# Part of the implementation is borrowed and modified from beit2, # publicly available at https://github.com/microsoft/unilm/tree/master/beit2 import collections.abc import itertools import math import os import warnings from functools import partial from typing import Dict, Sequence import torch import torch.nn as nn import torch.nn.functional as F from einops import rearrange from mmcls.models.backbones.base_backbone import BaseBackbone from mmcls.models.builder import BACKBONES from mmcv.cnn.bricks import DropPath, build_activation_layer, build_norm_layer from mmcv.runner import BaseModule from torch.nn.modules.batchnorm import _BatchNorm from ..utils import to_2tuple, trunc_normal_ 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) # commit this for the original BERT implement 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., window_size=None, attn_head_dim=None): super().__init__() self.num_heads = num_heads head_dim = dim // num_heads if attn_head_dim is not None: head_dim = attn_head_dim all_head_dim = head_dim * self.num_heads self.scale = qk_scale or head_dim**-0.5 self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False) if qkv_bias: self.q_bias = nn.Parameter(torch.zeros(all_head_dim)) self.v_bias = nn.Parameter(torch.zeros(all_head_dim)) else: self.q_bias = None self.v_bias = None if window_size: self.window_size = window_size self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3 self.relative_position_bias_table = nn.Parameter( torch.zeros(self.num_relative_distance, num_heads)) # 2*Wh-1 * 2*Ww-1, nH # cls to token & token 2 cls & cls to cls # get pair-wise relative position index for each token inside the window coords_h = torch.arange(window_size[0]) coords_w = torch.arange(window_size[1]) coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww relative_coords = relative_coords.permute( 1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 relative_coords[:, :, 0] += window_size[0] - 1 # shift to start from 0 relative_coords[:, :, 1] += window_size[1] - 1 relative_coords[:, :, 0] *= 2 * window_size[1] - 1 relative_position_index = \ torch.zeros(size=(window_size[0] * window_size[1] + 1, ) * 2, dtype=relative_coords.dtype) relative_position_index[1:, 1:] = relative_coords.sum( -1) # Wh*Ww, Wh*Ww relative_position_index[0, 0:] = self.num_relative_distance - 3 relative_position_index[0:, 0] = self.num_relative_distance - 2 relative_position_index[0, 0] = self.num_relative_distance - 1 self.register_buffer('relative_position_index', relative_position_index) else: self.window_size = None self.relative_position_bias_table = None self.relative_position_index = None self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(all_head_dim, dim) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x, rel_pos_bias=None, return_attention=False, return_qkv=False): B, N, C = x.shape qkv_bias = None if self.q_bias is not None: qkv_bias = torch.cat( (self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias)) # qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias) qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[ 2] # make torchscript happy (cannot use tensor as tuple) (B, H, N, C) q = q * self.scale attn = (q @ k.transpose(-2, -1)) if self.relative_position_bias_table is not None: relative_position_bias = \ self.relative_position_bias_table[self.relative_position_index.view(-1)].view( self.window_size[0] * self.window_size[1] + 1, self.window_size[0] * self.window_size[1] + 1, -1) # Wh*Ww,Wh*Ww,nH relative_position_bias = relative_position_bias.permute( 2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww attn = attn + relative_position_bias.unsqueeze(0) if rel_pos_bias is not None: attn = attn + rel_pos_bias attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) if return_attention: return attn x = (attn @ v).transpose(1, 2).reshape(B, N, -1) x = self.proj(x) x = self.proj_drop(x) if return_qkv: return x, qkv 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., init_values=None, act_layer=nn.GELU, norm_layer=nn.LayerNorm, window_size=None, attn_head_dim=None): 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, window_size=window_size, attn_head_dim=attn_head_dim) # 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) if init_values > 0: 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) else: self.gamma_1, self.gamma_2 = None, None def forward(self, x, rel_pos_bias=None, return_attention=False, return_qkv=False): if return_attention: return self.attn( self.norm1(x), rel_pos_bias=rel_pos_bias, return_attention=True) if return_qkv: y, qkv = self.attn( self.norm1(x), rel_pos_bias=rel_pos_bias, return_qkv=return_qkv) x = x + self.drop_path(self.gamma_1 * y) x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x))) return x, qkv if self.gamma_1 is None: x = x + self.drop_path( self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias)) x = x + self.drop_path(self.mlp(self.norm2(x))) else: x = x + self.drop_path(self.gamma_1 * self.attn( self.norm1(x), rel_pos_bias=rel_pos_bias)) x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x))) 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): 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.patch_shape = (img_size[0] // patch_size[0], img_size[1] // patch_size[1]) self.img_size = img_size self.patch_size = patch_size self.num_patches = num_patches self.proj = nn.Conv2d( in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) def forward(self, x, **kwargs): 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 RelativePositionBias(nn.Module): def __init__(self, window_size, num_heads): super().__init__() self.window_size = window_size self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3 self.relative_position_bias_table = nn.Parameter( torch.zeros(self.num_relative_distance, num_heads)) # 2*Wh-1 * 2*Ww-1, nH # cls to token & token 2 cls & cls to cls # get pair-wise relative position index for each token inside the window coords_h = torch.arange(window_size[0]) coords_w = torch.arange(window_size[1]) coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww relative_coords = relative_coords.permute( 1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 relative_coords[:, :, 0] += window_size[0] - 1 # shift to start from 0 relative_coords[:, :, 1] += window_size[1] - 1 relative_coords[:, :, 0] *= 2 * window_size[1] - 1 relative_position_index = \ torch.zeros(size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype) relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww relative_position_index[0, 0:] = self.num_relative_distance - 3 relative_position_index[0:, 0] = self.num_relative_distance - 2 relative_position_index[0, 0] = self.num_relative_distance - 1 self.register_buffer('relative_position_index', relative_position_index) # trunc_normal_(self.relative_position_bias_table, std=.02) def forward(self): relative_position_bias = \ self.relative_position_bias_table[self.relative_position_index.view(-1)].view( self.window_size[0] * self.window_size[1] + 1, self.window_size[0] * self.window_size[1] + 1, -1) # Wh*Ww,Wh*Ww,nH return relative_position_bias.permute( 2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww @BACKBONES.register_module() class BEiTv2(BaseBackbone): embed_dims = {'base': 768, 'large': 1024, 'huge': 1280, 'giant': 1408} depths = {'base': 12, 'large': 24, 'huge': 32, 'giant': 40} num_heads = {'base': 12, 'large': 16, 'huge': 16, 'giant': 16} mlp_ratios = {'base': 4, 'large': 4, 'huge': 4, 'giant': 6144 / 1408} """ Vision Transformer with support for patch or hybrid CNN input stage """ def __init__(self, arch='base', patch_size=16, img_size=224, in_chans=3, qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=partial(nn.LayerNorm, eps=1e-6), init_values=None, use_abs_pos_emb=True, use_rel_pos_bias=False, use_shared_rel_pos_bias=False, use_mean_pooling=True, init_scale=0.001, out_indices=-1, frozen_stages=-1, init_cfg=None): super().__init__(init_cfg=init_cfg) embed_dim = self.embed_dims[arch] depth = self.depths[arch] num_heads = self.num_heads[arch] mlp_ratio = self.mlp_ratios[arch] self.out_indices = out_indices self.frozen_stages = frozen_stages self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models self.patch_embed = PatchEmbed( 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.mask_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) if use_abs_pos_emb: self.pos_embed = nn.Parameter( torch.zeros(1, num_patches + 1, embed_dim)) else: self.pos_embed = None self.pos_drop = nn.Dropout(p=drop_rate) if use_shared_rel_pos_bias: self.rel_pos_bias = RelativePositionBias( window_size=self.patch_embed.patch_shape, num_heads=num_heads) else: self.rel_pos_bias = None dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth) ] # stochastic depth decay rule self.use_rel_pos_bias = use_rel_pos_bias 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, init_values=init_values, window_size=self.patch_embed.patch_shape if use_rel_pos_bias else None) for i in range(depth) ]) self.norm = nn.Identity() if use_mean_pooling else norm_layer( embed_dim) self.fc_norm = norm_layer(embed_dim) if use_mean_pooling else None if self.pos_embed is not None: trunc_normal_(self.pos_embed, std=.02) trunc_normal_(self.cls_token, std=.02) def init_weights(self): super(BEiTv2, self).init_weights() if (isinstance(self.init_cfg, dict) and self.init_cfg['type'] == 'Pretrained'): # Suppress default init if use pretrained model. return self.apply(self._init_weights) self.fix_init_weight() 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 fix_init_weight(self): def rescale(param, layer_id): param.div_(math.sqrt(2.0 * layer_id)) for layer_id, layer in enumerate(self.blocks): rescale(layer.attn.proj.weight.data, layer_id + 1) rescale(layer.mlp.fc2.weight.data, layer_id + 1) def interpolate_pos_encoding(self, x, w, h): npatch = x.shape[1] - 1 N = self.pos_embed.shape[1] - 1 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[0] h0 = h // self.patch_embed.patch_size[0] # 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 forward_features(self, x, return_patch_tokens=False, return_all_tokens=False, **kwargs): B, nc, w, h = x.shape x = self.patch_embed(x) batch_size, seq_len, _ = x.size() cls_tokens = self.cls_token.expand( batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks x = torch.cat((cls_tokens, x), dim=1) if self.pos_embed is not None: if x.shape[1] != self.pos_embed.shape[1]: x = x + self.interpolate_pos_encoding(x, w, h) else: x = x + self.pos_embed x = self.pos_drop(x) rel_pos_bias = self.rel_pos_bias( ) if self.rel_pos_bias is not None else None for blk in self.blocks: x = blk(x, rel_pos_bias=rel_pos_bias) x = self.norm(x) if self.fc_norm is not None: if return_all_tokens: return self.fc_norm(x) t = x[:, 1:, :] if return_patch_tokens: return self.fc_norm(t) else: return self.fc_norm(t.mean(1)) else: if return_all_tokens: return x elif return_patch_tokens: return x[:, 1:] else: return x[:, 0] def forward(self, x, return_patch_tokens=False, return_all_tokens=False, **kwargs): x = self.forward_features( x, return_patch_tokens=return_patch_tokens, return_all_tokens=return_all_tokens, **kwargs) return tuple([x]) def _freeze_stages(self): if self.frozen_stages > 0: self.patch_embed.eval() for param in self.patch_embed.parameters(): param.requires_grad = False for idx, layer in enumerate(self.blocks): if idx <= self.frozen_stages - 1: layer.eval() for param in layer.parameters(): param.requires_grad = False def train(self, mode=True): super(BEiTv2, self).train(mode) self._freeze_stages()