models_v4.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.

import torch
import torch.nn as nn
from functools import partial
import math
from timm.models.vision_transformer import Mlp, PatchEmbed , _cfg

from timm.models.layers import DropPath, to_2tuple, trunc_normal_
from timm.models.registry import register_model
import torch.nn.functional as F

class Attention(nn.Module):
    # taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
    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.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]
        
        q = q * self.scale

        attn = (q @ k.transpose(-2, -1))
        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):
    # taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
    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,Attention_block = Attention,Mlp_block=Mlp
                 ,init_values=1e-4):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention_block(
            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.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp_block(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

    def forward(self, x, lambda_v=1):
        # x = x + self.drop_path(self.attn(self.norm1(x))) # origin
        x = (2 - lambda_v) * x + lambda_v * self.drop_path(self.attn(self.norm1(x)))
        x = x + self.drop_path(self.mlp(self.norm2(x))) # origin
        # x = (2 - lambda_v) * x + lambda_v * self.drop_path(self.mlp(self.norm2(x)))
        return x
    
class Layer_scale_init_Block(nn.Module):
    # taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
    # with slight modifications
    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,Attention_block = Attention,Mlp_block=Mlp
                 ,init_values=1e-4):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention_block(
            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.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp_block(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
        self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
        self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)

    def forward(self, x):
        x = x + self.drop_path(self.gamma_1 * self.attn(self.norm1(x)))
        x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))
        return x

class Layer_scale_init_Block_paralx2(nn.Module):
    # taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
    # with slight modifications
    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,Attention_block = Attention,Mlp_block=Mlp
                 ,init_values=1e-4):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.norm11 = norm_layer(dim)
        self.attn = Attention_block(
            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.attn1 = Attention_block(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        self.norm21 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp_block(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
        self.mlp1 = Mlp_block(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
        self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
        self.gamma_1_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
        self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
        self.gamma_2_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
        
    def forward(self, x):
        x = x + self.drop_path(self.gamma_1*self.attn(self.norm1(x))) + self.drop_path(self.gamma_1_1 * self.attn1(self.norm11(x)))
        x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x))) + self.drop_path(self.gamma_2_1 * self.mlp1(self.norm21(x)))
        return x
        
class Block_paralx2(nn.Module):
    # taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
    # with slight modifications
    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,Attention_block = Attention,Mlp_block=Mlp
                 ,init_values=1e-4):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.norm11 = norm_layer(dim)
        self.attn = Attention_block(
            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.attn1 = Attention_block(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        self.norm21 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp_block(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
        self.mlp1 = Mlp_block(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))) + self.drop_path(self.attn1(self.norm11(x)))
        x = x + self.drop_path(self.mlp(self.norm2(x))) + self.drop_path(self.mlp1(self.norm21(x)))
        return x
        
        
class hMLP_stem(nn.Module):
    """ hMLP_stem: https://arxiv.org/pdf/2203.09795.pdf
    taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
    with slight modifications
    """
    def __init__(self, img_size=224,  patch_size=16, in_chans=3, embed_dim=768,norm_layer=nn.SyncBatchNorm):
        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
        self.proj = torch.nn.Sequential(*[nn.Conv2d(in_chans, embed_dim//4, kernel_size=4, stride=4),
                                          norm_layer(embed_dim//4),
                                          nn.GELU(),
                                          nn.Conv2d(embed_dim//4, embed_dim//4, kernel_size=2, stride=2),
                                          norm_layer(embed_dim//4),
                                          nn.GELU(),
                                          nn.Conv2d(embed_dim//4, embed_dim, kernel_size=2, stride=2),
                                          norm_layer(embed_dim),
                                         ])
        

    def forward(self, x):
        B, C, H, W = x.shape
        x = self.proj(x).flatten(2).transpose(1, 2)
        return x
    
class vit_models(nn.Module):
    """ Vision Transformer with LayerScale (https://arxiv.org/abs/2103.17239) support
    taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
    with slight modifications
    """
    def __init__(self, img_size=224,  patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,
                 num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
                 drop_path_rate=0., norm_layer=nn.LayerNorm, global_pool=None,
                 block_layers = Block,
                 Patch_layer=PatchEmbed,act_layer=nn.GELU,
                 Attention_block = Attention, Mlp_block=Mlp,
                dpr_constant=True,init_scale=1e-4,
                mlp_ratio_clstk = 4.0,**kwargs):
        super().__init__()
        
        self.dropout_rate = drop_rate

            
        self.num_classes = num_classes
        self.num_features = self.embed_dim = embed_dim

        self.patch_embed = Patch_layer(
                img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
        num_patches = self.patch_embed.num_patches

        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))

        #self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches+1, embed_dim)) # align timm

        dpr = [drop_path_rate for i in range(depth)]
        self.blocks = nn.ModuleList([
            block_layers(
                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
                drop=0.0, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
                act_layer=act_layer,Attention_block=Attention_block,Mlp_block=Mlp_block,init_values=init_scale)
            for i in range(depth)])
        

        
            
        self.norm = norm_layer(embed_dim)

        self.feature_info = [dict(num_chs=embed_dim, reduction=0, module='head')]
        self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()

        trunc_normal_(self.pos_embed, std=.02)
        trunc_normal_(self.cls_token, std=.02)
        self.apply(self._init_weights)
        # lambda: the momentum value
        self.lambda_v = None
        self.prune_layer = kwargs['prune_layer']
        self.done_layer = kwargs['done_layer']
        self.frozen_stages = kwargs['frozen_stages']
        self._freeze_stages()
        print(f'the pruning layer is {self.prune_layer}, the pruned layer is {self.done_layer}, frozen_stages is {self.frozen_stages}')

    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'}

    def get_classifier(self):
        return self.head
    
    def get_num_layers(self):
        return len(self.blocks)
    
    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, lambda_v=1):
        B = x.shape[0]
        x = self.patch_embed(x)

        cls_tokens = self.cls_token.expand(B, -1, -1)
        
        #x = x + self.pos_embed
        
        x = torch.cat((cls_tokens, x), dim=1)
        x = x + self.pos_embed # align timm
            
        for i , blk in enumerate(self.blocks):
            #if i in [1, 2, 3, 4]:
            if i in self.prune_layer:
                #print(i)
                x = blk(x,lambda_v)
            elif i in self.done_layer:
                x = blk(x,lambda_v=0)
            else:
                x = blk(x,lambda_v=1)
            
        x = self.norm(x)
        return x[:, 0]

    def forward(self, x, epoch=0):
        start_decay = 0
        if epoch <=start_decay:
            magic_lambda = 1.0
        else:
            magic_lambda = 1.0 - ( (epoch-start_decay) / 299)
        magic_lambda = 0 if magic_lambda < 0 else magic_lambda
        if self.lambda_v == None or self.lambda_v !=magic_lambda:
            print(f'Magic lambda is {magic_lambda}')
            self.lambda_v = magic_lambda
        x = self.forward_features(x,self.lambda_v)
        
        if self.dropout_rate:
            x = F.dropout(x, p=float(self.dropout_rate), training=self.training)
        x = self.head(x)
        
        return x

    def _freeze_stages(self):
        """Freeze stages param and norm stats."""

        if self.frozen_stages >= 0:
            self.cls_token.requires_grad = False
            self.pos_embed.requires_grad = False
            self.patch_embed.eval()
            for param in self.patch_embed.parameters():
                param.requires_grad = False
        
        for i in range(0,self.frozen_stages+1):
            self.blocks[i].eval()
            for p in self.blocks[i].parameters():
                p.requires_grad = False

#--------------------------- our models are hear ------------------

@register_model
def deit_base_patch16_224_attn(pretrained=False, img_size=224, pretrained_21k = False,  **kwargs):
    model = vit_models(
        img_size = img_size, patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    if pretrained:
        name = 'https://dl.fbaipublicfiles.com/deit/deit_3_base_'+str(img_size)+'_'
        if pretrained_21k:
            name+='21k.pth'
        else:
            name+='1k.pth'

        checkpoint = torch.hub.load_state_dict_from_url(
            url=name,
            map_location="cpu", check_hash=True
        )
        model.load_state_dict(checkpoint["model"])
    return model