CrossViT_base.py

import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.hub
from functools import partial
import torch.nn.functional as F
from itertools import repeat
import collections.abc

__all__ = ["CrossViT_Base"]

def drop_path(x, drop_prob: float = 0., training: bool = False):
    if drop_prob == 0. or not training:
        return x
    keep_prob = 1 - drop_prob
    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
    random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
    random_tensor.floor_()  # binarize
    output = x.div(keep_prob) * random_tensor
    return output


class DropPath(nn.Module):
    def __init__(self, drop_prob=None):
        super(DropPath, self).__init__()
        self.drop_prob = drop_prob

    def forward(self, x):
        return drop_path(x, self.drop_prob, self.training)



class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU,
                 norm_layer=None, bias=True, drop=0., use_conv=False):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        bias = to_2tuple(bias)
        drop_probs = to_2tuple(drop)
        linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear

        self.fc1 = linear_layer(in_features, hidden_features, bias=bias[0])
        self.act = act_layer()
        self.drop1 = nn.Dropout(drop_probs[0])
        self.norm = norm_layer(hidden_features) if norm_layer is not None else nn.Identity()
        self.fc2 = linear_layer(hidden_features, out_features, bias=bias[1])
        self.drop2 = nn.Dropout(drop_probs[1])

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop1(x)
        x = self.fc2(x)
        x = self.drop2(x)
        return x    
    

class Attention(nn.Module):
    def __init__(self, dim, num_heads=8, qkv_bias=False, attn_drop=0., proj_drop=0.):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = 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]   # make torchscript happy (cannot use tensor as tuple)

        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 Block(nn.Module):

    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, drop=0., attn_drop=0.,
                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop)
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        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 forward(self, x):
        x = x + self.drop_path(self.attn(self.norm1(x)))
        x = x + self.drop_path(self.mlp(self.norm2(x)))
        return x
    

def _ntuple(n):
    def parse(x):
        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
            return tuple(x)
        return tuple(repeat(x, n))
    return parse

to_2tuple = _ntuple(2)

def _trunc_normal_(tensor, mean, std, a, b):
    def norm_cdf(x):
        return (1. + math.erf(x / math.sqrt(2.))) / 2.

    if (mean < a - 2 * std) or (mean > b + 2 * std):
        warnings.warn("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
                      "The distribution of values may be incorrect.",
                      stacklevel=2)
    l = norm_cdf((a - mean) / std)
    u = norm_cdf((b - mean) / std)

    tensor.uniform_(2 * l - 1, 2 * u - 1)
    tensor.erfinv_()
    tensor.mul_(std * math.sqrt(2.))
    tensor.add_(mean)
    tensor.clamp_(min=a, max=b)
    return tensor


def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
    with torch.no_grad():
        return _trunc_normal_(tensor, mean, std, a, b)


class PatchEmbed(nn.Module):
    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, multi_conv=False):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
        self.img_size = img_size
        self.patch_size = patch_size
        self.num_patches = num_patches
        if multi_conv:
            if patch_size[0] == 12:
                self.proj = nn.Sequential(
                    nn.Conv2d(in_chans, embed_dim // 4, kernel_size=7, stride=4, padding=3),
                    nn.ReLU(inplace=True),
                    nn.Conv2d(embed_dim // 4, embed_dim // 2, kernel_size=3, stride=3, padding=0),
                    nn.ReLU(inplace=True),
                    nn.Conv2d(embed_dim // 2, embed_dim, kernel_size=3, stride=1, padding=1),
                )
            elif patch_size[0] == 16:
                self.proj = nn.Sequential(
                    nn.Conv2d(in_chans, embed_dim // 4, kernel_size=7, stride=4, padding=3),
                    nn.ReLU(inplace=True),
                    nn.Conv2d(embed_dim // 4, embed_dim // 2, kernel_size=3, stride=2, padding=1),
                    nn.ReLU(inplace=True),
                    nn.Conv2d(embed_dim // 2, embed_dim, kernel_size=3, stride=2, padding=1),
                )
        else:
            self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)

    def forward(self, x):
        B, C, H, W = x.shape
        # FIXME look at relaxing size constraints
        assert H == self.img_size[0] and W == self.img_size[1], \
            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
        x = self.proj(x).flatten(2).transpose(1, 2)
        return x


class CrossAttention(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.wq = nn.Linear(dim, dim, bias=qkv_bias)
        self.wk = nn.Linear(dim, dim, bias=qkv_bias)
        self.wv = nn.Linear(dim, dim, 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
        q = self.wq(x[:, 0:1, ...]).reshape(B, 1, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)  # B1C -> B1H(C/H) -> BH1(C/H)
        k = self.wk(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)  # BNC -> BNH(C/H) -> BHN(C/H)
        v = self.wv(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)  # BNC -> BNH(C/H) -> BHN(C/H)

        attn = (q @ k.transpose(-2, -1)) * self.scale  # BH1(C/H) @ BH(C/H)N -> BH1N
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B, 1, C)   # (BH1N @ BHN(C/H)) -> BH1(C/H) -> B1H(C/H) -> B1C
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class CrossAttentionBlock(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, has_mlp=True):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = CrossAttention(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.has_mlp = has_mlp
        if has_mlp:
            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 forward(self, x):
        x = x[:, 0:1, ...] + self.drop_path(self.attn(self.norm1(x)))
        if self.has_mlp:
            x = x + self.drop_path(self.mlp(self.norm2(x)))

        return x


class MultiScaleBlock(nn.Module):
    def __init__(self, dim, patches, depth, num_heads, mlp_ratio, qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
        super().__init__()

        num_branches = len(dim)
        self.num_branches = num_branches
        self.blocks = nn.ModuleList()
        for d in range(num_branches):
            tmp = []
            for i in range(depth[d]):
                tmp.append(
                    Block(dim=dim[d], num_heads=num_heads[d], mlp_ratio=mlp_ratio[d], qkv_bias=qkv_bias, 
                          drop=drop, attn_drop=attn_drop, drop_path=drop_path[i], norm_layer=norm_layer))
            if len(tmp) != 0:
                self.blocks.append(nn.Sequential(*tmp))

        if len(self.blocks) == 0:
            self.blocks = None

        self.projs = nn.ModuleList()
        for d in range(num_branches):
            if dim[d] == dim[(d+1) % num_branches] and False:
                tmp = [nn.Identity()]
            else:
                tmp = [norm_layer(dim[d]), act_layer(), nn.Linear(dim[d], dim[(d+1) % num_branches])]
            self.projs.append(nn.Sequential(*tmp))

        self.fusion = nn.ModuleList()
        for d in range(num_branches):
            d_ = (d+1) % num_branches
            nh = num_heads[d_]
            if depth[-1] == 0:  # backward capability:
                self.fusion.append(CrossAttentionBlock(dim=dim[d_], num_heads=nh, mlp_ratio=mlp_ratio[d], 
                                                       qkv_bias=qkv_bias, qk_scale=qk_scale,
                                                       drop=drop, attn_drop=attn_drop, drop_path=drop_path[-1], norm_layer=norm_layer,
                                                       has_mlp=False))
            else:
                tmp = []
                for _ in range(depth[-1]):
                    tmp.append(CrossAttentionBlock(dim=dim[d_], num_heads=nh, mlp_ratio=mlp_ratio[d], qkv_bias=qkv_bias, qk_scale=qk_scale,
                                                   drop=drop, attn_drop=attn_drop, drop_path=drop_path[-1], norm_layer=norm_layer,
                                                   has_mlp=False))
                self.fusion.append(nn.Sequential(*tmp))

        self.revert_projs = nn.ModuleList()
        for d in range(num_branches):
            if dim[(d+1) % num_branches] == dim[d] and False:
                tmp = [nn.Identity()]
            else:
                tmp = [norm_layer(dim[(d+1) % num_branches]), act_layer(), nn.Linear(dim[(d+1) % num_branches], dim[d])]
            self.revert_projs.append(nn.Sequential(*tmp))

    def forward(self, x):
        outs_b = [block(x_) for x_, block in zip(x, self.blocks)]
        # only take the cls token out
        proj_cls_token = [proj(x[:, 0:1]) for x, proj in zip(outs_b, self.projs)]
        # cross attention
        outs = []
        for i in range(self.num_branches):
            tmp = torch.cat((proj_cls_token[i], outs_b[(i + 1) % self.num_branches][:, 1:, ...]), dim=1)
            tmp = self.fusion[i](tmp)
            reverted_proj_cls_token = self.revert_projs[i](tmp[:, 0:1, ...])
            tmp = torch.cat((reverted_proj_cls_token, outs_b[i][:, 1:, ...]), dim=1)
            outs.append(tmp)
        return outs


def _compute_num_patches(img_size, patches):
    return [i // p * i // p for i, p in zip(img_size,patches)]



class CrossViT_Base(nn.Module):
    def __init__(self, img_size=[240, 224], 
                 patch_size=[12, 16], 
                 in_chans=3, num_classes=3, 
                 embed_dim=[384, 768], 
                 depth=[[1, 4, 0], [1, 4, 0], [1, 4, 0]],
                 num_heads=[12, 12],
                 mlp_ratio=[4, 4, 1],
                 qkv_bias=True,
                 qk_scale=None,
                 drop_rate=0.,
                 attn_drop_rate=0.,
                 drop_path_rate=0.,
                 hybrid_backbone=None,
                 norm_layer=partial(nn.LayerNorm, eps=1e-6),
                 multi_conv=False):
        super().__init__()

        self.num_classes = num_classes
        if not isinstance(img_size, list):
            img_size = to_2tuple(img_size)
        self.img_size = img_size

        num_patches = _compute_num_patches(img_size, patch_size)
        self.num_branches = len(patch_size)

        self.patch_embed = nn.ModuleList()
        if hybrid_backbone is None:
            self.pos_embed = nn.ParameterList([nn.Parameter(torch.zeros(1, 1 + num_patches[i], 
                                                                        embed_dim[i])) for i in range(self.num_branches)])
            for im_s, p, d in zip(img_size, patch_size, embed_dim):
                self.patch_embed.append(PatchEmbed(img_size=im_s, patch_size=p, in_chans=in_chans, embed_dim=d, multi_conv=multi_conv))
        else:
            self.pos_embed = nn.ParameterList()
            from .t2t import T2T, get_sinusoid_encoding
            tokens_type = 'transformer' if hybrid_backbone == 't2t' else 'performer'
            for idx, (im_s, p, d) in enumerate(zip(img_size, patch_size, embed_dim)):
                self.patch_embed.append(T2T(im_s, tokens_type=tokens_type, patch_size=p, embed_dim=d))
                self.pos_embed.append(nn.Parameter(data=get_sinusoid_encoding(n_position=1 + num_patches[idx], 
                                                                              d_hid=embed_dim[idx]), requires_grad=False))

            del self.pos_embed
            self.pos_embed = nn.ParameterList([nn.Parameter(torch.zeros(1, 1 + num_patches[i], 
                                                                        embed_dim[i])) for i in range(self.num_branches)])

        self.cls_token = nn.ParameterList([nn.Parameter(torch.zeros(1, 1, embed_dim[i])) for i in range(self.num_branches)])
        self.pos_drop = nn.Dropout(p=drop_rate)

        total_depth = sum([sum(x[-2:]) for x in depth])
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, total_depth)]  # stochastic depth decay rule
        dpr_ptr = 0
        self.blocks = nn.ModuleList()
        for idx, block_cfg in enumerate(depth):
            curr_depth = max(block_cfg[:-1]) + block_cfg[-1]
            dpr_ = dpr[dpr_ptr:dpr_ptr + curr_depth]
            blk = MultiScaleBlock(embed_dim, num_patches, block_cfg, num_heads=num_heads, mlp_ratio=mlp_ratio,
                                  qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr_,
                                  norm_layer=norm_layer)
            dpr_ptr += curr_depth
            self.blocks.append(blk)

        self.norm = nn.ModuleList([norm_layer(embed_dim[i]) for i in range(self.num_branches)])
        self.head = nn.ModuleList([nn.Linear(embed_dim[i], 
                                             num_classes) if num_classes > 0 else nn.Identity() for i in range(self.num_branches)])

        for i in range(self.num_branches):
            if self.pos_embed[i].requires_grad:
                trunc_normal_(self.pos_embed[i], std=.02)
            trunc_normal_(self.cls_token[i], std=.02)

        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)

    def forward_features(self, x):
        B, C, H, W = x.shape
        xs = []
        for i in range(self.num_branches):
            x_ = torch.nn.functional.interpolate(x, size=(self.img_size[i], self.img_size[i]), 
                                                 mode='bicubic') if H != self.img_size[i] else x
            tmp = self.patch_embed[i](x_)
            cls_tokens = self.cls_token[i].expand(B, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
            tmp = torch.cat((cls_tokens, tmp), dim=1)
            tmp = tmp + self.pos_embed[i]
            tmp = self.pos_drop(tmp)
            xs.append(tmp)

        for blk in self.blocks:
            xs = blk(xs)

        # NOTE: was before branch token section, move to here to assure all branch token are before layer norm
        xs = [self.norm[i](x) for i, x in enumerate(xs)]
        out = [x[:, 0] for x in xs]

        return out

    def forward(self, x):
        xs = self.forward_features(x)
        ce_logits = [self.head[i](x) for i, x in enumerate(xs)]
        ce_logits = torch.mean(torch.stack(ce_logits, dim=0), dim=0)
        return ce_logits



if __name__ == "__main__":
    model = CrossViT_Base()
    input = torch.randn(1,3,224,224)
    output = model(input)
    print("Model done")
    print(input.size())
    print(output.size())
    assert output.size()[-1] == 3
    print("Model done again")