modeling_mvanet.py

"""PyTorch MVANet model for semantic segmentation."""

import math
from typing import Optional, Tuple, Union

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint
from einops import rearrange
from huggingface_hub import hf_hub_download
from timm.layers import DropPath, to_2tuple, trunc_normal_
from timm.models import load_checkpoint
from transformers import PreTrainedModel
from transformers.modeling_outputs import SemanticSegmenterOutput

from mvanet.transformers.configuration_mvanet import MVANetConfig

# ============================================================================
# Helper Functions
# ============================================================================


def get_activation_fn(activation):
    """Return an activation function given a string"""
    if activation == "relu":
        return F.relu
    if activation == "gelu":
        return F.gelu
    if activation == "glu":
        return F.glu
    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")


def make_cbr(in_dim, out_dim):
    return nn.Sequential(
        nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1),
        nn.BatchNorm2d(out_dim),
        nn.PReLU(),
    )


def make_cbg(in_dim, out_dim):
    return nn.Sequential(
        nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1),
        nn.BatchNorm2d(out_dim),
        nn.GELU(),
    )


def rescale_to(x, scale_factor: float = 2, interpolation="nearest"):
    return F.interpolate(x, scale_factor=scale_factor, mode=interpolation)


def resize_as(x, y, interpolation="bilinear"):
    return F.interpolate(x, size=y.shape[-2:], mode=interpolation)


def image2patches(x):
    """b c (hg h) (wg w) -> (hg wg b) c h w"""
    b, c, h, w = x.shape
    if h % 2 != 0 or w % 2 != 0:
        x = F.interpolate(
            x, size=(h + h % 2, w + w % 2), mode="bilinear", align_corners=False
        )
    x = rearrange(x, "b c (hg h) (wg w) -> (hg wg b) c h w", hg=2, wg=2)
    return x


def patches2image(x):
    """(hg wg b) c h w -> b c (hg h) (wg w)"""
    patches_b, c, h, w = x.shape
    actual_b = patches_b // 4
    x = rearrange(x, "(hg wg b) c h w -> b c (hg h) (wg w)", hg=2, wg=2, b=actual_b)
    return x


# ============================================================================
# Position Embedding
# ============================================================================


class PositionEmbeddingSine(nn.Module):
    def __init__(
        self, num_pos_feats=64, temperature=10000, normalize=False, scale=None
    ):
        super().__init__()
        self.num_pos_feats = num_pos_feats
        self.temperature = temperature
        self.normalize = normalize
        if scale is not None and normalize is False:
            raise ValueError("normalize should be True if scale is passed")
        if scale is None:
            scale = 2 * math.pi
        self.scale = scale
        self.dim_t = torch.arange(
            0,
            self.num_pos_feats,
            dtype=torch.float32,
            device=torch.device("cuda" if torch.cuda.is_available() else "cpu"),
        )

    def __call__(self, b, h, w):
        mask = torch.zeros([b, h, w], dtype=torch.bool, device=self.dim_t.device)
        assert mask is not None
        not_mask = ~mask
        y_embed = not_mask.cumsum(dim=1, dtype=torch.float32)
        x_embed = not_mask.cumsum(dim=2, dtype=torch.float32)
        if self.normalize:
            eps = 1e-6
            y_embed = ((y_embed - 0.5) / (y_embed[:, -1:, :] + eps) * self.scale).to(
                mask.device
            )
            x_embed = ((x_embed - 0.5) / (x_embed[:, :, -1:] + eps) * self.scale).to(
                mask.device
            )

        dim_t = self.temperature ** (2 * (self.dim_t // 2) / self.num_pos_feats)

        pos_x = x_embed[:, :, :, None] / dim_t
        pos_y = y_embed[:, :, :, None] / dim_t
        pos_x = torch.stack(
            (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4
        ).flatten(3)
        pos_y = torch.stack(
            (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4
        ).flatten(3)
        return torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)


# ============================================================================
# Swin Transformer Components
# ============================================================================


class Mlp(nn.Module):
    """Multilayer perceptron."""

    def __init__(
        self,
        in_features,
        hidden_features=None,
        out_features=None,
        act_layer=nn.GELU,
        drop=0.0,
    ):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x


def window_partition(x, window_size):
    """
    Args:
        x: (B, H, W, C)
        window_size (int): window size

    Returns:
        windows: (num_windows*B, window_size, window_size, C)
    """
    B, H, W, C = x.shape
    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
    windows = (
        x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
    )
    return windows


def window_reverse(windows, window_size, H, W):
    """
    Args:
        windows: (num_windows*B, window_size, window_size, C)
        window_size (int): Window size
        H (int): Height of image
        W (int): Width of image

    Returns:
        x: (B, H, W, C)
    """
    B = int(windows.shape[0] / (H * W / window_size / window_size))
    x = windows.view(
        B, H // window_size, W // window_size, window_size, window_size, -1
    )
    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
    return x


class WindowAttention(nn.Module):
    """Window based multi-head self attention (W-MSA) module with relative position bias.
    It supports both of shifted and non-shifted window.

    Args:
        dim (int): Number of input channels.
        window_size (tuple[int]): The height and width of the window.
        num_heads (int): Number of attention heads.
        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
    """

    def __init__(
        self,
        dim,
        window_size,
        num_heads,
        qkv_bias=True,
        qk_scale=None,
        attn_drop=0.0,
        proj_drop=0.0,
    ):
        super().__init__()
        self.dim = dim
        self.window_size = window_size  # Wh, Ww
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim**-0.5

        # define a parameter table of relative position bias
        self.relative_position_bias_table = nn.Parameter(
            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)
        )  # 2*Wh-1 * 2*Ww-1, nH

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

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

        trunc_normal_(self.relative_position_bias_table, std=0.02)
        self.softmax = nn.Softmax(dim=-1)

    def forward(self, x, mask=None):
        """Forward function.

        Args:
            x: input features with shape of (num_windows*B, N, C)
            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
        """
        B_, N, C = x.shape
        qkv = (
            self.qkv(x)
            .reshape(B_, N, 3, self.num_heads, C // self.num_heads)
            .permute(2, 0, 3, 1, 4)
        )
        q, k, v = (
            qkv[0],
            qkv[1],
            qkv[2],
        )  # make torchscript happy (cannot use tensor as tuple)

        q = q * self.scale
        attn = q @ k.transpose(-2, -1)

        relative_position_index = self.relative_position_index
        assert isinstance(relative_position_index, torch.Tensor)
        relative_position_bias = self.relative_position_bias_table[
            relative_position_index.view(-1)
        ].view(
            self.window_size[0] * self.window_size[1],
            self.window_size[0] * self.window_size[1],
            -1,
        )  # Wh*Ww,Wh*Ww,nH
        relative_position_bias = relative_position_bias.permute(
            2, 0, 1
        ).contiguous()  # nH, Wh*Ww, Wh*Ww
        attn = attn + relative_position_bias.unsqueeze(0)

        if mask is not None:
            nW = mask.shape[0]
            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(
                1
            ).unsqueeze(0)
            attn = attn.view(-1, self.num_heads, N, N)
            attn = self.softmax(attn)
        else:
            attn = self.softmax(attn)

        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 SwinTransformerBlock(nn.Module):
    """Swin Transformer Block.

    Args:
        dim (int): Number of input channels.
        num_heads (int): Number of attention heads.
        window_size (int): Window size.
        shift_size (int): Shift size for SW-MSA.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
        drop (float, optional): Dropout rate. Default: 0.0
        attn_drop (float, optional): Attention dropout rate. Default: 0.0
        drop_path (float, optional): Stochastic depth rate. Default: 0.0
        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
    """

    def __init__(
        self,
        dim,
        num_heads,
        window_size=7,
        shift_size=0,
        mlp_ratio=4.0,
        qkv_bias=True,
        qk_scale=None,
        drop=0.0,
        attn_drop=0.0,
        drop_path=0.0,
        act_layer=nn.GELU,
        norm_layer=nn.LayerNorm,
    ):
        super().__init__()
        self.dim = dim
        self.num_heads = num_heads
        self.window_size = window_size
        self.shift_size = shift_size
        self.mlp_ratio = mlp_ratio
        assert 0 <= self.shift_size < self.window_size, (
            "shift_size must in 0-window_size"
        )

        self.norm1 = norm_layer(dim)
        self.attn = WindowAttention(
            dim,
            window_size=to_2tuple(self.window_size),
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            qk_scale=qk_scale,
            attn_drop=attn_drop,
            proj_drop=drop,
        )

        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(
            in_features=dim,
            hidden_features=mlp_hidden_dim,
            act_layer=act_layer,
            drop=drop,
        )

        self.H: int | None = None
        self.W: int | None = None

    def forward(self, x, mask_matrix):
        """Forward function.

        Args:
            x: Input feature, tensor size (B, H*W, C).
            H, W: Spatial resolution of the input feature.
            mask_matrix: Attention mask for cyclic shift.
        """
        B, L, C = x.shape
        H, W = self.H, self.W
        assert H is not None and W is not None, "H and W must be set before forward"
        assert L == H * W, "input feature has wrong size"

        shortcut = x
        x = self.norm1(x)
        x = x.view(B, H, W, C)

        # pad feature maps to multiples of window size
        pad_l = pad_t = 0
        pad_r = (self.window_size - W % self.window_size) % self.window_size
        pad_b = (self.window_size - H % self.window_size) % self.window_size
        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
        _, Hp, Wp, _ = x.shape

        # cyclic shift
        if self.shift_size > 0:
            shifted_x = torch.roll(
                x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)
            )
            attn_mask = mask_matrix
        else:
            shifted_x = x
            attn_mask = None

        # partition windows
        x_windows = window_partition(
            shifted_x, self.window_size
        )  # nW*B, window_size, window_size, C
        x_windows = x_windows.view(
            -1, self.window_size * self.window_size, C
        )  # nW*B, window_size*window_size, C

        # W-MSA/SW-MSA
        attn_windows = self.attn(
            x_windows, mask=attn_mask
        )  # nW*B, window_size*window_size, C

        # merge windows
        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C

        # reverse cyclic shift
        if self.shift_size > 0:
            x = torch.roll(
                shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
            )
        else:
            x = shifted_x

        if pad_r > 0 or pad_b > 0:
            x = x[:, :H, :W, :].contiguous()

        x = x.view(B, H * W, C)

        # FFN
        x = shortcut + self.drop_path(x)
        x = x + self.drop_path(self.mlp(self.norm2(x)))

        return x


class PatchMerging(nn.Module):
    """Patch Merging Layer

    Args:
        dim (int): Number of input channels.
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
    """

    def __init__(self, dim, norm_layer=nn.LayerNorm):
        super().__init__()
        self.dim = dim
        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
        self.norm = norm_layer(4 * dim)

    def forward(self, x, H, W):
        """Forward function.

        Args:
            x: Input feature, tensor size (B, H*W, C).
            H, W: Spatial resolution of the input feature.
        """
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size"

        x = x.view(B, H, W, C)

        # padding
        pad_input = (H % 2 == 1) or (W % 2 == 1)
        if pad_input:
            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))

        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C

        x = self.norm(x)
        x = self.reduction(x)

        return x


class BasicLayer(nn.Module):
    """A basic Swin Transformer layer for one stage.

    Args:
        dim (int): Number of feature channels
        depth (int): Depths of this stage.
        num_heads (int): Number of attention head.
        window_size (int): Local window size. Default: 7.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
        drop (float, optional): Dropout rate. Default: 0.0
        attn_drop (float, optional): Attention dropout rate. Default: 0.0
        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
    """

    def __init__(
        self,
        dim,
        depth,
        num_heads,
        window_size=7,
        mlp_ratio=4.0,
        qkv_bias=True,
        qk_scale=None,
        drop=0.0,
        attn_drop=0.0,
        drop_path=0.0,
        norm_layer=nn.LayerNorm,
        downsample=None,
        use_checkpoint=False,
    ):
        super().__init__()
        self.window_size = window_size
        self.shift_size = window_size // 2
        self.depth = depth
        self.use_checkpoint = use_checkpoint

        # build blocks
        self.blocks = nn.ModuleList(
            [
                SwinTransformerBlock(
                    dim=dim,
                    num_heads=num_heads,
                    window_size=window_size,
                    shift_size=0 if (i % 2 == 0) else window_size // 2,
                    mlp_ratio=mlp_ratio,
                    qkv_bias=qkv_bias,
                    qk_scale=qk_scale,
                    drop=drop,
                    attn_drop=attn_drop,
                    drop_path=drop_path[i]
                    if isinstance(drop_path, list)
                    else drop_path,
                    norm_layer=norm_layer,
                )
                for i in range(depth)
            ]
        )

        # patch merging layer
        if downsample is not None:
            self.downsample = downsample(dim=dim, norm_layer=norm_layer)
        else:
            self.downsample = None

    def forward(self, x, H, W):
        """Forward function.

        Args:
            x: Input feature, tensor size (B, H*W, C).
            H, W: Spatial resolution of the input feature.
        """

        # calculate attention mask for SW-MSA
        Hp = int(np.ceil(H / self.window_size)) * self.window_size
        Wp = int(np.ceil(W / self.window_size)) * self.window_size
        img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device)  # 1 Hp Wp 1
        h_slices = (
            slice(0, -self.window_size),
            slice(-self.window_size, -self.shift_size),
            slice(-self.shift_size, None),
        )
        w_slices = (
            slice(0, -self.window_size),
            slice(-self.window_size, -self.shift_size),
            slice(-self.shift_size, None),
        )
        cnt = 0
        for h in h_slices:
            for w in w_slices:
                img_mask[:, h, w, :] = cnt
                cnt += 1

        mask_windows = window_partition(
            img_mask, self.window_size
        )  # nW, window_size, window_size, 1
        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(
            attn_mask == 0, float(0.0)
        )

        for blk in self.blocks:
            blk.H, blk.W = H, W
            if self.use_checkpoint:
                x = checkpoint.checkpoint(blk, x, attn_mask)
            else:
                x = blk(x, attn_mask)
        if self.downsample is not None:
            x_down = self.downsample(x, H, W)
            Wh, Ww = (H + 1) // 2, (W + 1) // 2
            return x, H, W, x_down, Wh, Ww
        else:
            return x, H, W, x, H, W


class PatchEmbed(nn.Module):
    """Image to Patch Embedding

    Args:
        patch_size (int): Patch token size. Default: 4.
        in_chans (int): Number of input image channels. Default: 3.
        embed_dim (int): Number of linear projection output channels. Default: 96.
        norm_layer (nn.Module, optional): Normalization layer. Default: None
    """

    def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
        super().__init__()
        patch_size = to_2tuple(patch_size)
        self.patch_size = patch_size

        self.in_chans = in_chans
        self.embed_dim = embed_dim

        self.proj = nn.Conv2d(
            in_chans, embed_dim, kernel_size=patch_size, stride=patch_size
        )
        if norm_layer is not None:
            self.norm = norm_layer(embed_dim)
        else:
            self.norm = None

    def forward(self, x):
        """Forward function."""
        # padding
        _, _, H, W = x.size()
        if W % self.patch_size[1] != 0:
            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
        if H % self.patch_size[0] != 0:
            x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))

        x = self.proj(x)  # B C Wh Ww
        if self.norm is not None:
            Wh, Ww = x.size(2), x.size(3)
            x = x.flatten(2).transpose(1, 2)
            x = self.norm(x)
            x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)

        return x


class SwinTransformer(nn.Module):
    """Swin Transformer backbone.
        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
          https://arxiv.org/pdf/2103.14030

    Args:
        pretrain_img_size (int): Input image size for training the pretrained model,
            used in absolute postion embedding. Default 224.
        patch_size (int | tuple(int)): Patch size. Default: 4.
        in_chans (int): Number of input image channels. Default: 3.
        embed_dim (int): Number of linear projection output channels. Default: 96.
        depths (tuple[int]): Depths of each Swin Transformer stage.
        num_heads (tuple[int]): Number of attention head of each stage.
        window_size (int): Window size. Default: 7.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
        drop_rate (float): Dropout rate.
        attn_drop_rate (float): Attention dropout rate. Default: 0.
        drop_path_rate (float): Stochastic depth rate. Default: 0.2.
        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
        patch_norm (bool): If True, add normalization after patch embedding. Default: True.
        out_indices (Sequence[int]): Output from which stages.
        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
            -1 means not freezing any parameters.
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
    """

    def __init__(
        self,
        pretrain_img_size=224,
        patch_size=4,
        in_chans=3,
        embed_dim=96,
        depths=[2, 2, 6, 2],
        num_heads=[3, 6, 12, 24],
        window_size=7,
        mlp_ratio=4.0,
        qkv_bias=True,
        qk_scale=None,
        drop_rate=0.0,
        attn_drop_rate=0.0,
        drop_path_rate=0.2,
        norm_layer=nn.LayerNorm,
        ape=False,
        patch_norm=True,
        out_indices=(0, 1, 2, 3),
        frozen_stages=-1,
        use_checkpoint=False,
    ):
        super().__init__()

        self.pretrain_img_size = pretrain_img_size
        self.num_layers = len(depths)
        self.embed_dim = embed_dim
        self.ape = ape
        self.patch_norm = patch_norm
        self.out_indices = out_indices
        self.frozen_stages = frozen_stages

        # split image into non-overlapping patches
        self.patch_embed = PatchEmbed(
            patch_size=patch_size,
            in_chans=in_chans,
            embed_dim=embed_dim,
            norm_layer=norm_layer if self.patch_norm else None,
        )

        # absolute position embedding
        if self.ape:
            pretrain_img_size = to_2tuple(pretrain_img_size)
            patch_size = to_2tuple(patch_size)
            patches_resolution = [
                pretrain_img_size[0] // patch_size[0],
                pretrain_img_size[1] // patch_size[1],
            ]

            self.absolute_pos_embed = nn.Parameter(
                torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1])
            )
            trunc_normal_(self.absolute_pos_embed, std=0.02)

        self.pos_drop = nn.Dropout(p=drop_rate)

        # stochastic depth
        dpr = [
            x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))
        ]  # stochastic depth decay rule

        # build layers
        self.layers = nn.ModuleList()
        for i_layer in range(self.num_layers):
            layer = BasicLayer(
                dim=int(embed_dim * 2**i_layer),
                depth=depths[i_layer],
                num_heads=num_heads[i_layer],
                window_size=window_size,
                mlp_ratio=mlp_ratio,
                qkv_bias=qkv_bias,
                qk_scale=qk_scale,
                drop=drop_rate,
                attn_drop=attn_drop_rate,
                drop_path=dpr[sum(depths[:i_layer]) : sum(depths[: i_layer + 1])],
                norm_layer=norm_layer,
                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
                use_checkpoint=use_checkpoint,
            )
            self.layers.append(layer)

        num_features = [int(embed_dim * 2**i) for i in range(self.num_layers)]
        self.num_features = num_features

        # add a norm layer for each output
        for i_layer in out_indices:
            layer = norm_layer(num_features[i_layer])
            layer_name = f"norm{i_layer}"
            self.add_module(layer_name, layer)

        self._freeze_stages()

    def _freeze_stages(self):
        if self.frozen_stages >= 0:
            self.patch_embed.eval()
            for param in self.patch_embed.parameters():
                param.requires_grad = False

        if self.frozen_stages >= 1 and self.ape:
            self.absolute_pos_embed.requires_grad = False

        if self.frozen_stages >= 2:
            self.pos_drop.eval()
            for i in range(0, self.frozen_stages - 1):
                m = self.layers[i]
                m.eval()
                for param in m.parameters():
                    param.requires_grad = False

    def init_weights(self, pretrained=None):
        """Initialize the weights in backbone.

        Args:
            pretrained (str, optional): Path to pre-trained weights.
                Defaults to None.
        """

        def _init_weights(m):
            if isinstance(m, nn.Linear):
                trunc_normal_(m.weight, std=0.02)
                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)

        if isinstance(pretrained, str):
            self.apply(_init_weights)
            load_checkpoint(self, pretrained, strict=False)
        elif pretrained is None:
            self.apply(_init_weights)
        else:
            raise TypeError("pretrained must be a str or None")

    def forward(self, x):
        x = self.patch_embed(x)

        Wh, Ww = x.size(2), x.size(3)
        if self.ape:
            # interpolate the position embedding to the corresponding size
            absolute_pos_embed = F.interpolate(
                self.absolute_pos_embed, size=(Wh, Ww), mode="bicubic"
            )
            x = x + absolute_pos_embed  # B Wh*Ww C

        outs = [x.contiguous()]
        x = x.flatten(2).transpose(1, 2)
        x = self.pos_drop(x)
        for i in range(self.num_layers):
            layer = self.layers[i]
            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)

            if i in self.out_indices:
                norm_layer = getattr(self, f"norm{i}")
                x_out = norm_layer(x_out)

                out = (
                    x_out.view(-1, H, W, self.num_features[i])
                    .permute(0, 3, 1, 2)
                    .contiguous()
                )
                outs.append(out)

        return tuple(outs)

    def train(self, mode=True):
        """Convert the model into training mode while keep layers freezed."""
        super(SwinTransformer, self).train(mode)
        self._freeze_stages()


def SwinB(pretrained=True):
    model = SwinTransformer(
        embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12
    )
    if pretrained is True:
        state_dict_path = hf_hub_download(
            repo_id="creative-graphic-design/MVANet-checkpoints",
            filename="swin_base_patch4_window12_384_22kto1k.pth",
        )
        state_dict = torch.load(state_dict_path, map_location="cpu")
        model.load_state_dict(state_dict["model"], strict=False)

    return model


# ============================================================================
# Multi-field Cross Localization Module (MCLM)
# ============================================================================


class inf_MCLM(nn.Module):
    def __init__(self, d_model, num_heads, pool_ratios=[1, 4, 8]):
        super(inf_MCLM, self).__init__()
        self.attention = nn.ModuleList(
            [
                nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
                nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
                nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
                nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
                nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
            ]
        )

        self.linear1 = nn.Linear(d_model, d_model * 2)
        self.linear2 = nn.Linear(d_model * 2, d_model)
        self.linear3 = nn.Linear(d_model, d_model * 2)
        self.linear4 = nn.Linear(d_model * 2, d_model)
        self.norm1 = nn.LayerNorm(d_model)
        self.norm2 = nn.LayerNorm(d_model)
        self.dropout = nn.Dropout(0.1)
        self.dropout1 = nn.Dropout(0.1)
        self.dropout2 = nn.Dropout(0.1)
        self.activation = get_activation_fn("relu")
        self.pool_ratios = pool_ratios
        self.p_poses = None
        self.g_pos = None
        self.positional_encoding = PositionEmbeddingSine(
            num_pos_feats=d_model // 2, normalize=True
        )

    def forward(self, l, g):
        """
        l: 4,c,h,w
        g: 1,c,h,w
        """
        b, c, h, w = l.size()
        # 4,c,h,w -> 1,c,2h,2w
        concated_locs = rearrange(l, "(hg wg b) c h w -> b c (hg h) (wg w)", hg=2, wg=2)
        pools = []
        p_poses_list = []
        for pool_ratio in self.pool_ratios:
            # b,c,h,w
            tgt_hw = (round(h / pool_ratio), round(w / pool_ratio))
            pool = F.adaptive_avg_pool2d(concated_locs, tgt_hw)
            pools.append(rearrange(pool, "b c h w -> (h w) b c"))
            pos_emb = self.positional_encoding(
                pool.shape[0], pool.shape[2], pool.shape[3]
            )
            pos_emb = rearrange(pos_emb, "b c h w -> (h w) b c")
            p_poses_list.append(pos_emb)
        pools = torch.cat(pools, 0)
        p_poses = torch.cat(p_poses_list, dim=0)
        pos_emb = self.positional_encoding(g.shape[0], g.shape[2], g.shape[3])
        g_pos = rearrange(pos_emb, "b c h w -> (h w) b c")

        # attention between glb (q) & multisensory concated-locs (k,v)
        g_hw_b_c = rearrange(g, "b c h w -> (h w) b c")
        g_hw_b_c = g_hw_b_c + self.dropout1(
            self.attention[0](g_hw_b_c + g_pos, pools + p_poses, pools)[0]
        )
        g_hw_b_c = self.norm1(g_hw_b_c)
        g_hw_b_c = g_hw_b_c + self.dropout2(
            self.linear2(self.dropout(self.activation(self.linear1(g_hw_b_c)).clone()))
        )
        g_hw_b_c = self.norm2(g_hw_b_c)

        # attention between origin locs (q) & freashed glb (k,v)
        l_hw_b_c = rearrange(l, "b c h w -> (h w) b c")
        _g_hw_b_c = rearrange(g_hw_b_c, "(h w) b c -> h w b c", h=h, w=w)
        _g_hw_b_c = rearrange(
            _g_hw_b_c, "(ng h) (nw w) b c -> (h w) (ng nw b) c", ng=2, nw=2
        )
        outputs_re = []
        for i, (_l, _g) in enumerate(
            zip(l_hw_b_c.chunk(4, dim=1), _g_hw_b_c.chunk(4, dim=1))
        ):
            outputs_re.append(self.attention[i + 1](_l, _g, _g)[0])  # (h w) 1 c
        outputs_re = torch.cat(outputs_re, 1)  # (h w) 4 c

        l_hw_b_c = l_hw_b_c + self.dropout1(outputs_re)
        l_hw_b_c = self.norm1(l_hw_b_c)
        l_hw_b_c = l_hw_b_c + self.dropout2(
            self.linear4(self.dropout(self.activation(self.linear3(l_hw_b_c)).clone()))
        )
        l_hw_b_c = self.norm2(l_hw_b_c)

        l = torch.cat((l_hw_b_c, g_hw_b_c), 1)  # hw,b(5),c
        return rearrange(l, "(h w) b c -> b c h w", h=h, w=w)  ## (5,c,h*w)


# ============================================================================
# Multi-crop Refinement Module (MCRM)
# ============================================================================


class inf_MCRM(nn.Module):
    def __init__(self, d_model, num_heads, pool_ratios=[4, 8, 16], h=None):
        super(inf_MCRM, self).__init__()
        self.attention = nn.ModuleList(
            [
                nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
                nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
                nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
                nn.MultiheadAttention(d_model, num_heads, dropout=0.1),
            ]
        )

        self.linear3 = nn.Linear(d_model, d_model * 2)
        self.linear4 = nn.Linear(d_model * 2, d_model)
        self.norm1 = nn.LayerNorm(d_model)
        self.norm2 = nn.LayerNorm(d_model)
        self.dropout = nn.Dropout(0.1)
        self.dropout1 = nn.Dropout(0.1)
        self.dropout2 = nn.Dropout(0.1)
        self.sigmoid = nn.Sigmoid()
        self.activation = get_activation_fn("relu")
        self.sal_conv = nn.Conv2d(d_model, 1, 1)
        self.pool_ratios = pool_ratios
        self.positional_encoding = PositionEmbeddingSine(
            num_pos_feats=d_model // 2, normalize=True
        )

    def forward(self, x):
        total_b, c, h, w = x.size()
        # Total batch is 5*batch_size (4 local + 1 global)
        batch_size = total_b // 5

        # Split into local (4*batch_size) and global (batch_size)
        loc, glb = x.split([4 * batch_size, batch_size], dim=0)
        # loc: (4*batch_size, c, h, w), glb: (batch_size, c, h, w)
        patched_glb = rearrange(glb, "b c (hg h) (wg w) -> (hg wg b) c h w", hg=2, wg=2)

        # generate token attention map
        token_attention_map = self.sigmoid(self.sal_conv(glb))
        token_attention_map = F.interpolate(
            token_attention_map, size=patches2image(loc).shape[-2:], mode="nearest"
        )
        loc = loc * rearrange(
            token_attention_map, "b c (hg h) (wg w) -> (hg wg b) c h w", hg=2, wg=2
        )
        pools = []
        for pool_ratio in self.pool_ratios:
            tgt_hw = (round(h / pool_ratio), round(w / pool_ratio))
            pool = F.adaptive_avg_pool2d(patched_glb, tgt_hw)
            pools.append(rearrange(pool, "nl c h w -> nl c (h w)"))
        # pools: (4*batch_size, c, nphw) -> (4*batch_size, nphw, 1, c)
        pools = rearrange(torch.cat(pools, 2), "nl c nphw -> nl nphw 1 c")
        # Reshape to separate batch and patch dimensions: (4, batch_size, nphw, 1, c)
        # Note: image2patches outputs in order (hg wg b) where b changes fastest
        # So the order is: [p0_b0, p0_b1, ..., p1_b0, p1_b1, ..., p3_b0, p3_b1]
        pools = rearrange(pools, "(p b) nphw 1 c -> p b nphw 1 c", p=4, b=batch_size)

        # loc_: (4*batch_size, hw, 1, c) -> (4, batch_size, hw, 1, c)
        loc_ = rearrange(loc, "nl c h w -> nl (h w) 1 c")
        loc_ = rearrange(loc_, "(p b) hw 1 c -> p b hw 1 c", p=4, b=batch_size)

        # Apply attention for each of 4 patches (only 4 iterations, not batch_size!)
        # Each iteration processes all batch items simultaneously
        outputs = []
        for i in range(4):  # Only 4 iterations regardless of batch_size!
            # Extract patch i across all batch items: (batch_size, hw, 1, c)
            q = loc_[i, :, :, :, :]  # (b, hw, 1, c)
            v = pools[i, :, :, :, :]  # (b, nphw, 1, c)
            k = v

            # Reshape for MultiheadAttention: (seq, batch, dim)
            q = rearrange(q, "b hw 1 c -> hw b c")
            k = rearrange(k, "b nphw 1 c -> nphw b c")
            v = rearrange(v, "b nphw 1 c -> nphw b c")

            # Apply attention (processes all batch_size items in parallel)
            attn_out = self.attention[i](q, k, v)[0]  # (hw, b, c)
            outputs.append(attn_out)

        # Concatenate outputs: list of 4 x (hw, b, c) -> (hw, p*b, c)
        # Interleave to match (p b) order: [p0_b0, p0_b1, ..., p1_b0, p1_b1, ...]
        outputs = torch.stack(outputs, dim=2)  # (hw, b, 4, c)
        outputs = rearrange(outputs, "hw b p c -> hw (p b) c")  # (hw, 4*b, c)

        # Continue with existing operations using batch_size
        src = loc.view(4 * batch_size, c, -1).permute(2, 0, 1) + self.dropout1(outputs)
        src = self.norm1(src)
        src = src + self.dropout2(
            self.linear4(self.dropout(self.activation(self.linear3(src)).clone()))
        )
        src = self.norm2(src)

        src = src.permute(1, 2, 0).reshape(4 * batch_size, c, h, w)  # freshed loc
        glb = glb + F.interpolate(
            patches2image(src), size=glb.shape[-2:], mode="nearest"
        )  # freshed glb
        return torch.cat((src, glb), 0)


# ============================================================================
# MVANet Model for Image Segmentation
# ============================================================================


class MVANetForImageSegmentation(PreTrainedModel):
    """
    MVANet Model for image segmentation.

    This model is a direct reimplementation of inf_MVANet with transformers-compatible
    interface for semantic segmentation tasks.

    Args:
        config (:class:`~mvanet.transformers.MVANetConfig`): Model configuration class with all the parameters of the model.
            Initializing with a config file does not load the weights associated with the model, only the configuration.

    Example::\

        >>> from transformers import AutoModel, AutoImageProcessor
        >>> from PIL import Image

        >>> # Load model and processor
        >>> model = AutoModel.from_pretrained("creative-graphic-design/mvanet")
        >>> processor = AutoImageProcessor.from_pretrained("creative-graphic-design/mvanet")

        >>> # Load image
        >>> image = Image.open("image.png")

        >>> # Preprocess
        >>> inputs = processor(image, return_tensors="pt")

        >>> # Forward pass
        >>> outputs = model(**inputs)

        >>> # Post-process
        >>> masks = processor.post_process_semantic_segmentation(
        ...     outputs, target_sizes=[image.size[::-1]]
        ... )
    """

    config_class = MVANetConfig
    base_model_prefix = "mvanet"
    main_input_name = "pixel_values"
    supports_gradient_checkpointing = False
    _no_split_modules = []

    def __init__(self, config: MVANetConfig):
        super().__init__(config)
        self.config = config

        emb_dim = config.embedding_dim

        # Backbone: Swin Transformer
        self.backbone = SwinB(pretrained=config.backbone_pretrained)

        # Feature projection layers - use config values
        out_channels = config.backbone_out_channels
        self.output5 = make_cbr(out_channels[4], emb_dim)  # 1024 -> 128
        self.output4 = make_cbr(out_channels[3], emb_dim)  # 512 -> 128
        self.output3 = make_cbr(out_channels[2], emb_dim)  # 256 -> 128
        self.output2 = make_cbr(out_channels[1], emb_dim)  # 128 -> 128
        self.output1 = make_cbr(out_channels[0], emb_dim)  # 128 -> 128

        # Multi-field Cross Localization Module
        self.multifieldcrossatt = inf_MCLM(
            emb_dim, config.mclm_num_heads, config.mclm_pool_ratios
        )

        # Convolution blocks for decoder
        self.conv1 = make_cbr(emb_dim, emb_dim)
        self.conv2 = make_cbr(emb_dim, emb_dim)
        self.conv3 = make_cbr(emb_dim, emb_dim)
        self.conv4 = make_cbr(emb_dim, emb_dim)

        # Multi-crop Refinement Module decoder blocks
        self.dec_blk1 = inf_MCRM(
            emb_dim, config.mcrm_num_heads, config.mcrm_pool_ratios
        )
        self.dec_blk2 = inf_MCRM(
            emb_dim, config.mcrm_num_heads, config.mcrm_pool_ratios
        )
        self.dec_blk3 = inf_MCRM(
            emb_dim, config.mcrm_num_heads, config.mcrm_pool_ratios
        )
        self.dec_blk4 = inf_MCRM(
            emb_dim, config.mcrm_num_heads, config.mcrm_pool_ratios
        )

        # Instance mask head - use config value
        hidden_dim = config.insmask_hidden_dim
        self.insmask_head = nn.Sequential(
            nn.Conv2d(emb_dim, hidden_dim, kernel_size=3, padding=1),
            nn.BatchNorm2d(hidden_dim),
            nn.PReLU(),
            nn.Conv2d(hidden_dim, hidden_dim, kernel_size=3, padding=1),
            nn.BatchNorm2d(hidden_dim),
            nn.PReLU(),
            nn.Conv2d(hidden_dim, emb_dim, kernel_size=3, padding=1),
        )

        # Shallow feature extraction - use config value
        self.shallow = nn.Sequential(
            nn.Conv2d(config.num_channels, emb_dim, kernel_size=3, padding=1)
        )

        # Upsampling layers
        self.upsample1 = make_cbg(emb_dim, emb_dim)
        self.upsample2 = make_cbg(emb_dim, emb_dim)

        # Final output layer - use config value
        self.output = nn.Sequential(
            nn.Conv2d(emb_dim, config.num_labels, kernel_size=3, padding=1)
        )

        # Set inplace operations for ReLU and Dropout
        for m in self.modules():
            if isinstance(m, nn.ReLU) or isinstance(m, nn.Dropout):
                m.inplace = True

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        pixel_values: torch.FloatTensor,
        labels: Optional[torch.LongTensor] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        **kwargs,
    ) -> Union[Tuple, SemanticSegmenterOutput]:
        """
        Forward pass of the model.

        Args:
            pixel_values (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_channels, height, width)`):
                Pixel values. Pixel values can be obtained using :class:`~mvanet.transformers.MVANetImageProcessor`.
                See :meth:`~mvanet.transformers.MVANetImageProcessor.preprocess` for details.
            labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, height, width)`, `optional`):
                Ground truth semantic segmentation maps for computing the loss.
            output_hidden_states (:obj:`bool`, `optional`):
                Whether or not to return the hidden states of all layers. Currently not supported.
            return_dict (:obj:`bool`, `optional`):
                Whether or not to return a :class:`~transformers.modeling_outputs.SemanticSegmenterOutput` instead of
                a plain tuple.

        Returns:
            :class:`~transformers.modeling_outputs.SemanticSegmenterOutput` or :obj:`tuple`:
                A :class:`~transformers.modeling_outputs.SemanticSegmenterOutput` (if ``return_dict=True`` is passed or
                when ``config.use_return_dict=True``) or a tuple of :obj:`torch.FloatTensor`.

        Example::\

            >>> from mvanet.transformers import MVANetForImageSegmentation, MVANetImageProcessor
            >>> import torch
            >>> from PIL import Image

            >>> processor = MVANetImageProcessor()
            >>> model = MVANetForImageSegmentation.from_pretrained("creative-graphic-design/mvanet")

            >>> image = Image.open("image.png")
            >>> inputs = processor(image, return_tensors="pt")
            >>> outputs = model(**inputs)
            >>> logits = outputs.logits  # (batch_size, num_labels, height, width)
        """
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        batch_size = pixel_values.shape[0]

        # Extract shallow features
        shallow = self.shallow(pixel_values)

        # Create multi-view input: 4 local patches + 1 global view
        # Use config value for global view scale
        glb = rescale_to(
            pixel_values,
            scale_factor=self.config.global_view_scale,
            interpolation="bilinear",
        )
        loc = image2patches(pixel_values)
        input_views = torch.cat((loc, glb), dim=0)

        # Extract features through backbone
        feature = self.backbone(input_views)

        # Project features to embedding dimension
        e5 = self.output5(feature[4])  # (batch*5, 128, 16, 16)
        e4 = self.output4(feature[3])  # (batch*5, 128, 32, 32)
        e3 = self.output3(feature[2])  # (batch*5, 128, 64, 64)
        e2 = self.output2(feature[1])  # (batch*5, 128, 128, 128)
        e1 = self.output1(feature[0])  # (batch*5, 128, 128, 128)

        # Split local and global features at deepest level
        # Use config value for number of patches
        loc_e5, glb_e5 = e5.split(
            [batch_size * self.config.num_patches, batch_size], dim=0
        )

        # Apply multi-field cross attention
        e5_cat = self.multifieldcrossatt(loc_e5, glb_e5)  # (batch*5, 128, 16, 16)

        # Decode through MCRM blocks with skip connections
        e4 = self.conv4(self.dec_blk4(e4 + resize_as(e5_cat, e4)))
        e3 = self.conv3(self.dec_blk3(e3 + resize_as(e4, e3)))
        e2 = self.conv2(self.dec_blk2(e2 + resize_as(e3, e2)))
        e1 = self.conv1(self.dec_blk1(e1 + resize_as(e2, e1)))

        # Split local and global features
        # Use config value for number of patches
        loc_e1, glb_e1 = e1.split(
            [batch_size * self.config.num_patches, batch_size], dim=0
        )

        # Merge local patches back to image
        output1_cat = patches2image(loc_e1)

        # Add global features
        output1_cat = output1_cat + resize_as(glb_e1, output1_cat)

        # Apply instance mask head
        final_output = self.insmask_head(output1_cat)

        # Merge shallow features
        final_output = final_output + resize_as(shallow, final_output)
        final_output = self.upsample1(rescale_to(final_output))
        final_output = rescale_to(final_output + resize_as(shallow, final_output))
        final_output = self.upsample2(final_output)

        # Final output (logits before sigmoid)
        logits = self.output(final_output)

        loss = None
        if labels is not None:
            # Compute binary cross-entropy loss with logits
            # labels should be float with values in [0, 1]
            loss_fct = nn.BCEWithLogitsLoss()
            # Ensure labels have the same shape as logits
            if labels.dim() == 3:
                # (B, H, W) -> (B, 1, H, W)
                labels = labels.unsqueeze(1)
            loss = loss_fct(logits, labels.float())

        if not return_dict:
            output = (logits,)
            return ((loss,) + output) if loss is not None else output

        return SemanticSegmenterOutput(
            loss=loss,
            logits=logits,
            hidden_states=None,
            attentions=None,
        )

    def _init_weights(self, module):
        """
        Initialize weights.

        The backbone (SwinB) and other modules handle their own weight initialization,
        so we don't need to do anything here.
        """
        pass