slots/DINOSAUR.py

import warnings

import torch.nn as nn
import torch
import numpy as np
import math
import torch.nn.functional as F
from torch.nn import init
import random


# 残差块
class ResidualBlock(nn.Module):
    def __init__(self, channel):
        super(ResidualBlock, self).__init__()
        self.channel = channel
        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channels=channel,
                      out_channels=channel,
                      kernel_size=3,
                      stride=1,
                      padding=1),
            nn.BatchNorm2d(channel),
            nn.ReLU(inplace=True)
        )
        self.conv2 = nn.Sequential(
            nn.Conv2d(channel, channel, kernel_size=3, stride=1, padding=1),
            # nn.BatchNorm2d(channel)
        )

    def forward(self, x):
        out = self.conv1(x)
        out = self.conv2(out)
        out += x
        out = F.relu(out)
        return out

class ResNet(nn.Module):
    def __init__(self):
        super(ResNet, self).__init__()
        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channels=384, out_channels=448, kernel_size=5),  # (1,28,28)
            nn.BatchNorm2d(448),  # (32,24,24)
            nn.ReLU(),
            nn.MaxPool2d(2)  # (32,12,12)
        )
        self.conv2 = nn.Sequential(
            nn.Conv2d(in_channels=448, out_channels=640, kernel_size=5),  # (16,8,8)
            nn.BatchNorm2d(640),
            nn.ReLU(),
            nn.MaxPool2d(2)  # (16,4,4)
        )
        self.conv3 = nn.Sequential(
            nn.Conv2d(in_channels=640, out_channels=1024, kernel_size=5),  # (16,8,8)
            nn.BatchNorm2d(1024),
            nn.ReLU(),
            nn.MaxPool2d(2)  # (16,4,4)
        )
        self.reslayer1 = ResidualBlock(448)
        self.reslayer2 = ResidualBlock(640)
        self.reslayer3 = ResidualBlock(1024)


    def forward(self, x):
        out = self.conv1(x)
        out = self.reslayer1(out)

        out = self.conv2(out)
        out = self.reslayer2(out)

        out = self.conv3(out)
        out = self.reslayer3(out)

        return out


class Loss_Function(nn.Module):
    def __init__(self, args):
        super().__init__()
        self.mse = nn.MSELoss(reduction="none")

        self.token_num = args.model.SLOTS.token_num
        self.num_slots = args.model.SLOTS.num_slots

        self.epsilon = 1e-8

    def forward(self, reconstruction, masks, target):
        # :args reconstruction: (B, token, 1024)
        # :args masks: (B, S, token)
        # :args target: (B, token, 1024)

        target = target.detach()
        loss = self.mse(reconstruction, target.detach()).mean()

        return loss


class MLP(nn.Module):
    def __init__(self, input_dim, hidden_dim, output_dim, residual=False, layer_order="none"):
        super().__init__()
        self.residual = residual
        self.layer_order = layer_order
        if residual:
            assert input_dim == output_dim

        self.layer1 = nn.Linear(input_dim, hidden_dim)
        self.layer2 = nn.Linear(hidden_dim, output_dim)
        self.activation = nn.ReLU(inplace=True)
        self.dropout = nn.Dropout(p=0.1)

        if layer_order in ["pre", "post"]:
            self.norm = nn.LayerNorm(input_dim)
        else:
            assert layer_order == "none"

    def forward(self, x):
        input = x

        if self.layer_order == "pre":
            x = self.norm(x)

        x = self.layer1(x)
        x = self.activation(x)
        x = self.layer2(x)
        x = self.dropout(x)

        if self.residual:
            x = x + input
        if self.layer_order == "post":
            x = self.norm(x)

        return x


class Visual_Encoder(nn.Module):
    def __init__(self, args):
        super().__init__()

        self.resize_to = args.resize_to
        self.token_num = args.token_num

        self.encoder = args.encoder

        self.model = self.load_model(args)

    def load_model(self, args):
        assert args.resize_to[0] % args.patch_size == 0
        assert args.resize_to[1] % args.patch_size == 0

        if args.encoder == "dino-vitb-8":
            model = torch.hub.load("facebookresearch/dino:main", "dino_vitb8")
        elif args.encoder == "dino-vitb-16":
            model = torch.hub.load("facebookresearch/dino:main", "dino_vitb16")
        elif args.encoder == "dinov2-vitb-14":
            model = torch.hub.load("facebookresearch/dinov2", "dinov2_vitb14")
        elif args.encoder == "sup-vitb-16":
            model = timm.create_model("vit_base_patch16_224", pretrained=True,
                                      img_size=(args.resize_to[0], args.resize_to[1]))
        else:
            assert False

        for p in model.parameters():
            p.requires_grad = False

        # wget https://dl.fbaipublicfiles.com/dino/dino_vitbase8_pretrain/dino_vitbase8_pretrain.pth
        # wget https://dl.fbaipublicfiles.com/dino/dino_vitbase16_pretrain/dino_vitbase16_pretrain.pth
        # wget https://dl.fbaipublicfiles.com/dinov2/dinov2_vitb14/dinov2_vitb14_pretrain.pth

        return model

    @torch.no_grad()
    def forward(self, frames):
        # :arg frames:  (B, 3, H, W)
        #
        # :return x:  (B, token, 1024)

        B = frames.shape[0]

        self.model.eval()

        if self.encoder.startswith("dinov2-"):
            x = self.model.prepare_tokens_with_masks(frames)
        elif self.encoder.startswith("sup-"):
            x = self.model.patch_embed(frames)
            x = self.model._pos_embed(x)
        else:
            x = self.model.prepare_tokens(frames)

        for blk in self.model.blocks:
            x = blk(x)
        x = x[:, 1:]

        assert x.shape[0] == B
        assert x.shape[1] == self.token_num
        assert x.shape[2] == 1024

        return x


class Decoder(nn.Module):
    def __init__(self, args):
        super().__init__()

        # === Token calculations ===
        slot_dim = args['slot_dim']
        hidden_dim = 2048

        # === MLP Based Decoder ===
        self.layer1 = nn.Linear(slot_dim, hidden_dim)
        self.layer2 = nn.Linear(hidden_dim, hidden_dim)
        self.layer3 = nn.Linear(hidden_dim, hidden_dim)
        self.layer4 = nn.Linear(hidden_dim, 1024 + 1)
        self.relu = nn.ReLU(inplace=True)

    def forward(self, slot_maps):
        # :arg slot_maps: (B * S, token, D_slot)

        slot_maps = self.relu(self.layer1(slot_maps))  #  (B * S, token, D_hidden)
        slot_maps = self.relu(self.layer2(slot_maps))  #  (B * S, token, D_hidden)
        slot_maps = self.relu(self.layer3(slot_maps))  #  (B * S, token, D_hidden)

        slot_maps = self.layer4(slot_maps)  #  (B * S, token, 1024 + 1)

        return slot_maps, slot_maps


class Decoder_to_DINOV2(nn.Module):
    def __init__(self, args):
        super().__init__()

        # === Token calculations ===
        slot_dim = args['slot_dim']
        hidden_dim = 2048

        # === MLP Based Decoder ===
        self.layer1 = nn.Linear(slot_dim, hidden_dim)
        self.layer2 = nn.Linear(hidden_dim, hidden_dim)
        self.layer3 = nn.Linear(hidden_dim, hidden_dim)
        self.layer4 = nn.Linear(hidden_dim, 1024 + 1)

        self.layer_to_dinov2 = nn.Linear(hidden_dim, 768)
        self.relu = nn.ReLU(inplace=True)

    def forward(self, slot_maps):
        # :arg slot_maps: (B * S, token, D_slot)
        slot_maps =  self.relu(self.layer1(slot_maps))  #  (B * S, token, D_hidden)
        x_dinov2 = self.layer_to_dinov2(slot_maps)
        slot_maps = self.relu(self.layer2(slot_maps))  #  (B * S, token, D_hidden)
        slot_maps = self.relu(self.layer3(slot_maps))  #  (B * S, token, D_hidden)

        slot_maps = self.layer4(slot_maps)  #  (B * S, token, 1024 + 1)

        return slot_maps, x_dinov2

from torch.nn.init import trunc_normal_
class DINOHead(nn.Module):
    def __init__(self, in_dim, out_dim, use_bn=False, norm_last_layer=True, nlayers=3, hidden_dim=2048, bottleneck_dim=768):
        super().__init__()
        nlayers = max(nlayers, 1)
        if nlayers == 1:
            self.mlp = nn.Linear(in_dim, bottleneck_dim)
        else:
            layers = [nn.Linear(in_dim, hidden_dim)]
            if use_bn:
                layers.append(nn.BatchNorm1d(hidden_dim))
            layers.append(nn.GELU())
            for _ in range(nlayers - 2):
                layers.append(nn.Linear(hidden_dim, hidden_dim))
                if use_bn:
                    layers.append(nn.BatchNorm1d(hidden_dim))
                layers.append(nn.GELU())
            layers.append(nn.Linear(hidden_dim, bottleneck_dim))
            self.mlp = nn.Sequential(*layers)
        self.apply(self._init_weights)
        self.gelu = nn.GELU()
        self.last_layer1 = nn.Linear(bottleneck_dim, bottleneck_dim)
        self.last_layer2 = nn.Linear(bottleneck_dim, out_dim)

        # self.last_layer = nn.utils.weight_norm(nn.Linear(bottleneck_dim, out_dim, bias=False))
        # self.last_layer.weight_g.data.fill_(1)
        # if norm_last_layer:
        #     self.last_layer.weight_g.requires_grad = False

    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)

    def forward(self, x):
        x_dinov2 = self.mlp(x)
        # x = nn.functional.normalize(x, dim=-1, p=2)
        x = self.gelu(self.last_layer1(x_dinov2))
        x = self.last_layer2(x)
        return x, x_dinov2


class ISA(nn.Module):
    def __init__(self, args, input_dim):
        super().__init__()

        self.num_slots = args.num_slots
        self.scale = args.slot_dim ** -0.5
        self.iters = args.slot_att_iter
        self.slot_dim = args.slot_dim
        self.query_opt = args.query_opt

        self.res_h = args.resize_to[0] // args.patch_size
        self.res_w = args.resize_to[1] // args.patch_size
        self.token = int(self.res_h * self.res_w)

        # === abs_grid ===
        self.sigma = 5
        xs = torch.linspace(-1, 1, steps=self.res_w)  # (C_x)
        ys = torch.linspace(-1, 1, steps=self.res_h)  # (C_y)

        xs, ys = torch.meshgrid(xs, ys, indexing='xy')  # (C_x, C_y), (C_x, C_y)
        xs = xs.reshape(1, 1, -1, 1)  # (1, 1, C_x * C_y, 1)
        ys = ys.reshape(1, 1, -1, 1)  # (1, 1, C_x * C_y, 1)
        self.abs_grid = nn.Parameter(torch.cat([xs, ys], dim=-1), requires_grad=False)  # (1, 1, token, 2)
        assert self.abs_grid.shape[2] == self.token

        self.h = nn.Linear(2, self.slot_dim)
        # === === ===

        # === Slot related ===
        if self.query_opt:
            self.slots = nn.Parameter(torch.Tensor(1, self.num_slots, self.slot_dim))
            init.xavier_uniform_(self.slots)
        else:
            self.slots_mu = nn.Parameter(torch.randn(1, 1, self.slot_dim))
            self.slots_logsigma = nn.Parameter(torch.zeros(1, 1, self.slot_dim))
            init.xavier_uniform_(self.slots_mu)
            init.xavier_uniform_(self.slots_logsigma)

        self.S_s = nn.Parameter(torch.Tensor(1, self.num_slots, 1, 2))  #  (1, S, 1, 2)
        self.S_p = nn.Parameter(torch.Tensor(1, self.num_slots, 1, 2))  #  (1, S, 1, 2)

        init.normal_(self.S_s, mean=0., std=.02)
        init.normal_(self.S_p, mean=0., std=.02)
        # === === ===

        # === Slot Attention related ===
        self.Q = nn.Linear(self.slot_dim, self.slot_dim, bias=False)
        self.norm = nn.LayerNorm(self.slot_dim)
        self.gru = nn.GRUCell(self.slot_dim, self.slot_dim)
        self.mlp = MLP(self.slot_dim, 4 * self.slot_dim, self.slot_dim,
                       residual=True, layer_order="pre")
        # === === ===

        # === Query & Key & Value ===
        self.K = nn.Linear(self.slot_dim, self.slot_dim, bias=False)
        self.V = nn.Linear(self.slot_dim, self.slot_dim, bias=False)

        self.g = nn.Linear(2, self.slot_dim)
        self.f = nn.Sequential(nn.Linear(self.slot_dim, self.slot_dim),
                               nn.ReLU(inplace=True),
                               nn.Linear(self.slot_dim, self.slot_dim))
        # === === ===

        # Note: starts and ends with LayerNorm
        self.initial_mlp = nn.Sequential(nn.LayerNorm(input_dim),
                                         nn.Linear(input_dim, input_dim),
                                         nn.ReLU(inplace=True),
                                         nn.Linear(input_dim, self.slot_dim),
                                         nn.LayerNorm(self.slot_dim))

        self.final_layer = nn.Linear(self.slot_dim, self.slot_dim)

    def get_rel_grid(self, attn):
        # :arg attn: (B, S, token)
        #
        # :return: (B, S, N, D_slot)

        B, S = attn.shape[:2]
        attn = attn.unsqueeze(dim=2)  # (B, S, 1, token)

        abs_grid = self.abs_grid.expand(B, S, self.token, 2)  # (B, S, token, 2)

        S_p = torch.einsum('bsjd,bsij->bsd', abs_grid, attn)  # (B, S, token, 2) x (B, S, 1, token) -> (B, S, 2)
        S_p = S_p.unsqueeze(dim=2)  # (B, S, 1, 2)

        values_ss = torch.pow(abs_grid - S_p, 2)  # (B, S, token, 2)
        S_s = torch.einsum('bsjd,bsij->bsd', values_ss, attn)  # (B, S, token, 2) x (B, S, 1, token) -> (B, S, 2)
        S_s = torch.sqrt(S_s)  # (B, S, 2)
        S_s = S_s.unsqueeze(dim=2)  # (B, S, 1, 2)

        rel_grid = (abs_grid - S_p) / (S_s * self.sigma)  # (B, S, token, 2)
        rel_grid = self.h(rel_grid)  # (B, S, token, D_slot)

        return rel_grid

    def forward(self, inputs):
        # :arg inputs:              (B, token, D)
        #
        # :return slots:            (B, S, D_slot)
        # :return attn:             (B, S, token)

        B, N, D = inputs.shape
        S = self.num_slots
        D_slot = self.slot_dim
        epsilon = 1e-8

        if self.query_opt:
            slots = self.slots.expand(B, S, D_slot)  # (B, S, D_slot)
        else:
            mu = self.slots_mu.expand(B, S, D_slot)
            sigma = self.slots_logsigma.exp().expand(B, S, D_slot)
            slots = mu + sigma * torch.randn(mu.shape, device=sigma.device, dtype=sigma.dtype)

        slots_init = slots
        inputs = self.initial_mlp(inputs).unsqueeze(dim=1)  # (B, 1, token, D_slot)
        inputs = inputs.expand(B, S, N, D_slot)  # (B * F, S, N', D_slot)

        abs_grid = self.abs_grid.expand(B, S, self.token, 2)  #  (B, S, token, 2)

        assert torch.sum(torch.isnan(abs_grid)) == 0

        S_s = self.S_s.expand(B, S, 1, 2)  #  (B, S, 1, 2)
        S_p = self.S_p.expand(B, S, 1, 2)  #  (B, S, 1, 2)

        for t in range(self.iters + 1):
            # last iteration for S_s and S_p: t = self.iters
            # last meaningful iteration: t = self.iters - 1

            assert torch.sum(torch.isnan(slots)) == 0, f"Iteration {t}"
            assert torch.sum(torch.isnan(S_s)) == 0, f"Iteration {t}"
            assert torch.sum(torch.isnan(S_p)) == 0, f"Iteration {t}"

            if self.query_opt and (t == self.iters - 1):
                slots = slots.detach() + slots_init - slots_init.detach()

            slots_prev = slots
            slots = self.norm(slots)

            # === key and value calculation using rel_grid ===
            rel_grid = (abs_grid - S_p) / (S_s * self.sigma)  # (B, S, token, 2)
            k = self.f(self.K(inputs) + self.g(rel_grid))  # (B, S, token, D_slot)
            v = self.f(self.V(inputs) + self.g(rel_grid))  # (B, S, token, D_slot)

            # === Calculate attention ===
            q = self.Q(slots).unsqueeze(dim=-1)  # (B, S, D_slot, 1)

            dots = torch.einsum('bsdi,bsjd->bsj', q, k)  # (B, S, D_slot, 1) x (B, S, token, D_slot) -> (B, S, token)
            dots *= self.scale  # (B, S, token)
            attn = dots.softmax(dim=1) + epsilon  # (B, S, token)

            # === Weighted mean ===
            attn = attn / attn.sum(dim=-1, keepdim=True)  # (B, S, token)
            attn = attn.unsqueeze(dim=2)  # (B, S, 1, token)
            updates = torch.einsum('bsjd,bsij->bsd', v,
                                   attn)  # (B, S, token, D_slot) x (B, S, 1, token) -> (B, S, D_slot)

            #  === Update S_p and S_s ===
            S_p = torch.einsum('bsjd,bsij->bsd', abs_grid, attn)  # (B, S, token, 2) x (B, S, 1, token) -> (B, S, 2)
            S_p = S_p.unsqueeze(dim=2)  # (B, S, 1, 2)

            values_ss = torch.pow(abs_grid - S_p, 2)  # (B, S, token, 2)
            S_s = torch.einsum('bsjd,bsij->bsd', values_ss, attn)  # (B, S, token, 2) x (B, S, 1, token) -> (B, S, 2)
            S_s = torch.sqrt(S_s)  # (B, S, 2)
            S_s = S_s.unsqueeze(dim=2)  # (B, S, 1, 2)

            # === Update ===
            if t != self.iters:
                slots = self.gru(
                    updates.reshape(-1, self.slot_dim),
                    slots_prev.reshape(-1, self.slot_dim))

                slots = slots.reshape(B, -1, self.slot_dim)
                slots = self.mlp(slots)

        slots = self.final_layer(slots_prev)  # (B, S, D_slot)
        attn = attn.squeeze(dim=2)  # (B, S, token)

        return slots, attn


class SA(nn.Module):
    def __init__(self, args, input_dim):

        super().__init__()
        self.num_slots = args['num_slots']
        self.scale = args['num_slots'] ** -0.5
        self.iters = args['slot_att_iter']
        self.slot_dim = args['slot_dim']
        self.query_opt = args['query_opt']

        # === Slot related ===
        if self.query_opt:
            self.slots = nn.Parameter(torch.Tensor(1, self.num_slots, self.slot_dim))
            init.xavier_uniform_(self.slots)
        else:
            self.slots_mu = nn.Parameter(torch.randn(1, 1, self.slot_dim))
            self.slots_logsigma = nn.Parameter(torch.zeros(1, 1, self.slot_dim))
            init.xavier_uniform_(self.slots_mu)
            init.xavier_uniform_(self.slots_logsigma)

        # === Slot Attention related ===
        self.Q = nn.Linear(self.slot_dim, self.slot_dim, bias=False)
        self.norm = nn.LayerNorm(self.slot_dim)
        # self.update_norm = nn.LayerNorm(self.slot_dim)
        self.gru = nn.GRUCell(self.slot_dim, self.slot_dim)
        self.mlp = MLP(self.slot_dim, 4 * self.slot_dim, self.slot_dim,
                       residual=True, layer_order="pre")
        # === === ===

        # === Query & Key & Value ===
        self.K = nn.Linear(self.slot_dim, self.slot_dim, bias=False)
        self.V = nn.Linear(self.slot_dim, self.slot_dim, bias=False)

        # self.f = nn.Sequential(nn.Linear(self.slot_dim, self.slot_dim),
        #                        nn.ReLU(inplace=True),
        #                        nn.Linear(self.slot_dim, self.slot_dim))
        # === === ===

        # Note: starts and ends with LayerNorm
        self.initial_mlp = nn.Sequential(nn.LayerNorm(input_dim),
                                         nn.Linear(input_dim, input_dim),
                                         nn.ReLU(inplace=True),
                                         nn.Linear(input_dim, self.slot_dim),
                                         nn.LayerNorm(self.slot_dim))

        self.final_layer = nn.Linear(self.slot_dim, self.slot_dim)

    def forward(self, inputs):
        # :arg inputs:              (B, token, D)
        #
        # :return slots:            (B, S, D_slot)

        B = inputs.shape[0]
        S = self.num_slots
        D_slot = self.slot_dim
        epsilon = 1e-8

        if self.query_opt:
            slots = self.slots.expand(B, S, D_slot)  # (B, S, D_slot)
        else:
            mu = self.slots_mu.expand(B, S, D_slot)
            sigma = self.slots_logsigma.exp().expand(B, S, D_slot)
            slots = mu + sigma * torch.randn(mu.shape, device=sigma.device, dtype=sigma.dtype)

        slots_init = slots
        inputs = self.initial_mlp(inputs)  # (B, token, D_slot)

        keys = self.K(inputs)  # (B, token, D_slot)
        values = self.V(inputs)  # (B, token, D_slot)

        for t in range(self.iters):
            assert torch.sum(torch.isnan(slots)) == 0, f"Iteration {t}"

            if t == self.iters - 1 and self.query_opt:
                slots = slots.detach() + slots_init - slots_init.detach()

            slots_prev = slots
            slots = self.norm(slots)
            queries = self.Q(slots)  # (B, S, D_slot)

            dots = torch.einsum('bsd,btd->bst', queries, keys)  # (B, S, token)
            dots *= self.scale  # (B, S, token)
            attn = dots.softmax(dim=1) + epsilon  # (B, S, token)
            attn = attn / attn.sum(dim=-1, keepdim=True)  # (B, S, token)

            updates = torch.einsum('bst,btd->bsd', attn, values)  # (B, S, D_slot)

            slots = self.gru(
                updates.reshape(-1, self.slot_dim),
                slots_prev.reshape(-1, self.slot_dim))

            slots = slots.reshape(B, -1, self.slot_dim)
            slots = self.mlp(slots)

        self.final_layer(slots)

        return slots


class DINOSAURpp(nn.Module):
    def __init__(self, args):
        super().__init__()

        self.slot_dim = args['slot_dim']
        self.slot_num = args['num_slots']
        self.token_num = args['token_num']

        self.ISA = args['ISA']
        if self.ISA:
            self.slot_encoder = ISA(args, input_dim=1024)
        else:
            self.slot_encoder = SA(args, input_dim=1024)

        self.slot_decoder = Decoder(args) #ori easy mlp
        # self.slot_decoder = DINOHead(in_dim=256, out_dim=1024+1, nlayers=3, bottleneck_dim=768) #ori easy mlp
        # self.slot_decoder = Decoder_to_DINOV2(args) #ori easy mlp

        self.pos_dec = nn.Parameter(torch.Tensor(1, self.token_num, self.slot_dim))
        init.normal_(self.pos_dec, mean=0., std=.02)

    def sbd_slots(self, slots):
        # :arg slots: (B, S, D_slot)
        #
        # :return slots: (B, S, token, D_slot)

        B, S, D_slot = slots.shape

        slots = slots.view(-1, 1, D_slot)  # (B * S, 1, D_slot)
        slots = slots.tile(1, self.token_num, 1)  # (B * S, token, D_slot)

        pos_embed = self.pos_dec.expand(slots.shape)
        slots = slots + pos_embed  #  (B * S, token, D_slot)
        slots = slots.view(B, S, self.token_num, D_slot)
        pos_embed = pos_embed.view(B, S, self.token_num, D_slot)

        return slots, pos_embed

    def reconstruct_feature_map(self, slot_maps):
        # :arg slot_maps: (B, S, token, 1024 + 1)
        #
        # :return reconstruction: (B, token, 1024)
        # :return masks: (B, S, token)

        B = slot_maps.shape[0]

        channels, masks = torch.split(slot_maps, [1024, 1], dim=-1)  # (B, S, token, 1024), (B, S, token, 1)
        masks = masks.softmax(dim=1)  # (B, S, token, 1)

        reconstruction = torch.sum(channels * masks, dim=1)  # (B, token, 1024)
        masks = masks.squeeze(dim=-1)  # (B, S, token)

        return reconstruction, masks

    def forward(self, features):
        # :arg features: (B, token, 1024)
        #
        # :return reconstruction: (B, token, 1024)
        # :return slots: (B, S, D_slot)
        # :return masks: (B, S, token)

        B, token, _ = features.shape

        if self.ISA:
            slots, attn = self.slot_encoder(features)  # (B, S, D_slot), (B, S, token)
            assert torch.sum(torch.isnan(slots)) == 0
            assert torch.sum(torch.isnan(attn)) == 0

            rel_grid = self.slot_encoder.get_rel_grid(attn)  # (B, S, token, D_slot)

            slot_maps = self.sbd_slots(slots) + rel_grid  # (B, S, token, D_slot)
            slot_maps, x_dinov2 = self.slot_decoder(slot_maps)  # (B, S, token, 1024 + 1)

        else:
            slots = self.slot_encoder(features)  # (B, S, D_slot), (B, S, token)
            assert torch.sum(torch.isnan(slots)) == 0

            slot_maps, pos_maps = self.sbd_slots(slots)
            slot_maps, x_dinov2 = self.slot_decoder(slot_maps)  # (B, S, token, 1024 + 1)

        reconstruction, masks = self.reconstruct_feature_map(slot_maps)  # (B, token, 1024), (B, S, token)

        return reconstruction, slots, masks, x_dinov2