Source code for mmseg.models.backbones.beit

# Copyright (c) OpenMMLab. All rights reserved.
import warnings

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from mmcv.cnn import build_norm_layer
from mmcv.cnn.bricks.drop import build_dropout
from mmengine.model import BaseModule, ModuleList
from mmengine.model.weight_init import (constant_init, kaiming_init,
from mmengine.runner.checkpoint import _load_checkpoint
from scipy import interpolate
from torch.nn.modules.batchnorm import _BatchNorm
from torch.nn.modules.utils import _pair as to_2tuple

from mmseg.registry import MODELS
from ..utils import PatchEmbed
from .vit import TransformerEncoderLayer as VisionTransformerEncoderLayer

class BEiTAttention(BaseModule):
    """Window based multi-head self-attention (W-MSA) module with relative
    position bias.

        embed_dims (int): Number of input channels.
        num_heads (int): Number of attention heads.
        window_size (tuple[int]): The height and width of the window.
        bias (bool): The option to add leanable bias for q, k, v. If bias is
            True, it will add leanable bias. If bias is 'qv_bias', it will only
            add leanable bias for q, v. If bias is False, it will not add bias
            for q, k, v. Default to 'qv_bias'.
        qk_scale (float | None, optional): Override default qk scale of
            head_dim ** -0.5 if set. Default: None.
        attn_drop_rate (float): Dropout ratio of attention weight.
            Default: 0.0
        proj_drop_rate (float): Dropout ratio of output. Default: 0.
        init_cfg (dict | None, optional): The Config for initialization.
            Default: None.

    def __init__(self,
        self.embed_dims = embed_dims
        self.num_heads = num_heads
        head_embed_dims = embed_dims // num_heads
        self.bias = bias
        self.scale = qk_scale or head_embed_dims**-0.5

        qkv_bias = bias
        if bias == 'qv_bias':
            qkv_bias = False

        self.window_size = window_size

        self.qkv = nn.Linear(embed_dims, embed_dims * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop_rate)
        self.proj = nn.Linear(embed_dims, embed_dims)
        self.proj_drop = nn.Dropout(proj_drop_rate)

    def _init_qv_bias(self):
        self.q_bias = nn.Parameter(torch.zeros(self.embed_dims))
        self.v_bias = nn.Parameter(torch.zeros(self.embed_dims))

    def _init_rel_pos_embedding(self):
        Wh, Ww = self.window_size
        # cls to token & token 2 cls & cls to cls
        self.num_relative_distance = (2 * Wh - 1) * (2 * Ww - 1) + 3
        # relative_position_bias_table shape is (2*Wh-1 * 2*Ww-1 + 3, nH)
        self.relative_position_bias_table = nn.Parameter(
            torch.zeros(self.num_relative_distance, self.num_heads))

        # get pair-wise relative position index for
        # each token inside the window
        coords_h = torch.arange(Wh)
        coords_w = torch.arange(Ww)
        # coords shape is (2, Wh, Ww)
        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))
        # coords_flatten shape is (2, Wh*Ww)
        coords_flatten = torch.flatten(coords, 1)
        relative_coords = (
            coords_flatten[:, :, None] - coords_flatten[:, None, :])
        # relative_coords shape is (Wh*Ww, Wh*Ww, 2)
        relative_coords = relative_coords.permute(1, 2, 0).contiguous()
        # shift to start from 0
        relative_coords[:, :, 0] += Wh - 1
        relative_coords[:, :, 1] += Ww - 1
        relative_coords[:, :, 0] *= 2 * Ww - 1
        relative_position_index = torch.zeros(
            size=(Wh * Ww + 1, ) * 2, dtype=relative_coords.dtype)
        # relative_position_index shape is (Wh*Ww, Wh*Ww)
        relative_position_index[1:, 1:] = relative_coords.sum(-1)
        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


    def init_weights(self):
        trunc_normal_(self.relative_position_bias_table, std=0.02)

    def forward(self, x):
            x (tensor): input features with shape of (num_windows*B, N, C).
        B, N, C = x.shape

        if self.bias == 'qv_bias':
            k_bias = torch.zeros_like(self.v_bias, requires_grad=False)
            qkv_bias =, k_bias, self.v_bias))
            qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
            qkv = self.qkv(x)

        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]
        q = q * self.scale
        attn = (q @ k.transpose(-2, -1))
        if self.relative_position_bias_table is not None:
            Wh = self.window_size[0]
            Ww = self.window_size[1]
            relative_position_bias = self.relative_position_bias_table[
                    Wh * Ww + 1, Wh * Ww + 1, -1)
            relative_position_bias = relative_position_bias.permute(
                2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
            attn = attn + relative_position_bias.unsqueeze(0)
        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 BEiTTransformerEncoderLayer(VisionTransformerEncoderLayer):
    """Implements one encoder layer in Vision Transformer.

        embed_dims (int): The feature dimension.
        num_heads (int): Parallel attention heads.
        feedforward_channels (int): The hidden dimension for FFNs.
        attn_drop_rate (float): The drop out rate for attention layer.
            Default: 0.0.
        drop_path_rate (float): Stochastic depth rate. Default 0.0.
        num_fcs (int): The number of fully-connected layers for FFNs.
            Default: 2.
        bias (bool): The option to add leanable bias for q, k, v. If bias is
            True, it will add leanable bias. If bias is 'qv_bias', it will only
            add leanable bias for q, v. If bias is False, it will not add bias
            for q, k, v. Default to 'qv_bias'.
        act_cfg (dict): The activation config for FFNs.
            Default: dict(type='GELU').
        norm_cfg (dict): Config dict for normalization layer.
            Default: dict(type='LN').
        window_size (tuple[int], optional): The height and width of the window.
            Default: None.
        init_values (float, optional): Initialize the values of BEiTAttention
            and FFN with learnable scaling. Default: None.

    def __init__(self,
        attn_cfg.update(dict(window_size=window_size, qk_scale=None))


        # NOTE: drop path for stochastic depth, we shall see if
        # this is better than dropout here
        dropout_layer = dict(type='DropPath', drop_prob=drop_path_rate)
        self.drop_path = build_dropout(
            dropout_layer) if dropout_layer else nn.Identity()
        self.gamma_1 = nn.Parameter(
            init_values * torch.ones(embed_dims), requires_grad=True)
        self.gamma_2 = nn.Parameter(
            init_values * torch.ones(embed_dims), requires_grad=True)

    def build_attn(self, attn_cfg):
        self.attn = BEiTAttention(**attn_cfg)

    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.ffn(self.norm2(x)))
        return x

[docs]@MODELS.register_module() class BEiT(BaseModule): """BERT Pre-Training of Image Transformers. Args: img_size (int | tuple): Input image size. Default: 224. patch_size (int): The patch size. Default: 16. in_channels (int): Number of input channels. Default: 3. embed_dims (int): Embedding dimension. Default: 768. num_layers (int): Depth of transformer. Default: 12. num_heads (int): Number of attention heads. Default: 12. mlp_ratio (int): Ratio of mlp hidden dim to embedding dim. Default: 4. out_indices (list | tuple | int): Output from which stages. Default: -1. qv_bias (bool): Enable bias for qv if True. Default: True. attn_drop_rate (float): The drop out rate for attention layer. Default 0.0 drop_path_rate (float): Stochastic depth rate. Default 0.0. norm_cfg (dict): Config dict for normalization layer. Default: dict(type='LN') act_cfg (dict): The activation config for FFNs. Default: dict(type='GELU'). patch_norm (bool): Whether to add a norm in PatchEmbed Block. Default: False. final_norm (bool): Whether to add a additional layer to normalize final feature map. Default: False. num_fcs (int): The number of fully-connected layers for FFNs. Default: 2. norm_eval (bool): Whether to set norm layers to eval mode, namely, freeze running stats (mean and var). Note: Effect on Batch Norm and its variants only. Default: False. pretrained (str, optional): Model pretrained path. Default: None. init_values (float): Initialize the values of BEiTAttention and FFN with learnable scaling. init_cfg (dict or list[dict], optional): Initialization config dict. Default: None. """ def __init__(self, img_size=224, patch_size=16, in_channels=3, embed_dims=768, num_layers=12, num_heads=12, mlp_ratio=4, out_indices=-1, qv_bias=True, attn_drop_rate=0., drop_path_rate=0., norm_cfg=dict(type='LN'), act_cfg=dict(type='GELU'), patch_norm=False, final_norm=False, num_fcs=2, norm_eval=False, pretrained=None, init_values=0.1, init_cfg=None): super().__init__(init_cfg=init_cfg) if isinstance(img_size, int): img_size = to_2tuple(img_size) elif isinstance(img_size, tuple): if len(img_size) == 1: img_size = to_2tuple(img_size[0]) assert len(img_size) == 2, \ f'The size of image should have length 1 or 2, ' \ f'but got {len(img_size)}' assert not (init_cfg and pretrained), \ 'init_cfg and pretrained cannot be set at the same time' if isinstance(pretrained, str): warnings.warn('DeprecationWarning: pretrained is deprecated, ' 'please use "init_cfg" instead') self.init_cfg = dict(type='Pretrained', checkpoint=pretrained) elif pretrained is not None: raise TypeError('pretrained must be a str or None') self.in_channels = in_channels self.img_size = img_size self.patch_size = patch_size self.norm_eval = norm_eval self.pretrained = pretrained self.num_layers = num_layers self.embed_dims = embed_dims self.num_heads = num_heads self.mlp_ratio = mlp_ratio self.attn_drop_rate = attn_drop_rate self.drop_path_rate = drop_path_rate self.num_fcs = num_fcs self.qv_bias = qv_bias self.act_cfg = act_cfg self.norm_cfg = norm_cfg self.patch_norm = patch_norm self.init_values = init_values self.window_size = (img_size[0] // patch_size, img_size[1] // patch_size) self.patch_shape = self.window_size self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dims)) self._build_patch_embedding() self._build_layers() if isinstance(out_indices, int): if out_indices == -1: out_indices = num_layers - 1 self.out_indices = [out_indices] elif isinstance(out_indices, list) or isinstance(out_indices, tuple): self.out_indices = out_indices else: raise TypeError('out_indices must be type of int, list or tuple') self.final_norm = final_norm if final_norm: self.norm1_name, norm1 = build_norm_layer( norm_cfg, embed_dims, postfix=1) self.add_module(self.norm1_name, norm1) def _build_patch_embedding(self): """Build patch embedding layer.""" self.patch_embed = PatchEmbed( in_channels=self.in_channels, embed_dims=self.embed_dims, conv_type='Conv2d', kernel_size=self.patch_size, stride=self.patch_size, padding=0, norm_cfg=self.norm_cfg if self.patch_norm else None, init_cfg=None) def _build_layers(self): """Build transformer encoding layers.""" dpr = [ x.item() for x in torch.linspace(0, self.drop_path_rate, self.num_layers) ] self.layers = ModuleList() for i in range(self.num_layers): self.layers.append( BEiTTransformerEncoderLayer( embed_dims=self.embed_dims, num_heads=self.num_heads, feedforward_channels=self.mlp_ratio * self.embed_dims, attn_drop_rate=self.attn_drop_rate, drop_path_rate=dpr[i], num_fcs=self.num_fcs, bias='qv_bias' if self.qv_bias else False, act_cfg=self.act_cfg, norm_cfg=self.norm_cfg, window_size=self.window_size, init_values=self.init_values)) @property def norm1(self): return getattr(self, self.norm1_name) def _geometric_sequence_interpolation(self, src_size, dst_size, sequence, num): """Get new sequence via geometric sequence interpolation. Args: src_size (int): Pos_embedding size in pre-trained model. dst_size (int): Pos_embedding size in the current model. sequence (tensor): The relative position bias of the pretrain model after removing the extra tokens. num (int): Number of attention heads. Returns: new_sequence (tensor): Geometric sequence interpolate the pre-trained relative position bias to the size of the current model. """ def geometric_progression(a, r, n): return a * (1.0 - r**n) / (1.0 - r) # Here is a binary function. left, right = 1.01, 1.5 while right - left > 1e-6: q = (left + right) / 2.0 gp = geometric_progression(1, q, src_size // 2) if gp > dst_size // 2: right = q else: left = q # The position of each interpolated point is determined # by the ratio obtained by dichotomy. dis = [] cur = 1 for i in range(src_size // 2): dis.append(cur) cur += q**(i + 1) r_ids = [-_ for _ in reversed(dis)] x = r_ids + [0] + dis y = r_ids + [0] + dis t = dst_size // 2.0 dx = np.arange(-t, t + 0.1, 1.0) dy = np.arange(-t, t + 0.1, 1.0) # Interpolation functions are being executed and called. new_sequence = [] for i in range(num): z = sequence[:, i].view(src_size, src_size).float().numpy() f = interpolate.interp2d(x, y, z, kind='cubic') new_sequence.append( torch.Tensor(f(dx, dy)).contiguous().view(-1, 1).to(sequence)) new_sequence =, dim=-1) return new_sequence
[docs] def resize_rel_pos_embed(self, checkpoint): """Resize relative pos_embed weights. This function is modified from # noqa: E501 Copyright (c) Microsoft Corporation Licensed under the MIT License Args: checkpoint (dict): Key and value of the pretrain model. Returns: state_dict (dict): Interpolate the relative pos_embed weights in the pre-train model to the current model size. """ if 'state_dict' in checkpoint: state_dict = checkpoint['state_dict'] else: state_dict = checkpoint all_keys = list(state_dict.keys()) for key in all_keys: if 'relative_position_index' in key: state_dict.pop(key) # In order to keep the center of pos_bias as consistent as # possible after interpolation, and vice versa in the edge # area, the geometric sequence interpolation method is adopted. if 'relative_position_bias_table' in key: rel_pos_bias = state_dict[key] src_num_pos, num_attn_heads = rel_pos_bias.size() dst_num_pos, _ = self.state_dict()[key].size() dst_patch_shape = self.patch_shape if dst_patch_shape[0] != dst_patch_shape[1]: raise NotImplementedError() # Count the number of extra tokens. num_extra_tokens = dst_num_pos - ( dst_patch_shape[0] * 2 - 1) * ( dst_patch_shape[1] * 2 - 1) src_size = int((src_num_pos - num_extra_tokens)**0.5) dst_size = int((dst_num_pos - num_extra_tokens)**0.5) if src_size != dst_size: extra_tokens = rel_pos_bias[-num_extra_tokens:, :] rel_pos_bias = rel_pos_bias[:-num_extra_tokens, :] new_rel_pos_bias = self._geometric_sequence_interpolation( src_size, dst_size, rel_pos_bias, num_attn_heads) new_rel_pos_bias = (new_rel_pos_bias, extra_tokens), dim=0) state_dict[key] = new_rel_pos_bias return state_dict
[docs] def init_weights(self): def _init_weights(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) self.apply(_init_weights) if (isinstance(self.init_cfg, dict) and self.init_cfg.get('type') == 'Pretrained'): checkpoint = _load_checkpoint( self.init_cfg['checkpoint'], logger=None, map_location='cpu') state_dict = self.resize_rel_pos_embed(checkpoint) self.load_state_dict(state_dict, False) elif self.init_cfg is not None: super().init_weights() else: # We only implement the 'jax_impl' initialization implemented at # # noqa: E501 # Copyright 2019 Ross Wightman # Licensed under the Apache License, Version 2.0 (the "License") trunc_normal_(self.cls_token, std=.02) for n, m in self.named_modules(): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if m.bias is not None: if 'ffn' in n: nn.init.normal_(m.bias, mean=0., std=1e-6) else: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.Conv2d): kaiming_init(m, mode='fan_in', bias=0.) elif isinstance(m, (_BatchNorm, nn.GroupNorm, nn.LayerNorm)): constant_init(m, val=1.0, bias=0.)
[docs] def forward(self, inputs): B = inputs.shape[0] x, hw_shape = self.patch_embed(inputs) # stole cls_tokens impl from Phil Wang, thanks cls_tokens = self.cls_token.expand(B, -1, -1) x =, x), dim=1) outs = [] for i, layer in enumerate(self.layers): x = layer(x) if i == len(self.layers) - 1: if self.final_norm: x = self.norm1(x) if i in self.out_indices: # Remove class token and reshape token for decoder head out = x[:, 1:] B, _, C = out.shape out = out.reshape(B, hw_shape[0], hw_shape[1], C).permute(0, 3, 1, 2).contiguous() outs.append(out) return tuple(outs)
[docs] def train(self, mode=True): super().train(mode) if mode and self.norm_eval: for m in self.modules(): if isinstance(m, nn.LayerNorm): m.eval()
Read the Docs v: latest
On Read the Docs
Project Home

Free document hosting provided by Read the Docs.