models_mae_temporal.py

# --------------------------------------------------------
# References:
# MAE: https://github.com/facebookresearch/mae
# timm: https://github.com/rwightman/pytorch-image-models/tree/master/timm
# DeiT: https://github.com/facebookresearch/deit
# --------------------------------------------------------

from functools import partial

import torch
import torch.nn as nn
from torchvision.utils import save_image

from timm.models.vision_transformer import PatchEmbed, Block

from util.pos_embed import get_2d_sincos_pos_embed, get_1d_sincos_pos_embed_from_grid_torch


class MaskedAutoencoderViT(nn.Module):
    """ Masked Autoencoder 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=nn.LayerNorm, norm_pix_loss=False, same_mask=False):
        super().__init__()

        # --------------------------------------------------------------------------
        # MAE encoder specifics
        self.patch_embed = PatchEmbed(img_size, patch_size, in_chans, embed_dim)
        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 - 384), requires_grad=False)  # fixed sin-cos embedding

        self.blocks = nn.ModuleList([
            Block(embed_dim, num_heads, mlp_ratio, qkv_bias=True, qk_scale=None, norm_layer=norm_layer)
            for i in range(depth)])
        self.norm = norm_layer(embed_dim)
        # --------------------------------------------------------------------------

        # --------------------------------------------------------------------------
        # MAE decoder specifics
        self.decoder_embed = nn.Linear(embed_dim, decoder_embed_dim, bias=True)

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

        self.decoder_pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, decoder_embed_dim - 192), requires_grad=False)  # fixed sin-cos embedding

        self.decoder_blocks = nn.ModuleList([
            Block(decoder_embed_dim, decoder_num_heads, mlp_ratio, qkv_bias=True, qk_scale=None, 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
        # --------------------------------------------------------------------------

        self.norm_pix_loss = norm_pix_loss

        self.same_mask = same_mask
        self.initialize_weights()
        self.counter = 0

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

        decoder_pos_embed = get_2d_sincos_pos_embed(self.decoder_pos_embed.shape[-1], int(self.patch_embed.num_patches**.5), cls_token=True)
        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.)
        torch.nn.init.normal_(self.cls_token, std=.02)
        torch.nn.init.normal_(self.mask_token, std=.02)

        # initialize nn.Linear and nn.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, mask=None):
        """
        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]
        
        if self.same_mask:
            L2 = L // 3
            assert 3 * L2 == L
            noise = torch.rand(N, L2, device=x.device)  # noise in [0, 1]
            ids_shuffle = torch.argsort(noise, dim=1)
            ids_shuffle = [ids_shuffle + i * L2 for i in range(3)]
            ids_shuffle_keep = [z[: ,:int(L2 * (1 - mask_ratio))] for z in ids_shuffle]
            ids_shuffle_disc = [z[: ,int(L2 * (1 - mask_ratio)):] for z in ids_shuffle]
            ids_shuffle = []
            for z in ids_shuffle_keep:
                ids_shuffle.append(z)
            for z in ids_shuffle_disc:
                ids_shuffle.append(z)
            ids_shuffle = torch.cat(ids_shuffle, dim=1)
            # print(ids_shuffle[0])
            # assert False
        else:
            if mask is None:
                # sort noise for each sample
                ids_shuffle = torch.argsort(noise, dim=1)  # ascend: small is keep, large is remove
            else:
                ids_shuffle = mask
        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 x_masked, mask, ids_restore

    def forward_encoder(self, x, timestamps, mask_ratio, mask=None):
        # embed patches
        x1 = self.patch_embed(x[:, 0])
        x2 = self.patch_embed(x[:, 1])
        x3 = self.patch_embed(x[:, 2])
        x = torch.cat([x1, x2, x3], dim=1)
        

        # print(timestamps.shape, x.shape)
        ts_embed = torch.cat([get_1d_sincos_pos_embed_from_grid_torch(128, timestamps.reshape(-1, 3)[:, 0].float()),
                   get_1d_sincos_pos_embed_from_grid_torch(128, timestamps.reshape(-1, 3)[:, 1].float()),
                   get_1d_sincos_pos_embed_from_grid_torch(128, timestamps.reshape(-1, 3)[:, 2].float())], dim=1).float()
        
        # print(ts_embed, ts_embed.shape)
        
        ts_embed = ts_embed.reshape(-1, 3, ts_embed.shape[-1]).unsqueeze(2)
        # print(ts_embed.shape)
        ts_embed = ts_embed.expand(-1, -1, x.shape[1] // 3, -1).reshape(x.shape[0], -1, ts_embed.shape[-1])
        # print(ts_embed.shape)
        # ts_embed = torch.zeros_like(ts_embed)


        # add pos embed w/o cls token
        x = x + torch.cat([self.pos_embed[:, 1:, :].repeat(ts_embed.shape[0], 3, 1), ts_embed], dim=-1)

        # masking: length -> length * mask_ratio
        x, mask, ids_restore = self.random_masking(x, mask_ratio, mask=mask)

        # append cls token
        cls_token = self.cls_token #+ self.pos_embed[:, :1, :]
        cls_tokens = cls_token.expand(x.shape[0], -1, -1)
        x = torch.cat((cls_tokens, x), dim=1)
        # print(x.dtype)

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

        return x, mask, ids_restore

    def forward_decoder(self, x, timestamps, 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[:, 1:, :], 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 = torch.cat([x[:, :1, :], x_], dim=1)  # append cls token


        ts_embed = torch.cat([get_1d_sincos_pos_embed_from_grid_torch(64, timestamps.reshape(-1, 3)[:, 0].float()),
                   get_1d_sincos_pos_embed_from_grid_torch(64, timestamps.reshape(-1, 3)[:, 1].float()),
                   get_1d_sincos_pos_embed_from_grid_torch(64, timestamps.reshape(-1, 3)[:, 2].float())], dim=1).float()
        
        ts_embed = ts_embed.reshape(-1, 3, ts_embed.shape[-1]).unsqueeze(2)
        ts_embed = ts_embed.expand(-1, -1, x.shape[1] // 3, -1).reshape(x.shape[0], -1, ts_embed.shape[-1])

        ts_embed = torch.cat([torch.zeros((ts_embed.shape[0], 1, ts_embed.shape[2]), device=ts_embed.device), ts_embed], dim=1)

        # ts_embed = torch.zeros_like(ts_embed)

        # add pos embed
        x = x + torch.cat(
            [torch.cat([self.decoder_pos_embed[:, :1, :], self.decoder_pos_embed[:, 1:, :].repeat(1, 3, 1)], dim=1).expand(ts_embed.shape[0], -1, -1),
             ts_embed], dim=-1)

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

        # predictor projection
        x = self.decoder_pred(x)

        # remove cls token
        x = x[:, 1:, :]

        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, 
        """
        target1 = self.patchify(imgs[:, 0])
        target2 = self.patchify(imgs[:, 1])
        target3 = self.patchify(imgs[:, 2])
        target = torch.cat([target1, target2, target3], dim=1)
        previous_target = target
        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

        # viz code
        '''
        m = torch.tensor([0.4182007312774658, 0.4214799106121063, 0.3991275727748871]).reshape(1, 3, 1, 1)
        std = torch.tensor([0.28774282336235046, 0.27541765570640564, 0.2764017581939697]).reshape(1, 3, 1, 1)
        
        image = (pred * (var + 1.e-6)**.5) + mean
        bs = image.shape[0]
        image = image.reshape(bs, 3, -1, image.shape[-1])[0]
        image = self.unpatchify(image).detach().cpu()
        image = image * std + m

        save_image(image, f'viz1/viz_{self.counter}.png')
        masked_image = self.patchify(image)
        masked_image.reshape(-1, 768)[mask[0].bool()] = 0.5
        masked_image = self.unpatchify(masked_image.reshape(3, -1 ,768))
        save_image(masked_image, f'viz1/viz_mask_{self.counter}.png')

        previous_target = previous_target.reshape(bs, 3, -1, previous_target.shape[-1])[0]
        previous_target = self.unpatchify(previous_target).detach().cpu()
        previous_target = previous_target * std + m
        save_image(previous_target, f'viz1/target_{self.counter}.png')

        masked_image = self.patchify(previous_target)
        masked_image.reshape(-1, 768)[mask[0].bool()] = 0.5
        masked_image = self.unpatchify(masked_image.reshape(3, -1 ,768))
        save_image(masked_image, f'viz1/viz_target_mask_{self.counter}.png')
        # print(image.shape)
        # assert False
        self.counter += 1
        '''

        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(self, imgs, timestamps, mask_ratio=0.75, mask=None):
        latent, mask, ids_restore = self.forward_encoder(imgs, timestamps, mask_ratio, mask=mask)
        pred = self.forward_decoder(latent, timestamps, ids_restore)  # [N, L, p*p*3]
        loss = self.forward_loss(imgs, pred, mask)
        return loss, pred, mask


def mae_vit_base_patch16_dec512d8b(**kwargs):
    model = MaskedAutoencoderViT(
        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(nn.LayerNorm, eps=1e-6), **kwargs)
    return model


def mae_vit_large_patch16_dec512d8b(**kwargs):
    model = MaskedAutoencoderViT(
        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(nn.LayerNorm, eps=1e-6), **kwargs)
    return model

def mae_vit_large_patch16_dec512d8b_samemask(**kwargs):
    model = MaskedAutoencoderViT(
        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(nn.LayerNorm, eps=1e-6), same_mask=True, **kwargs)
    return model


def mae_vit_huge_patch14_dec512d8b(**kwargs):
    model = MaskedAutoencoderViT(
        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(nn.LayerNorm, eps=1e-6), **kwargs)
    return model


# set recommended archs
mae_vit_base_patch16 = mae_vit_base_patch16_dec512d8b  # decoder: 512 dim, 8 blocks
mae_vit_large_patch16 = mae_vit_large_patch16_dec512d8b  # decoder: 512 dim, 8 blocks
mae_vit_large_patch16_samemask = mae_vit_large_patch16_dec512d8b_samemask
# from models_mae import mae_vit_large_patch16_dec512d8b
# mae_vit_large_patch16_nontemp = mae_vit_large_patch16_dec512d8b
mae_vit_huge_patch14 = mae_vit_huge_patch14_dec512d8b  # decoder: 512 dim, 8 blocks