p2t.py

from os import sep
from pickle import TRUE
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 numpy as np

__all__ = [
    'p2t_tiny', 'p2t_small', 'p2t_base', 'p2t_large'
]


class IRB(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, ksize=3, act_layer=nn.Hardswish, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Conv2d(in_features, hidden_features, 1, 1, 0)
        self.act = act_layer()
        self.conv = nn.Conv2d(hidden_features, hidden_features, kernel_size=ksize, padding=ksize//2, stride=1, groups=hidden_features)
        self.fc2 = nn.Conv2d(hidden_features, out_features, 1, 1, 0)
        self.drop = nn.Dropout(drop)
    
    def forward(self, x, H, W):
        B, N, C = x.shape
        x = x.permute(0,2,1).reshape(B, C, H, W)
        x = self.fc1(x)
        x = self.act(x)
        x = self.conv(x)
        x = self.act(x)
        x = self.fc2(x)
        return x.reshape(B, C, -1).permute(0,2,1)


class PoolingAttention(nn.Module):
    def __init__(self, dim, num_heads=2, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., 
        pool_ratios=[1,2,3,6]):

        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.num_elements = np.array([t*t for t in pool_ratios]).sum()
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5

        self.q = nn.Sequential(nn.Linear(dim, dim, bias=qkv_bias))
        self.kv = nn.Sequential(nn.Linear(dim, dim * 2, bias=qkv_bias))
        
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

        self.pool_ratios = pool_ratios
        self.pools = nn.ModuleList()
        
        self.norm = nn.LayerNorm(dim)

    def forward(self, x, H, W, d_convs=None):
        B, N, C = x.shape
        
        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
        pools = []
        x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
        for (pool_ratio, l) in zip(self.pool_ratios, d_convs):
            pool = F.adaptive_avg_pool2d(x_, (round(H/pool_ratio), round(W/pool_ratio)))
            pool = pool + l(pool) # fix backward bug in higher torch versions when training
            pools.append(pool.view(B, C, -1))
        
        pools = torch.cat(pools, dim=2)
        pools = self.norm(pools.permute(0,2,1))
        
        kv = self.kv(pools).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)   
        x = x.transpose(1,2).contiguous().reshape(B, N, C)
        
        x = self.proj(x)

        return x


class Block(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, pool_ratios=[12,16,20,24]):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = PoolingAttention(
                    dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, 
                    attn_drop=attn_drop, proj_drop=drop, pool_ratios=pool_ratios)
        
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()

        self.norm2 = norm_layer(dim)
        self.mlp = IRB(in_features=dim, hidden_features=int(dim * mlp_ratio), act_layer=nn.Hardswish, drop=drop, ksize=3)
    
    def forward(self, x, H, W, d_convs=None):
        x = x + self.drop_path(self.attn(self.norm1(x), H, W, d_convs=d_convs))
        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))

        return x

class PatchEmbed(nn.Module):
    """ (Overlapped) Image to Patch Embedding
    """

    def __init__(self, img_size=224, patch_size=16, kernel_size=3, in_chans=3, embed_dim=768, overlap=True):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)

        self.img_size = img_size
        self.patch_size = patch_size
        assert img_size[0] % patch_size[0] == 0 and img_size[1] % patch_size[1] == 0, \
            f"img_size {img_size} should be divided by patch_size {patch_size}."
        self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
        self.num_patches = self.H * self.W
        if not overlap:
            self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
        else:
            self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=kernel_size, stride=patch_size, padding=kernel_size//2)
        
        self.norm = nn.LayerNorm(embed_dim)

    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 PyramidPoolingTransformer(nn.Module):
    def __init__(self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, embed_dims=[64, 128, 320, 512],
                 num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True, qk_scale=None, drop_rate=0.,
                 attn_drop_rate=0., drop_path_rate=0.1, norm_layer=partial(nn.LayerNorm, eps=1e-6),
                 depths=[2, 2, 9, 3], **kwargs): #
        super().__init__()
        self.num_classes = num_classes
        self.depths = depths

        self.embed_dims = embed_dims

        # pyramid pooling ratios for each stage
        pool_ratios = [[12,16,20,24], [6,8,10,12], [3,4,5,6], [1,2,3,4]]
        
        self.patch_embed1 = PatchEmbed(img_size=img_size, patch_size=4, kernel_size=7, in_chans=in_chans,
                                       embed_dim=embed_dims[0], overlap=True)

        self.patch_embed2 = PatchEmbed(img_size=img_size // 4, patch_size=2, in_chans=embed_dims[0],
                                    embed_dim=embed_dims[1], overlap=True)
        self.patch_embed3 = PatchEmbed(img_size=img_size // 8, patch_size=2, in_chans=embed_dims[1],
                                    embed_dim=embed_dims[2], overlap=True)
        self.patch_embed4 = PatchEmbed(img_size=img_size // 16, patch_size=2, in_chans=embed_dims[2],
                                    embed_dim=embed_dims[3], overlap=True)
        
        self.d_convs1 = nn.ModuleList([nn.Conv2d(embed_dims[0], embed_dims[0], kernel_size=3, stride=1, padding=1, groups=embed_dims[0]) for temp in pool_ratios[0]])
        self.d_convs2 = nn.ModuleList([nn.Conv2d(embed_dims[1], embed_dims[1], kernel_size=3, stride=1, padding=1, groups=embed_dims[1]) for temp in pool_ratios[1]])
        self.d_convs3 = nn.ModuleList([nn.Conv2d(embed_dims[2], embed_dims[2], kernel_size=3, stride=1, padding=1, groups=embed_dims[2]) for temp in pool_ratios[2]])
        self.d_convs4 = nn.ModuleList([nn.Conv2d(embed_dims[3], embed_dims[3], kernel_size=3, stride=1, padding=1, groups=embed_dims[3]) for temp in pool_ratios[3]])

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


        ksize = 3

        self.block1 = nn.ModuleList([Block(
            dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale,
            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, pool_ratios=pool_ratios[0])
            for i in range(depths[0])])
        

        cur += depths[0]
        self.block2 = nn.ModuleList([Block(
            dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale,
            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, pool_ratios=pool_ratios[1])
            for i in range(depths[1])])

        cur += depths[1]

        
        self.block3 = nn.ModuleList([Block(
            dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale,
            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, pool_ratios=pool_ratios[2])
            for i in range(depths[2])])

        cur += depths[2]

        self.block4 = nn.ModuleList([Block(
            dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, pool_ratios=pool_ratios[3])
            for i in range(depths[3])])
        
        # classification head
        self.head = nn.Linear(embed_dims[3], num_classes) if num_classes > 0 else nn.Identity()
        self.gap = nn.AdaptiveAvgPool1d(1)

        self.apply(self._init_weights)

        #print(self)

    def reset_drop_path(self, drop_path_rate):
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))]
        cur = 0
        for i in range(self.depths[0]):
            self.block1[i].drop_path.drop_prob = dpr[cur + i]

        cur += self.depths[0]
        for i in range(self.depths[1]):
            self.block2[i].drop_path.drop_prob = dpr[cur + i]

        cur += self.depths[1]
        for i in range(self.depths[2]):
            self.block3[i].drop_path.drop_prob = dpr[cur + i]

        cur += self.depths[2]
        for i in range(self.depths[3]):
            self.block4[i].drop_path.drop_prob = dpr[cur + i]
        

    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 {'pos_embed', 'cls_token'} # has pos_embed may be better
        return {'cls_token'}

    def get_classifier(self):
        return self.head

    def reset_classifier(self, num_classes, global_pool=''):
        self.num_classes = num_classes
        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
    
    def forward_features(self, x):
        B = x.shape[0]

        # stage 1
        x, (H, W) = self.patch_embed1(x)
        
        for idx, blk in enumerate(self.block1):
            x = blk(x, H, W, self.d_convs1)
        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2)

        # stage 2
        x, (H, W) = self.patch_embed2(x)

        for idx, blk in enumerate(self.block2):
            x = blk(x, H, W, self.d_convs2)
        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2)

        # stage 3
        x, (H, W) = self.patch_embed3(x)

        for idx, blk in enumerate(self.block3):
            x = blk(x, H, W, self.d_convs3)
        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2)
        
        # stage 4
        x, (H, W) = self.patch_embed4(x)

        for idx, blk in enumerate(self.block4):
            x = blk(x, H, W, self.d_convs4)
        
        return x
    
    def forward_features_for_fpn(self, x):
        outs = []

        B = x.shape[0]

        # stage 1
        x, (H, W) = self.patch_embed1(x)
        
        for idx, blk in enumerate(self.block1):
            x = blk(x, H, W, self.d_convs1)
        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2)
        outs.append(x)

        # stage 2
        x, (H, W) = self.patch_embed2(x)

        for idx, blk in enumerate(self.block2):
            x = blk(x, H, W, self.d_convs2)
        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2)
        outs.append(x)

        x, (H, W) = self.patch_embed3(x)

        for idx, blk in enumerate(self.block3):
            x = blk(x, H, W, self.d_convs3)
        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2)
        outs.append(x)
        
        # stage 4
        x, (H, W) = self.patch_embed4(x)

        for idx, blk in enumerate(self.block4):
            x = blk(x, H, W, self.d_convs4)
        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2)
        outs.append(x)
        
        return outs

    def forward(self, x):
        x = self.forward_features(x)
        x = torch.mean(x, dim=1)
        x = self.head(x)

        return x
    
    def forward_for_fpn(self, x):
        return self.forward_features_for_fpn(x)


def _conv_filter(state_dict, patch_size=16):
    """ convert patch embedding weight from manual patchify + linear proj to conv"""
    out_dict = {}
    for k, v in state_dict.items():
        if 'patch_embed.proj.weight' in k:
            v = v.reshape((v.shape[0], 3, patch_size, patch_size))
        out_dict[k] = v

    return out_dict


@register_model
def p2t_tiny(pretrained=False, **kwargs):
    model = PyramidPoolingTransformer(
        patch_size=4, embed_dims=[48, 96, 240, 384], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 6, 3],
        **kwargs)
    model.default_cfg = _cfg()

    return model

@register_model
def p2t_small(pretrained=True, **kwargs):
    model = PyramidPoolingTransformer(
        patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 9, 3], **kwargs)
    model.default_cfg = _cfg()

    return model

@register_model
def p2t_base(pretrained=False, **kwargs):
    model = PyramidPoolingTransformer(
        patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3],
        **kwargs)
    model.default_cfg = _cfg()

    return model

@register_model
def p2t_medium(pretrained=False, **kwargs):
    model = PyramidPoolingTransformer(
        patch_size=4, embed_dims=[64, 128, 384, 512], num_heads=[1, 2, 6, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 15, 3],
        **kwargs)
    model.default_cfg = _cfg()

    return model

@register_model
def p2t_large(pretrained=False, **kwargs):
    model = PyramidPoolingTransformer(
        patch_size=4, embed_dims=[64, 128, 320, 640], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3],
        **kwargs)
    model.default_cfg = _cfg()

    return model