vision_transformer_attn.py

"""
Mostly copy-paste from timm library.
https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
"""
import math
import torch
import torch.nn as nn
from functools import partial

from utils import trunc_normal_
from loss_functions import generate_foreground_mask

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):
    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
    """
    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, 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

class Attention(nn.Module):
    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., attn_wo_soft = False):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5
        self.attn_wo_soft = attn_wo_soft

        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, curr_layer_mask=None):
        """
        current version only cuts off relationship between cls token and background tokens 
        """
        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]
        attn = (q @ k.transpose(-2, -1)) * self.scale #[B, H, N, N]

        if curr_layer_mask is not None:
            attn[:, :, 0, 1:] -= 1e9 * (1-curr_layer_mask).unsqueeze(1)
            attn[:, :, 1:, 0] -= 1e9 * (1-curr_layer_mask).unsqueeze(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, attn
        
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, attn_wo_soft = False):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, attn_wo_soft = attn_wo_soft)
        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, return_attention=False, curr_layer_mask = None):
        y, attn = self.attn(self.norm1(x), curr_layer_mask)
        x = x + self.drop_path(y)
        x = x + self.drop_path(self.mlp(self.norm2(x)))
        if return_attention:
            return x, attn
        return x

class PatchEmbed(nn.Module):
    """ Image to Patch Embedding
    """
    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, frozen_patch_embed = False):
        super().__init__()
        num_patches = (img_size // patch_size) * (img_size // patch_size)
        self.img_size = img_size
        self.patch_size = patch_size
        self.num_patches = num_patches
        if frozen_patch_embed:
            print("frozen patch embedding !")
            with torch.no_grad():
                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=patch_size, stride=patch_size)
        
    def forward(self, x):
        B, C, H, W = x.shape
        x = self.proj(x)  # [B, C, W, H]
        x = x.flatten(2).transpose(1, 2)
        return x

class VisionTransformer(nn.Module):
    """ Vision Transformer """
    def __init__(self, img_size=[224], patch_size=16, in_chans=3, num_classes=0, 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, is_student=False, frozen_patch_embed = False, attn_wo_soft = False, **kwargs):
        super().__init__()
        
        self.img_size = img_size
        self.patch_size = patch_size 
        self.in_chans = in_chans 
        self.frozen_patch_embed = frozen_patch_embed
        
        self.is_student = is_student
        self.attn_wo_soft = attn_wo_soft
        self.num_features = self.embed_dim = embed_dim

        self.patch_embed = PatchEmbed(
            img_size=img_size[0], patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, frozen_patch_embed = frozen_patch_embed)
        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 + 1, embed_dim))
        self.pos_drop = nn.Dropout(p=drop_rate)

        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
        self.blocks = nn.ModuleList([
            Block(
                dim=embed_dim, 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[i], norm_layer=norm_layer, attn_wo_soft = attn_wo_soft)
            for i in range(depth)])
        self.norm = norm_layer(embed_dim)

        # Classifier 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)

    def reset_patch_embed_size(self, patch_size):
        self.patch_size = patch_size
        self.patch_embed = PatchEmbed(
            img_size=self.img_size[0], patch_size=self.patch_size, in_chans=self.in_chans, embed_dim=self.embed_dim, frozen_patch_embed = self.frozen_patch_embed)#.cuda()
        num_patches = self.patch_embed.num_patches        
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, self.embed_dim))
        trunc_normal_(self.pos_embed, std=.02)
        
    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 interpolate_pos_encoding(self, x, w, h):
        npatch = x.shape[1] - 1 # 144
        N = self.pos_embed.shape[1] - 1 #784
        if npatch == N and w == h:
            return self.pos_embed
        class_pos_embed = self.pos_embed[:, 0]
        patch_pos_embed = self.pos_embed[:, 1:]
        dim = x.shape[-1]
        w0 = w // self.patch_embed.patch_size #12
        h0 = h // self.patch_embed.patch_size #12
        # we add a small number to avoid floating point error in the interpolation
        # see discussion at https://github.com/facebookresearch/dino/issues/8
        w0, h0 = w0 + 0.1, h0 + 0.1
        patch_pos_embed = nn.functional.interpolate(
            patch_pos_embed.reshape(1, int(math.sqrt(N)), int(math.sqrt(N)), dim).permute(0, 3, 1, 2),
            scale_factor=(w0 / math.sqrt(N), h0 / math.sqrt(N)),
            mode='bicubic',
        )
        assert int(w0) == patch_pos_embed.shape[-2] and int(h0) == patch_pos_embed.shape[-1]
        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
        return torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1)

    def prepare_tokens(self, x):
        B, nc, w, h = x.shape
        x = self.patch_embed(x)  # patch linear embedding

        # add the [CLS] token to the embed patch tokens
        cls_tokens = self.cls_token.expand(B, -1, -1)
        x = torch.cat((cls_tokens, x), dim=1)

        # add positional encoding to each token
        x = x + self.interpolate_pos_encoding(x, w, h)
 
        return self.pos_drop(x)

    def forward(self, x, keep_last_k_attns = False, last_k_attns = 1, keep_last_k_cls = False, last_k_cls = 1, fixed_mask = None, rollout_teacher = False, rollout_student=False, foreground_threshold=0.6):
        x = self.prepare_tokens(x) # [B, N, N-1]
        if fixed_mask is not None:
            x = x[:, fixed_mask, :] # select unmasked indexes (with patch drop ratio)
            
        if keep_last_k_attns:
            last_k_attns_list = []
        if keep_last_k_cls:
            last_k_cls_list = []

        for i, blk in enumerate(self.blocks):
            if not self.is_student:
                x = blk(x)
            else:
                if rollout_teacher:
                    if i==0:
                        x, attn = blk(x, return_attention=True)
                        B, H, N, _ = attn.shape
                        rollout_matrix = attn.mean(axis = 1) + torch.eye(N).unsqueeze(0).repeat(B, 1, 1).cuda() 
                    elif i < len(self.blocks) - 1:
                        x, attn = blk(x, return_attention=True)
                        attention_matrix = attn.mean(axis = 1) + torch.eye(N).unsqueeze(0).repeat(B, 1, 1).cuda() 
                        rollout_matrix = torch.bmm(attention_matrix, rollout_matrix)
                    else:
                        cls_rollout = rollout_matrix[:,0,1:]
                        th_attn = generate_foreground_mask(attn=cls_rollout, gaussianblur_kernel_size=1, blur_sigma=1,
                                                    foreground_threshold=foreground_threshold, remove_component_less_than_pixels=1)
                        x, attn = blk(x, return_attention=True, curr_layer_mask = th_attn)
                elif rollout_student:
                    if i < len(self.blocks) - 1:
                        x = blk(x)
                    else:
                        x, attn = blk(x, return_attention=True, curr_layer_mask = fixed_mask)                    
                else:
                    if i < len(self.blocks) - 1 and i < len(self.blocks) - max(last_k_attns, last_k_cls):
                        x = blk(x, curr_layer_mask = None)
                    elif (i < len(self.blocks) - 1 and i >= len(self.blocks) - last_k_attns and keep_last_k_attns) or \
                         (i < len(self.blocks) - 1 and i >= len(self.blocks) - last_k_cls and keep_last_k_cls):
                        x, attn = blk(x, return_attention=True, curr_layer_mask = None)
                        
                        if keep_last_k_attns:
                            last_k_attns_list.append(attn[:,:,0,1:])
                        if keep_last_k_cls:
                            last_k_cls_list.append(self.norm(x)[:,0])
                    else:
                        x, attn= blk(x, return_attention=True, curr_layer_mask = None)
                        if keep_last_k_attns:
                            last_k_attns_list.append(attn[:,:,0,1:])
                        
        x = self.norm(x)
        if not self.is_student:
            return x[:,0]
        else:
            if keep_last_k_attns:
                return x, attn, last_k_attns_list, None
            elif keep_last_k_cls:
                return x, attn, None, last_k_cls_list 
            else:
                return x, attn, None, None

    def get_last_selfattention(self, x, rollout=False, foreground_threshold=0.5):
        x = self.prepare_tokens(x)
        for i, blk in enumerate(self.blocks):
            if rollout:
                if i==0:
                    x, attn = blk(x, return_attention=True)
                    B, H, N, _ = attn.shape
                    rollout_matrix = attn.mean(axis = 1) + torch.eye(N).unsqueeze(0).repeat(B, 1, 1).cuda() 
                elif i < len(self.blocks) - 1:
                    x, attn = blk(x, return_attention=True)
                    attention_matrix = attn.mean(axis = 1) + torch.eye(N).unsqueeze(0).repeat(B, 1, 1).cuda() 
                    rollout_matrix = torch.bmm(attention_matrix, rollout_matrix)
                else:
                    cls_rollout = rollout_matrix[:,0,1:]
                    th_attn = generate_foreground_mask(attn=cls_rollout, gaussianblur_kernel_size=1, blur_sigma=1,
                                                foreground_threshold=foreground_threshold, remove_component_less_than_pixels=1)
                    x, attn = blk(x, return_attention=True, curr_layer_mask = th_attn)
                    return x, attn
            else:
                if i < len(self.blocks) - 1:
                    x = blk(x)    
                else:
                    return blk(x, return_attention=True)

    def get_intermediate_layers(self, x, n=1, rollout=False, foreground_threshold=0.5):
        x = self.prepare_tokens(x)
        # we return the output tokens from the `n` last blocks
        output = []
        for i, blk in enumerate(self.blocks):
            if rollout:
                if i==0:
                    x, attn = blk(x, return_attention=True)
                    B, H, N, _ = attn.shape
                    rollout_matrix = attn.mean(axis = 1) + torch.eye(N).unsqueeze(0).repeat(B, 1, 1).cuda() 
                elif len(self.blocks) - i > n:
                    x, attn = blk(x, return_attention=True)
                    attention_matrix = attn.mean(axis = 1) + torch.eye(N).unsqueeze(0).repeat(B, 1, 1).cuda() 
                    rollout_matrix = torch.bmm(attention_matrix, rollout_matrix)
                elif len(self.blocks) - i <= n and len(self.blocks) - i > 1:
                    x, attn = blk(x, return_attention=True)
                    attention_matrix = attn.mean(axis = 1) + torch.eye(N).unsqueeze(0).repeat(B, 1, 1).cuda() 
                    rollout_matrix = torch.bmm(attention_matrix, rollout_matrix)
                    output.append(self.norm(x))
                elif len(self.blocks) - i == 1: # last layer 
                    cls_rollout = rollout_matrix[:,0,1:]
                    th_attn = generate_foreground_mask(attn=cls_rollout, gaussianblur_kernel_size=1, blur_sigma=1,
                                            foreground_threshold=foreground_threshold, remove_component_less_than_pixels=5)
                    x, attn = blk(x, return_attention=True, curr_layer_mask = th_attn)
                    output.append(self.norm(x))
            else:
                x = blk(x)
                if len(self.blocks) - i <= n:
                    output.append(self.norm(x))
        return output

    def get_last_k_selfattention(self, x, k=1, rollout=False, foreground_threshold=0.5):
        x = self.prepare_tokens(x)
        if rollout:
            for i, blk in enumerate(self.blocks):
                if i==0:
                    x, attn = blk(x, return_attention=True)
                    B, H, N, _ = attn.shape
                    rollout_matrix = attn.mean(axis = 1) + torch.eye(N).unsqueeze(0).repeat(B, 1, 1).cuda() 
                elif i < len(self.blocks) - k:
                    x, attn = blk(x, return_attention=True)
                    attention_matrix = attn.mean(axis = 1) + torch.eye(N).unsqueeze(0).repeat(B, 1, 1).cuda() 
                    rollout_matrix = torch.bmm(attention_matrix, rollout_matrix)
                else:
                    cls_rollout = rollout_matrix[:,0,1:]
                    th_attn = generate_foreground_mask(attn=cls_rollout, gaussianblur_kernel_size=1, blur_sigma=1,
                                                foreground_threshold=foreground_threshold, remove_component_less_than_pixels=1)
                    x, attn = blk(x, return_attention=True, curr_layer_mask = th_attn)
                    return x, attn
        else:
            for i, blk in enumerate(self.blocks):
                if i < len(self.blocks) - k:
                    x = blk(x)
                else:
                    return blk(x, return_attention=True)
            
def vit_tiny(patch_size=16, **kwargs):
    model = VisionTransformer(
        patch_size=patch_size, embed_dim=192, depth=12, num_heads=3, mlp_ratio=4,
        qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    return model

def vit_small(patch_size=16, **kwargs):
    model = VisionTransformer(
        patch_size=patch_size, embed_dim=384, depth=12, num_heads=6, mlp_ratio=4,
        qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    return model

def vit_small_fixed_num_patches(num_patches=16, **kwargs):
    model = VisionTransformer(
        num_patches=num_patches, embed_dim=384, depth=12, num_heads=6, mlp_ratio=4,
        qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    return model

def vit_base(patch_size=16, **kwargs):
    model = VisionTransformer(
        patch_size=patch_size, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4,
        qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    return model