models_sim.py

# ------------------------------------------------------------------------
# SiameseIM
# Copyright (c) SenseTime. All Rights Reserved.
# ------------------------------------------------------------------------
# Modified from MAE (https://github.com/facebookresearch/mae)
# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved.
# ------------------------------------------------------------------------
# References:
# timm: https://github.com/rwightman/pytorch-image-models/tree/master/timm
# DeiT: https://github.com/facebookresearch/deit
# ------------------------------------------------------------------------


import random
from functools import partial
from turtle import update
import math

import torch
import torch.nn as nn

from util.pos_embed import get_2d_sincos_pos_embed, get_2d_sincos_pos_embed_relative
from util.misc import LayerNorm
from models_vit import Block, CrossBlock, PatchEmbed 


class PermuteBN(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.bn = nn.BatchNorm1d(dim)

    @torch.cuda.amp.autocast(enabled=False)
    def forward(self, x):
        x = x.permute(0, 2, 1) # N, L, C -> N, C, L
        x = x.float()
        x = self.bn(x)
        x = x.permute(0, 2, 1) # N, C, L -> N, L, C

        return x


class SiameseIMViT(nn.Module):
    """  SiameseIM with VisionTransformer backbone
    """
    def __init__(self, img_size=224, patch_size=16, in_chans=3,
                 embed_dim=1024, depth=24, num_heads=16,
                 decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16,
                 mlp_ratio=4., norm_layer=LayerNorm, norm_pix_loss=False, args=None):
        super().__init__()
        self.norm_pix_loss = norm_pix_loss
        self.args = args
        decoder_embed_dim = args.decoder_embed_dim

        # --------------------------------------------------------------------------
        # encoder specifics
        self.patch_embed = PatchEmbed(img_size, patch_size, in_chans, embed_dim)
        num_patches = self.patch_embed.num_patches
        self.num_patches = num_patches

        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
        if args.use_abs_pos_emb:
            if hasattr(self, 'cls_token'):
                self.pos_embed = nn.Parameter(torch.zeros(1, num_patches+1, embed_dim), requires_grad=False)  # fixed sin-cos embedding
            else:
                self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim), requires_grad=False)  # fixed sin-cos embedding

        dpr = [x.item() for x in torch.linspace(0, args.drop_path_rate, depth)]
        self.blocks = nn.ModuleList([
            Block(embed_dim, num_heads, mlp_ratio, qkv_bias=True, norm_layer=norm_layer,
                    drop_path=dpr[i], init_values=args.init_values)
            for i in range(depth)])
        # --------------------------------------------------------------------------

        # --------------------------------------------------------------------------
        # decoder specifics
        if args.loss_type in ['mae']:
            self.decoder_embed = nn.Linear(embed_dim, decoder_embed_dim, bias=True)

            self.mask_token = nn.Parameter(torch.zeros(1, 1, decoder_embed_dim))

            if hasattr(self, 'cls_token'):
                self.decoder_pos_embed = nn.Parameter(torch.zeros(1, num_patches+1, decoder_embed_dim), requires_grad=False)  # fixed sin-cos embedding
            else:
                self.decoder_pos_embed = nn.Parameter(torch.zeros(1, num_patches, decoder_embed_dim), requires_grad=False)

            self.decoder_blocks = nn.ModuleList([
                Block(decoder_embed_dim, decoder_num_heads, mlp_ratio, qkv_bias=True, norm_layer=norm_layer)
                for i in range(decoder_depth)])

            self.decoder_norm = norm_layer(decoder_embed_dim)
            self.decoder_pred = nn.Linear(decoder_embed_dim, patch_size**2 * in_chans, bias=True) # decoder to patch
        elif args.loss_type in ['sim',]:
            self.decoder_embed = nn.Linear(embed_dim, decoder_embed_dim, bias=True)

            self.mask_token = nn.Parameter(torch.zeros(1, 1, decoder_embed_dim))

            if hasattr(self, 'cls_token'):
                self.decoder_pos_embed = nn.Parameter(torch.zeros(1, num_patches+1, decoder_embed_dim), requires_grad=False)  # fixed sin-cos embedding
            else:
                self.decoder_pos_embed = nn.Parameter(torch.zeros(1, num_patches, decoder_embed_dim), requires_grad=False)

            if args.projector_depth > 0:
                self.projector_decoder_blocks = nn.ModuleList([
                    Block(decoder_embed_dim, decoder_num_heads, mlp_ratio, qkv_bias=True,
                          norm_layer=norm_layer if args.use_proj_ln else PermuteBN)
                    for i in range(args.projector_depth)])

            self.predictor_decoder_blocks = nn.ModuleList([
                Block(decoder_embed_dim, decoder_num_heads, mlp_ratio, qkv_bias=True,
                      norm_layer=norm_layer if args.use_pred_ln else PermuteBN)
                for i in range(args.predictor_depth)])

            self.decoder_pred = nn.Linear(decoder_embed_dim, decoder_embed_dim, bias=True) # decoder to patch
            if args.online_ln:
                self.student_norm = LayerNorm(decoder_embed_dim)
                for p in self.student_norm.parameters():
                    p.requires_grad = False
            else:
                self.student_norm = nn.Identity()
        # --------------------------------------------------------------------------
    
        # ---------------------------------------------------------------------------
        # decoder pos embed change dim
        if self.args.loss_type in ['sim',]:
            self.decoder_pos_mlp = nn.Linear(decoder_embed_dim*2, decoder_embed_dim)
        # ---------------------------------------------------------------------------

        self.initialize_weights()

        # build momentum branch
        if self.args.loss_type in ['sim',]:
            self.build_momentum_target(img_size, patch_size, in_chans, embed_dim, num_heads,
                                        mlp_ratio, norm_layer, depth, decoder_embed_dim, decoder_num_heads)

        # stop grad for patch embedding
        if (not args.train_patch_embed):
            self.patch_embed.proj.weight.requires_grad = False
            self.patch_embed.proj.bias.requires_grad = False

    def build_momentum_target(self, img_size, patch_size, in_chans, embed_dim, num_heads,
                                mlp_ratio, norm_layer, depth, decoder_embed_dim, decoder_num_heads):
        # --------------------------------------------------------------------------
        # momentum encoder specifics
        self.mm_patch_embed = PatchEmbed(img_size, patch_size, in_chans, embed_dim)

        if hasattr(self, 'cls_token'):
            self.mm_cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))

        self.mm_blocks = nn.ModuleList([
            Block(embed_dim, num_heads, mlp_ratio, qkv_bias=True, norm_layer=norm_layer, init_values=self.args.init_values)
            for i in range(depth)])
        
        # load weight
        self.mm_patch_embed.load_state_dict(self.patch_embed.state_dict())
        for p in self.mm_patch_embed.parameters():
            p.requires_grad = False

        self.mm_cls_token.data.copy_(self.cls_token.data)
        self.mm_cls_token.requires_grad = False

        self.mm_blocks.load_state_dict(self.blocks.state_dict())
        for p in self.mm_blocks.parameters():
            p.requires_grad = False
        # --------------------------------------------------------------------------
 
        # --------------------------------------------------------------------------
        # momentum decoder specifics
        self.mm_decoder_embed = nn.Linear(embed_dim, decoder_embed_dim, bias=True)

        self.mm_mask_token = nn.Parameter(torch.zeros(1, 1, decoder_embed_dim))

        if self.args.projector_depth > 0:
            self.mm_projector_decoder_blocks = nn.ModuleList([
                Block(decoder_embed_dim, decoder_num_heads, mlp_ratio, qkv_bias=True, norm_layer=norm_layer if self.args.use_proj_ln else PermuteBN)
                for i in range(self.args.projector_depth)])

        # load weight
        self.mm_decoder_embed.load_state_dict(self.decoder_embed.state_dict())
        for p in self.mm_decoder_embed.parameters():
            p.requires_grad = False
        
        self.mm_mask_token.data.copy_(self.mask_token.data)
        self.mm_mask_token.requires_grad = False
        
        if self.args.projector_depth > 0:
            self.mm_projector_decoder_blocks.load_state_dict(self.projector_decoder_blocks.state_dict())
            for p in self.mm_projector_decoder_blocks.parameters():
                p.requires_grad = False
        # ---------------------------------------------------------------------------

        if self.args.loss_type in ['sim',]:
            self.teacher_norm = LayerNorm(decoder_embed_dim, elementwise_affine=False)
            for p in self.teacher_norm.parameters():
                p.requires_grad = False

    def initialize_weights(self):
        # initialization
        # initialize (and freeze) pos_embed by sin-cos embedding
        if self.args.use_abs_pos_emb:
            pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.num_patches**.5), cls_token=hasattr(self, 'cls_token'))
            self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))

        if hasattr(self, 'decoder_pos_embed'):
            decoder_pos_embed = get_2d_sincos_pos_embed(self.decoder_pos_embed.shape[-1], int(self.num_patches**.5), cls_token=hasattr(self, 'cls_token'))
            self.decoder_pos_embed.data.copy_(torch.from_numpy(decoder_pos_embed).float().unsqueeze(0))

        # initialize patch_embed like nn.Linear (instead of nn.Conv2d)
        w = self.patch_embed.proj.weight.data
        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))

        # timm's trunc_normal_(std=.02) is effectively normal_(std=0.02) as cutoff is too big (2.)
        if hasattr(self, 'cls_token'):
            torch.nn.init.normal_(self.cls_token, std=.02)
        if hasattr(self, 'mask_token'):
            torch.nn.init.normal_(self.mask_token, std=.02)

        # initialize nn.Linear and LayerNorm
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            # we use xavier_uniform following official JAX ViT:
            torch.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 patchify(self, imgs):
        """
        imgs: (N, 3, H, W)
        x: (N, L, patch_size**2 *3)
        """
        p = self.patch_embed.patch_size[0]
        assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0

        h = w = imgs.shape[2] // p
        x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))
        x = torch.einsum('nchpwq->nhwpqc', x)
        x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3))
        return x

    def unpatchify(self, x):
        """
        x: (N, L, patch_size**2 *3)
        imgs: (N, 3, H, W)
        """
        p = self.patch_embed.patch_size[0]
        h = w = int(x.shape[1]**.5)
        assert h * w == x.shape[1]
        
        x = x.reshape(shape=(x.shape[0], h, w, p, p, 3))
        x = torch.einsum('nhwpqc->nchpwq', x)
        imgs = x.reshape(shape=(x.shape[0], 3, h * p, h * p))
        return imgs

    def random_masking(self, x, mask_ratio):
        """
        Perform per-sample random masking by per-sample shuffling.
        Per-sample shuffling is done by argsort random noise.
        x: [N, L, D], sequence
        """
        N, L, D = x.shape  # batch, length, dim
        len_keep = int(L * (1 - mask_ratio))
        
        noise = torch.rand(N, L, device=x.device)  # noise in [0, 1]
        
        # sort noise for each sample
        ids_shuffle = torch.argsort(noise, dim=1)  # ascend: small is keep, large is remove
        ids_restore = torch.argsort(ids_shuffle, dim=1)

        # keep the first subset
        # ids_keep = ids_shuffle[:, :len_keep]
        # x_masked = torch.gather(x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))

        # generate the binary mask: 0 is keep, 1 is remove
        mask = torch.ones([N, L], device=x.device)
        mask[:, :len_keep] = 0
        # unshuffle to get the binary mask
        mask = torch.gather(mask, dim=1, index=ids_restore)

        return mask, ids_restore

    @torch.cuda.amp.autocast(enabled=False)
    def mm_update(self, mm):
        for param_q, param_k in zip(self.patch_embed.parameters(), self.mm_patch_embed.parameters()):
            param_k.data = param_k.data * mm + param_q.data * (1. - mm)
        for param_q, param_k in zip(self.blocks.parameters(), self.mm_blocks.parameters()):
            param_k.data = param_k.data * mm + param_q.data * (1. - mm)
        if hasattr(self, 'mm_cls_token'):
            self.mm_cls_token.data = self.mm_cls_token.data * mm + self.cls_token.data * (1. - mm)
        if hasattr(self, 'mm_norm'):
            for param_q, param_k in zip(self.norm.parameters(), self.mm_norm.parameters()):
                param_k.data = param_k.data * mm + param_q.data * (1. - mm)
        if hasattr(self, 'mm_projector'):
            for param_q, param_k in zip(self.projector.parameters(), self.mm_projector.parameters()):
                param_k.data = param_k.data * mm + param_q.data * (1. - mm)
        if hasattr(self, 'mm_decoder_embed'):
            for param_q, param_k in zip(self.decoder_embed.parameters(), self.mm_decoder_embed.parameters()):
                param_k.data = param_k.data * mm + param_q.data * (1. - mm)
        if hasattr(self, 'mm_mask_token'):
            self.mm_mask_token.data = self.mm_mask_token.data * mm + self.mask_token.data * (1. - mm)
        if hasattr(self, 'mm_decoder_blocks'):
            for param_q, param_k in zip(self.decoder_blocks.parameters(), self.mm_decoder_blocks.parameters()):
                param_k.data = param_k.data * mm + param_q.data * (1. - mm)
        if hasattr(self, 'mm_projector_decoder_blocks'):
            for param_q, param_k in zip(self.projector_decoder_blocks.parameters(), self.mm_projector_decoder_blocks.parameters()):
                param_k.data = param_k.data * mm + param_q.data * (1. - mm)
        if hasattr(self, 'mm_decoder_norm'):
            for param_q, param_k in zip(self.decoder_norm.parameters(), self.mm_decoder_norm.parameters()):
                param_k.data = param_k.data * mm + param_q.data * (1. - mm)
        if hasattr(self, 'mm_decoder_pred'):
            for param_q, param_k in zip(self.decoder_pred.parameters(), self.mm_decoder_pred.parameters()):
                param_k.data = param_k.data * mm + param_q.data * (1. - mm)

    def forward_encoder(self, x, mask_ratio):
        # embed patches
        x = self.patch_embed(x)

        # add pos embed w/o cls token
        x = x + self.pos_embed

        # masking: length -> length * mask_ratio
        mask, ids_restore = self.random_masking(x, mask_ratio)
        x = x[~mask.bool()].view(x.shape[0], -1, x.shape[-1])

        # apply Transformer blocks
        for blk in self.blocks:
            x = blk(x)
        x = self.norm(x)

        return x, mask, ids_restore

    def forward_decoder(self, x, ids_restore):
        # embed tokens
        x = self.decoder_embed(x)

        # append mask tokens to sequence
        mask_tokens = self.mask_token.repeat(x.shape[0], ids_restore.shape[1] + 1 - x.shape[1], 1)
        x_ = torch.cat([x, mask_tokens], dim=1)  # no cls token
        x_ = torch.gather(x_, dim=1, index=ids_restore.unsqueeze(-1).repeat(1, 1, x.shape[2]))  # unshuffle
        x = x_

        # add pos embed
        x = x + self.decoder_pos_embed

        # apply Transformer blocks
        for blk in self.decoder_blocks:
            x = blk(x)
        x = self.decoder_norm(x)

        # predictor projection
        x = self.decoder_pred(x)

        return x

    def forward_loss(self, imgs, pred, mask):
        """
        imgs: [N, 3, H, W]
        pred: [N, L, p*p*3]
        mask: [N, L], 0 is keep, 1 is remove, 
        """
        target = self.patchify(imgs)
        if self.norm_pix_loss:
            mean = target.mean(dim=-1, keepdim=True)
            var = target.var(dim=-1, keepdim=True)
            target = (target - mean) / (var + 1.e-6)**.5

        loss = (pred - target) ** 2
        loss = loss.mean(dim=-1)  # [N, L], mean loss per patch

        loss = (loss * mask).sum() / mask.sum()  # mean loss on removed patches
        return loss

    def forward_mae(self, imgs, mask_ratio=0.75):
        latent, mask, ids_restore = self.forward_encoder(imgs, mask_ratio)
        pred = self.forward_decoder(latent, ids_restore)  # [N, L, p*p*3]
        loss = self.forward_loss(imgs, pred, mask)
        return loss, pred, mask
    
    def forward(self, *args, **kwargs):
        if self.args.loss_type in ['sim',]:
            return self.forward_sim(*args, **kwargs)
        else:
            return self.forward_mae(*args, **kwargs)

    def forward_sim(self, x1, x2, rel_pos_21, mm, update_mm, mask=None):
        # forward online encoder
        if self.args.with_blockwise_mask:
            assert mask is not None, 'mask should not be None when mask_type is block'
            mask = mask.view(mask.shape[0], -1)
        else:
            assert False
            mask, ids_restore1 = self.random_masking(online_x1, self.args.mask_ratio)
        
        online_x1 = self.patch_embed(x1)
        online_x1 = online_x1 + self.pos_embed[:, 1:, :]
        online_x1 = online_x1[~mask.bool()].view(online_x1.shape[0], -1, online_x1.shape[-1])
        # add cls token
        cls_tokens = self.cls_token.expand(online_x1.shape[0], -1, -1) + self.pos_embed[:, 0, :].unsqueeze(1)
        online_x1 = torch.cat((cls_tokens, online_x1), dim=1)

        # forward online encoder
        for blk in self.blocks:
            online_x1 = blk(online_x1)

        # forward online projector
        online_x1 = self.decoder_embed(online_x1)
        if self.args.projector_depth > 0:
            for blk in self.projector_decoder_blocks:
                online_x1 = blk(online_x1)

        # calculate decoder pos embed
        cls_pos_embed = self.decoder_pos_embed[:, 0, :].unsqueeze(1)
        x1_vis_embed = self.decoder_pos_embed[:, 1:, :].repeat(online_x1.shape[0], 1, 1)[~mask.bool()].view(online_x1.shape[0], -1, self.decoder_pos_embed.shape[-1])
        x2_embed = get_2d_sincos_pos_embed_relative(*rel_pos_21, self.decoder_pos_embed.shape[-1],
                                                    int(self.num_patches ** .5))
        x2_embed = self.decoder_pos_mlp(x2_embed)

        # append mask tokens to sequence
        cls_token = online_x1[:, 0, :].unsqueeze(1)
        x1_vis_tokens = online_x1[:, 1:, :]
        mask_tokens = self.mask_token.repeat(x2.shape[0], x2_embed.shape[1], 1)
        x = torch.cat([cls_token+cls_pos_embed, x1_vis_tokens+x1_vis_embed, mask_tokens+x2_embed], dim=1)

        # forward online decoder
        for blk in self.predictor_decoder_blocks:
            x = blk(x)

        # predictor projection
        x = self.decoder_pred(x)
        pred = x[:, -x2_embed.shape[1]:]

        # forward target encoder
        with torch.no_grad():
            if update_mm:
                self.mm_update(mm)

            target_x2 = self.mm_patch_embed(x2)
            mm_cls_tokens = self.mm_cls_token.expand(target_x2.shape[0], -1, -1)
            target_x2 = torch.cat((mm_cls_tokens, target_x2), dim=1)
            target_x2 = target_x2 + self.pos_embed

            # forward target encoder
            for blk in self.mm_blocks:
                target_x2 = blk(target_x2)

            # forward target projector
            target_x2 = self.mm_decoder_embed(target_x2)
            if self.args.projector_depth > 0:
                for blk in self.mm_projector_decoder_blocks:
                    target_x2 = blk(target_x2)

            target = target_x2[:, 1:, :]

        # compute loss
        outputs = {}
        with torch.cuda.amp.autocast(enabled=False):
            loss = self.compute_unigrad_loss(pred.float(), target.float())
        outputs['loss_sim'] = loss.item()

        return loss, outputs

    def compute_unigrad_loss(self, pred, target):
        pred = self.student_norm(pred)
        with torch.no_grad():
            target = self.teacher_norm(target)
        
        dense_pred = pred.reshape(-1, pred.shape[-1])
        dense_target = target.reshape(-1, target.shape[-1])

        # compute pos term
        pos_term = ((dense_pred - dense_target)**2).sum(-1).mean()

        # compute neg term
        correlation = (dense_target.T @ dense_target) / dense_target.shape[0]
        torch.distributed.all_reduce(correlation)
        correlation = correlation / torch.distributed.get_world_size()
        
        neg_term = torch.diagonal(dense_pred @ correlation @ dense_pred.T).mean()

        loss = (pos_term + self.args.neg_weight * neg_term) / pred.shape[-1]

        return loss


def sim_vit_base_patch16_dec512d8b(**kwargs):
    model = SiameseIMViT(
        patch_size=16, embed_dim=768, depth=12, num_heads=12,
        decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16,
        mlp_ratio=4, norm_layer=partial(LayerNorm, eps=1e-6), **kwargs)
    return model


def sim_vit_large_patch16_dec512d8b(**kwargs):
    model = SiameseIMViT(
        patch_size=16, embed_dim=1024, depth=24, num_heads=16,
        decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16,
        mlp_ratio=4, norm_layer=partial(LayerNorm, eps=1e-6), **kwargs)
    return model


def sim_vit_huge_patch14_dec512d8b(**kwargs):
    model = SiameseIMViT(
        patch_size=14, embed_dim=1280, depth=32, num_heads=16,
        decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16,
        mlp_ratio=4, norm_layer=partial(LayerNorm, eps=1e-6), **kwargs)
    return model


# set recommended archs
sim_vit_base_patch16 = sim_vit_base_patch16_dec512d8b  # decoder: 512 dim, 8 blocks
sim_vit_large_patch16 = sim_vit_large_patch16_dec512d8b  # decoder: 512 dim, 8 blocks
sim_vit_huge_patch14 = sim_vit_huge_patch14_dec512d8b  # decoder: 512 dim, 8 blocks