torch.nn as nn from spikingjelly.clock_driven.neuron import MultiStepParametricLIFNode, MultiStepLIFNode from timm.models.layers import to_2tuple, trunc_normal_, DropPath from timm.models.registry import register_model from timm.models.vision_transformer import _cfg from functools import partial from timm.models import create_model from .delay_synaptic_func_inter import DelayConv __all__ = ['delay_QKFormer'] class MLP(nn.Module): def __init__(self, in_features, hidden_features=None, out_features=None, drop=0.): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features self.mlp1_conv = nn.Conv2d(in_features, hidden_features, kernel_size=1, stride=1) self.mlp1_bn = nn.BatchNorm2d(hidden_features) self.mlp1_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') self.mlp2_conv = nn.Conv2d(hidden_features, out_features, kernel_size=1, stride=1) self.mlp2_bn = nn.BatchNorm2d(out_features) self.mlp2_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') self.c_hidden = hidden_features self.c_output = out_features def forward(self, x): T, B, C, H, W = x.shape x = self.mlp1_conv(x.flatten(0, 1)) x = self.mlp1_bn(x).reshape(T, B, self.c_hidden, H, W) x = self.mlp1_lif(x) x = self.mlp2_conv(x.flatten(0, 1)) x = self.mlp2_bn(x).reshape(T, B, C, H, W) x = self.mlp2_lif(x) return x class Token_QK_Attention(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1, k=16): 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 self.q_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1, bias=False) self.q_bn = nn.BatchNorm1d(dim) self.q_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') # self.k_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1, bias=False) self.k_proj_delay = DelayConv(in_c=self.dim, k=k) self.k_bn = nn.BatchNorm1d(dim) self.k_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') self.attn_lif = MultiStepLIFNode(tau=2.0, v_threshold=0.5, detach_reset=True, backend='cupy') self.proj_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1) self.proj_bn = nn.BatchNorm1d(dim) self.proj_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') def forward(self, x): T, B, C, H, W = x.shape x = x.flatten(3) T, B, C, N = x.shape x_for_qkv = x.flatten(0, 1) q_conv_out = self.q_conv(x_for_qkv) q_conv_out = self.q_bn(q_conv_out).reshape(T, B, C, N) q_conv_out = self.q_lif(q_conv_out) q = q_conv_out.unsqueeze(2).reshape(T, B, self.num_heads, C // self.num_heads, N) # k_conv_out = self.k_conv(x_for_qkv) k_conv_out = self.k_proj_delay(x_for_qkv.reshape(T,B,C,N)) k_conv_out = self.k_bn(k_conv_out).reshape(T, B, C, N) k_conv_out = self.k_lif(k_conv_out) k = k_conv_out.unsqueeze(2).reshape(T, B, self.num_heads, C // self.num_heads, N) q = torch.sum(q, dim=3, keepdim=True) attn = self.attn_lif(q) x = torch.mul(attn, k) x = x.flatten(2, 3) x = self.proj_bn(self.proj_conv(x.flatten(0, 1))).reshape(T, B, C, H, W) x = self.proj_lif(x) return x class Spiking_Self_Attention(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1, k=16): 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 = 0.125 self.q_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1, bias=False) self.q_bn = nn.BatchNorm1d(dim) self.q_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') # self.k_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1, bias=False) self.k_proj_delay = DelayConv(in_c=self.dim, k=k) self.k_bn = nn.BatchNorm1d(dim) self.k_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') # self.v_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1, bias=False) self.v_proj_delay = DelayConv(in_c=self.dim, k=k) self.v_bn = nn.BatchNorm1d(dim) self.v_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') self.attn_lif = MultiStepLIFNode(tau=2.0, v_threshold=0.5, detach_reset=True, backend='cupy') self.proj_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1) self.proj_bn = nn.BatchNorm1d(dim) self.proj_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') self.qkv_mp = nn.MaxPool1d(4) def forward(self, x): T, B, C, H, W = x.shape x = x.flatten(3) T, B, C, N = x.shape x_for_qkv = x.flatten(0, 1) q_conv_out = self.q_conv(x_for_qkv) q_conv_out = self.q_bn(q_conv_out).reshape(T, B, C, N).contiguous() q_conv_out = self.q_lif(q_conv_out) q = q_conv_out.transpose(-1, -2).reshape(T, B, N, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4).contiguous() k_conv_out = self.k_proj_delay(x_for_qkv.reshape(T,B,C,N)) k_conv_out = self.k_bn(k_conv_out).reshape(T, B, C, N).contiguous() k_conv_out = self.k_lif(k_conv_out) k = k_conv_out.transpose(-1, -2).reshape(T, B, N, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4).contiguous() v_conv_out = self.v_proj_delay(x_for_qkv.reshape(T,B,C,N)) v_conv_out = self.v_bn(v_conv_out).reshape(T, B, C, N).contiguous() v_conv_out = self.v_lif(v_conv_out) v = v_conv_out.transpose(-1, -2).reshape(T, B, N, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4).contiguous() x = k.transpose(-2, -1) @ v x = (q @ x) * self.scale x = x.transpose(3, 4).reshape(T, B, C, N).contiguous() x = self.attn_lif(x) x = x.flatten(0, 1) x = self.proj_lif(self.proj_bn(self.proj_conv(x))).reshape(T, B, C, H, W) return x class TokenSpikingTransformer(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., norm_layer=nn.LayerNorm, sr_ratio=1): super().__init__() self.tssa = Token_QK_Attention(dim, num_heads) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = MLP(in_features= dim, hidden_features=mlp_hidden_dim, drop=drop) def forward(self, x): x = x + self.tssa(x) x = x + self.mlp(x) return x class SpikingTransformer(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., norm_layer=nn.LayerNorm, sr_ratio=1): super().__init__() self.ssa = Spiking_Self_Attention(dim, num_heads) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = MLP(in_features= dim, hidden_features=mlp_hidden_dim, drop=drop) def forward(self, x): x = x + self.ssa(x) x = x + self.mlp(x) return x class PatchEmbedInit(nn.Module): def __init__(self, img_size_h=128, img_size_w=128, patch_size=4, in_channels=2, embed_dims=256): super().__init__() self.image_size = [img_size_h, img_size_w] patch_size = to_2tuple(patch_size) self.patch_size = patch_size self.C = in_channels self.H, self.W = self.image_size[0] // patch_size[0], self.image_size[1] // patch_size[1] self.num_patches = self.H * self.W self.proj_conv = nn.Conv2d(in_channels, embed_dims // 8, kernel_size=3, stride=1, padding=1, bias=False) self.proj_bn = nn.BatchNorm2d(embed_dims // 8) self.proj_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') self.proj1_conv = nn.Conv2d(embed_dims // 8, embed_dims // 4, kernel_size=3, stride=1, padding=1, bias=False) self.proj1_bn = nn.BatchNorm2d(embed_dims // 4) self.maxpool1 = torch.nn.MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) self.proj1_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') self.proj2_conv = nn.Conv2d(embed_dims//4, embed_dims // 2, kernel_size=3, stride=1, padding=1, bias=False) self.proj2_bn = nn.BatchNorm2d(embed_dims // 2) self.maxpool2 = torch.nn.MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) self.proj2_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') self.proj3_conv = nn.Conv2d(embed_dims // 2, embed_dims, kernel_size=3, stride=1, padding=1, bias=False) self.proj3_bn = nn.BatchNorm2d(embed_dims) self.maxpool3 = torch.nn.MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) self.proj3_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') self.proj_res_conv = nn.Conv2d(embed_dims // 4, embed_dims, kernel_size=1, stride=4, padding=0, bias=False) self.proj_res_bn = nn.BatchNorm2d(embed_dims) self.proj_res_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') def forward(self, x): T, B, C, H, W = x.shape # Downsampling + Res # x_feat = x.flatten(0, 1) x = self.proj_conv(x.flatten(0, 1)) x = self.proj_bn(x).reshape(T, B, -1, H, W) x = self.proj_lif(x).flatten(0, 1).contiguous() x = self.proj1_conv(x) x = self.proj1_bn(x) x = self.maxpool1(x) _, _, H1, W1 = x.shape x = x.reshape(T, B, -1, H1, W1).contiguous() x = self.proj1_lif(x).flatten(0, 1).contiguous() x_feat = x x = self.proj2_conv(x) x = self.proj2_bn(x) x = self.maxpool2(x) _, _, H2, W2 = x.shape x = x.reshape(T, B, -1, H2, W2).contiguous() x = self.proj2_lif(x).flatten(0, 1).contiguous() x = self.proj3_conv(x) x = self.proj3_bn(x) x = self.maxpool3(x) _, _, H3, W3 = x.shape x = x.reshape(T, B, -1, H3, W3).contiguous() x = self.proj3_lif(x) x_feat = self.proj_res_conv(x_feat) x_feat = self.proj_res_bn(x_feat) _, _, Hres, Wres = x_feat.shape x_feat = x_feat.reshape(T, B, -1, Hres, Wres).contiguous() x_feat = self.proj_res_lif(x_feat) x = x + x_feat # shortcut return x class PatchEmbeddingStage(nn.Module): def __init__(self, img_size_h=128, img_size_w=128, patch_size=4, in_channels=2, embed_dims=256): super().__init__() self.image_size = [img_size_h, img_size_w] patch_size = to_2tuple(patch_size) self.patch_size = patch_size self.C = in_channels self.H, self.W = self.image_size[0] // patch_size[0], self.image_size[1] // patch_size[1] self.num_patches = self.H * self.W self.proj_conv = nn.Conv2d(embed_dims//2, embed_dims, kernel_size=3, stride=1, padding=1, bias=False) self.proj_bn = nn.BatchNorm2d(embed_dims) self.proj_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') self.proj4_conv = nn.Conv2d(embed_dims, embed_dims, kernel_size=3, stride=1, padding=1, bias=False) self.proj4_bn = nn.BatchNorm2d(embed_dims) self.proj4_maxpool = torch.nn.MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) self.proj4_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') self.proj_res_conv = nn.Conv2d(embed_dims//2, embed_dims, kernel_size=1, stride=2, padding=0, bias=False) self.proj_res_bn = nn.BatchNorm2d(embed_dims) self.proj_res_lif = MultiStepLIFNode(tau=2.0, detach_reset=True, backend='cupy') def forward(self, x): T, B, C, H, W = x.shape # Downsampling + Res x = x.flatten(0, 1).contiguous() x_feat = x x = self.proj_conv(x) x = self.proj_bn(x).reshape(T, B, -1, H, W).contiguous() x = self.proj_lif(x).flatten(0, 1).contiguous() x = self.proj4_conv(x) x = self.proj4_bn(x) x = self.proj4_maxpool(x) _, _, H4, W4 = x.shape x = x.reshape(T, B, -1, H4, W4).contiguous() x = self.proj4_lif(x) x_feat = self.proj_res_conv(x_feat) x_feat = self.proj_res_bn(x_feat) _, _, Hres, Wres = x_feat.shape x_feat = x_feat.reshape(T, B, -1, Hres, Wres).contiguous() x_feat = self.proj_res_lif(x_feat) x = x + x_feat # shortcut return x class vit_snn(nn.Module): def __init__(self, img_size_h=128, img_size_w=128, patch_size=16, in_channels=2, num_classes=11, embed_dims=[64, 128, 256], num_heads=[1, 2, 4], mlp_ratios=[4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, depths=[6, 8, 6], sr_ratios=[8, 4, 2], T=4, pretrained_cfg=None, in_chans = 3, no_weight_decay = None ): super().__init__() self.num_classes = num_classes self.depths = depths self.T = T num_heads = [16, 16, 16] dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depths)] # stochastic depth decay rule # patch_embed1 = PatchEmbedInit(img_size_h=img_size_h, img_size_w=img_size_w, patch_size=patch_size, in_channels=in_channels, embed_dims=embed_dims // 2) stage1 = nn.ModuleList([TokenSpikingTransformer( dim=embed_dims // 2, num_heads=num_heads[0], mlp_ratio=mlp_ratios, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[j], norm_layer=norm_layer, sr_ratio=sr_ratios) for j in range(1)]) patch_embed2 = PatchEmbeddingStage(img_size_h=img_size_h, img_size_w=img_size_w, patch_size=patch_size, in_channels=in_channels, embed_dims=embed_dims) stage2 = nn.ModuleList([SpikingTransformer( dim=embed_dims, num_heads=num_heads[1], mlp_ratio=mlp_ratios, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[j], norm_layer=norm_layer, sr_ratio=sr_ratios) for j in range(1)]) setattr(self, f"patch_embed1", patch_embed1) setattr(self, f"stage1", stage1) setattr(self, f"patch_embed2", patch_embed2) setattr(self, f"stage2", stage2) # classification head self.head = nn.Linear(embed_dims, num_classes) if num_classes > 0 else nn.Identity() self.apply(self._init_weights) @torch.jit.ignore def no_weight_decay(self): return {'pose_embed'} @torch.jit.ignore def _get_pos_embed(self, pos_embed, patch_embed, H, W): return None 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) def forward_features(self, x): stage1 = getattr(self, f"stage1") patch_embed1 = getattr(self, f"patch_embed1") stage2 = getattr(self, f"stage2") patch_embed2 = getattr(self, f"patch_embed2") x = patch_embed1(x) for blk in stage1: x = blk(x) x = patch_embed2(x) for blk in stage2: x = blk(x) return x.flatten(3).mean(3) def forward(self, x): x = x.permute(1, 0, 2, 3, 4) # [T, N, 2, *, *] # print("torch.unique", torch.unique(x)) # print("torch.count_nonzero", torch.count_nonzero(x)) # print("numel()", x.numel()) x = self.forward_features(x) x = self.head(x.mean(0)) return x @register_model def delay_QKFormer(pretrained=False, **kwargs): model = vit_snn( patch_size=16, embed_dims=256, num_heads=16, mlp_ratios=4, in_channels=2, num_classes=101, qkv_bias=False, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=4, sr_ratios=1, **kwargs ) model.default_cfg = _cfg() return model from timm.models import create_model if __name__ == '__main__': x = torch.randn(1, 1, 2, 128, 128).cuda() model = create_model( 'delay_QKFormer', pretrained=False, drop_rate=0, drop_path_rate=0.1, drop_block_rate=None, ).cuda() model.eval() from torchinfo import summary summary(model, input_size=(1, 1, 2, 128, 128)) # y = model(x) # print(y.shape) # print('Test Good!')