mmpretrain/models/necks/cae_neck.py

# Copyright (c) OpenMMLab. All rights reserved.
from typing import Tuple

import torch
import torch.nn as nn
from mmcv.cnn import build_norm_layer
from mmcv.cnn.bricks import DropPath
from mmcv.cnn.bricks.transformer import FFN
from mmengine.model import BaseModule
from mmengine.model.weight_init import trunc_normal_

from mmpretrain.models.backbones.beit import BEiTTransformerEncoderLayer
from mmpretrain.registry import MODELS
from ..utils import CrossMultiheadAttention


class CAETransformerRegressorLayer(BaseModule):
    """Transformer layer for the regressor of CAE.

    This module is different from conventional transformer encoder layer, for
    its queries are the masked tokens, but its keys and values are the
    concatenation of the masked and unmasked tokens.

    Args:
        embed_dims (int): The feature dimension.
        num_heads (int): The number of heads in multi-head attention.
        feedforward_channels (int): The hidden dimension of FFNs.
            Defaults: 1024.
        num_fcs (int, optional): The number of fully-connected layers in
            FFNs. Default: 2.
        qkv_bias (bool): If True, add a learnable bias to q, k, v.
            Defaults to True.
        qk_scale (float, optional): Override default qk scale of
            ``head_dim ** -0.5`` if set. Defaults to None.
        drop_rate (float): The dropout rate. Defaults to 0.0.
        attn_drop_rate (float): The drop out rate for attention output weights.
            Defaults to 0.
        drop_path_rate (float): Stochastic depth rate. Defaults to 0.
        layer_scale_init_value (float): The init value of gamma.
            Defaults to 0.0.
        act_cfg (dict): The activation config for FFNs.
            Defaluts to ``dict(type='GELU')``.
        norm_cfg (dict): Config dict for normalization layer.
            Defaults to ``dict(type='LN')``.
    """

    def __init__(
        self,
        embed_dims: int,
        num_heads: int,
        feedforward_channels: int,
        num_fcs: int = 2,
        qkv_bias: bool = False,
        qk_scale: float = None,
        drop_rate: float = 0.,
        attn_drop_rate: float = 0.,
        drop_path_rate: float = 0.,
        layer_scale_init_value: float = 0.0,
        act_cfg: dict = dict(type='GELU'),
        norm_cfg: dict = dict(type='LN', eps=1e-6)
    ) -> None:
        super().__init__()

        # NOTE: cross attention
        _, self.norm1_q_cross = build_norm_layer(
            norm_cfg, embed_dims, postfix=2)
        _, self.norm1_k_cross = build_norm_layer(
            norm_cfg, embed_dims, postfix=2)
        _, self.norm1_v_cross = build_norm_layer(
            norm_cfg, embed_dims, postfix=2)
        _, self.norm2_cross = build_norm_layer(norm_cfg, embed_dims, postfix=2)
        self.cross_attn = CrossMultiheadAttention(
            embed_dims,
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            qk_scale=qk_scale,
            attn_drop=attn_drop_rate,
            proj_drop=drop_rate)

        self.ffn = FFN(
            embed_dims=embed_dims,
            feedforward_channels=feedforward_channels,
            num_fcs=num_fcs,
            ffn_drop=drop_rate,
            dropout_layer=None,
            act_cfg=act_cfg,
            add_identity=False)

        self.drop_path = DropPath(drop_prob=drop_path_rate)

        if layer_scale_init_value > 0:
            self.gamma_1_cross = nn.Parameter(
                layer_scale_init_value * torch.ones((embed_dims)),
                requires_grad=True)
            self.gamma_2_cross = nn.Parameter(
                layer_scale_init_value * torch.ones((embed_dims)),
                requires_grad=True)
        else:
            self.gamma_1_cross = nn.Parameter(
                torch.ones((embed_dims)), requires_grad=False)
            self.gamma_2_cross = nn.Parameter(
                torch.ones((embed_dims)), requires_grad=False)

    def forward(self, x_q: torch.Tensor, x_kv: torch.Tensor,
                pos_q: torch.Tensor, pos_k: torch.Tensor) -> torch.Tensor:
        """Forward function."""
        x = x_q + self.drop_path(self.gamma_1_cross * self.cross_attn(
            self.norm1_q_cross(x_q + pos_q),
            k=self.norm1_k_cross(x_kv + pos_k),
            v=self.norm1_v_cross(x_kv)))
        x = self.norm2_cross(x)
        x = x + self.drop_path(self.gamma_2_cross * self.ffn(x))

        return x


@MODELS.register_module()
class CAENeck(BaseModule):
    """Neck for CAE Pre-training.

    This module construct the latent prediction regressor and the decoder
    for the latent prediction and final prediction.

    Args:
        num_classes (int): The number of classes for final prediction. Defaults
            to 8192.
        embed_dims (int): The embed dims of latent feature in regressor and
            decoder. Defaults to 768.
        regressor_depth (int): The number of regressor blocks. Defaults to 6.
        decoder_depth (int): The number of decoder blocks. Defaults to 8.
        num_heads (int): The number of head in multi-head attention. Defaults
            to 12.
        mlp_ratio (int): The expand ratio of latent features in MLP. defaults
            to 4.
        qkv_bias (bool): Whether or not to use qkv bias. Defaults to True.
        qk_scale (float, optional): The scale applied to the results of qk.
            Defaults to None.
        drop_rate (float): The dropout rate. Defaults to 0.
        attn_drop_rate (float): The dropout rate in attention block. Defaults
            to 0.
        norm_cfg (dict): The config of normalization layer. Defaults to
            dict(type='LN', eps=1e-6).
        layer_scale_init_value (float, optional): The init value of gamma.
            Defaults to None.
        mask_tokens_num (int): The number of mask tokens. Defaults to 75.
        init_cfg (dict, optional): Initialization config dict.
            Defaults to None.
    """

    def __init__(self,
                 num_classes: int = 8192,
                 embed_dims: int = 768,
                 regressor_depth: int = 6,
                 decoder_depth: int = 8,
                 num_heads: int = 12,
                 mlp_ratio: int = 4,
                 qkv_bias: bool = True,
                 qk_scale: float = None,
                 drop_rate: float = 0.,
                 attn_drop_rate: float = 0.,
                 drop_path_rate: float = 0.,
                 norm_cfg: dict = dict(type='LN', eps=1e-6),
                 layer_scale_init_value: float = None,
                 mask_tokens_num: int = 75,
                 init_cfg: dict = None) -> None:
        super().__init__(init_cfg=init_cfg)

        self.num_features = self.embed_dim = embed_dims
        self.mask_token_num = mask_tokens_num

        # regressor
        regressor_drop_path_rates = [
            x.item()
            for x in torch.linspace(0, drop_path_rate, regressor_depth)
        ]
        self.regressors = nn.ModuleList([
            CAETransformerRegressorLayer(
                embed_dims=embed_dims,
                num_heads=num_heads,
                feedforward_channels=mlp_ratio * embed_dims,
                qkv_bias=qkv_bias,
                qk_scale=qk_scale,
                drop_rate=drop_rate,
                attn_drop_rate=attn_drop_rate,
                drop_path_rate=regressor_drop_path_rates[i],
                norm_cfg=norm_cfg,
                layer_scale_init_value=layer_scale_init_value)
            for i in range(regressor_depth)
        ])

        # decoder
        decoder_drop_path_rates = [
            x.item() for x in torch.linspace(0, drop_path_rate, decoder_depth)
        ]
        self.decoders = nn.ModuleList([
            BEiTTransformerEncoderLayer(
                embed_dims=embed_dims,
                num_heads=num_heads,
                feedforward_channels=mlp_ratio * embed_dims,
                layer_scale_init_value=layer_scale_init_value,
                window_size=None,
                # setting `use_rel_pos_bias` to False ignores the `window_size`
                use_rel_pos_bias=False,
                drop_rate=drop_rate,
                attn_drop_rate=attn_drop_rate,
                drop_path_rate=decoder_drop_path_rates[i],
                norm_cfg=norm_cfg) for i in range(decoder_depth)
        ])

        _, self.norm_regressor = build_norm_layer(
            norm_cfg, embed_dims, postfix=2)
        _, self.norm_decoder = build_norm_layer(
            norm_cfg, embed_dims, postfix=2)

        self.head = nn.Linear(
            embed_dims, num_classes) if num_classes > 0 else nn.Identity()
        self.mask_token = nn.Parameter(torch.zeros(1, 1, embed_dims))

    def init_weights(self) -> None:
        """Initialization."""
        super().init_weights()
        self.apply(self._init_weights)
        trunc_normal_(self.mask_token, std=0.02)
        trunc_normal_(self.head.weight, std=0.02)

    def _init_weights(self, m: nn.Module) -> None:
        """Initialization."""
        if isinstance(m, nn.Linear):
            nn.init.xavier_uniform_(m.weight)
            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 forward(
            self, x_unmasked: torch.Tensor, pos_embed_masked: torch.Tensor,
            pos_embed_unmasked: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Get the latent prediction and final prediction.

        Args:
            x_unmasked (torch.Tensor): Features of unmasked tokens.
            pos_embed_masked (torch.Tensor): Position embedding of masked
                tokens.
            pos_embed_unmasked (torch.Tensor): Position embedding of unmasked
                tokens.

        Returns:
            Tuple[torch.Tensor, torch.Tensor]:
              - ``logits``: Final prediction.
              - ``latent_pred``: Latent prediction.
        """
        x_masked = self.mask_token.expand(x_unmasked.shape[0],
                                          self.mask_token_num, -1)
        # regressor
        for regressor in self.regressors:
            x_masked = regressor(
                x_masked, torch.cat([x_unmasked, x_masked], dim=1),
                pos_embed_masked,
                torch.cat([pos_embed_unmasked, pos_embed_masked], dim=1))
        x_masked = self.norm_regressor(x_masked)
        latent_pred = x_masked

        # decoder
        x_masked = x_masked + pos_embed_masked
        for decoder in self.decoders:
            x_masked = decoder(x_masked, rel_pos_bias=None)
        x_masked = self.norm_decoder(x_masked)

        logits = self.head(x_masked)

        return logits, latent_pred