|
"""Implementation of MSCAN from SegNeXt: Rethinking Convolutional Attention Design for Semantic |
|
Segmentation (NeurIPS 2022) adapted from |
|
|
|
https://github.com/Visual-Attention-Network/SegNeXt/blob/main/mmseg/models/backbones/mscan.py |
|
|
|
|
|
Light Hamburger Decoder adapted from: |
|
|
|
https://github.com/Visual-Attention-Network/SegNeXt/blob/main/mmseg/models/decode_heads/ham_head.py |
|
""" |
|
|
|
from typing import Dict, Tuple |
|
|
|
import torch |
|
import torch.nn as nn |
|
import torch.nn.functional as F |
|
from torch.nn.modules.utils import _pair as to_2tuple |
|
|
|
|
|
|
|
|
|
|
|
class ConvModule(nn.Module): |
|
"""Replacement for mmcv.cnn.ConvModule to avoid mmcv dependency.""" |
|
|
|
def __init__( |
|
self, |
|
in_channels: int, |
|
out_channels: int, |
|
kernel_size: int, |
|
padding: int = 0, |
|
use_norm: bool = False, |
|
bias: bool = True, |
|
): |
|
"""Simple convolution block. |
|
|
|
Args: |
|
in_channels (int): Input channels. |
|
out_channels (int): Output channels. |
|
kernel_size (int): Kernel size. |
|
padding (int, optional): Padding. Defaults to 0. |
|
use_norm (bool, optional): Whether to use normalization. Defaults to False. |
|
bias (bool, optional): Whether to use bias. Defaults to True. |
|
""" |
|
super().__init__() |
|
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, padding=padding, bias=bias) |
|
self.bn = nn.BatchNorm2d(out_channels) if use_norm else nn.Identity() |
|
self.activate = nn.ReLU(inplace=True) |
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor: |
|
"""Forward pass.""" |
|
x = self.conv(x) |
|
x = self.bn(x) |
|
return self.activate(x) |
|
|
|
|
|
class ResidualConvUnit(nn.Module): |
|
"""Residual convolution module.""" |
|
|
|
def __init__(self, features): |
|
"""Simple residual convolution block. |
|
|
|
Args: |
|
features (int): number of features |
|
""" |
|
super().__init__() |
|
|
|
self.conv1 = nn.Conv2d(features, features, kernel_size=3, stride=1, padding=1, bias=True) |
|
self.conv2 = nn.Conv2d(features, features, kernel_size=3, stride=1, padding=1, bias=True) |
|
|
|
self.relu = torch.nn.ReLU(inplace=True) |
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor: |
|
"""Forward pass.""" |
|
out = self.relu(x) |
|
out = self.conv1(out) |
|
out = self.relu(out) |
|
out = self.conv2(out) |
|
return out + x |
|
|
|
|
|
class FeatureFusionBlock(nn.Module): |
|
"""Feature fusion block.""" |
|
|
|
def __init__(self, features: int, unit2only=False, upsample=True): |
|
"""Feature fusion block. |
|
|
|
Args: |
|
features (int): Number of features. |
|
unit2only (bool, optional): Whether to use only the second unit. Defaults to False. |
|
upsample (bool, optional): Whether to upsample. Defaults to True. |
|
""" |
|
super().__init__() |
|
self.upsample = upsample |
|
|
|
if not unit2only: |
|
self.resConfUnit1 = ResidualConvUnit(features) |
|
self.resConfUnit2 = ResidualConvUnit(features) |
|
|
|
def forward(self, *xs: torch.Tensor) -> torch.Tensor: |
|
"""Forward pass.""" |
|
output = xs[0] |
|
|
|
if len(xs) == 2: |
|
output = output + self.resConfUnit1(xs[1]) |
|
|
|
output = self.resConfUnit2(output) |
|
|
|
if self.upsample: |
|
output = F.interpolate(output, scale_factor=2, mode="bilinear", align_corners=False) |
|
|
|
return output |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
class NMF2D(nn.Module): |
|
"""Non-negative Matrix Factorization (NMF) for 2D data.""" |
|
|
|
def __init__(self): |
|
"""Non-negative Matrix Factorization (NMF) for 2D data.""" |
|
super().__init__() |
|
self.S, self.D, self.R = 1, 512, 64 |
|
self.train_steps = 6 |
|
self.eval_steps = 7 |
|
self.inv_t = 1 |
|
|
|
def _build_bases(self, B: int, S: int, D: int, R: int, device: str = "cpu") -> torch.Tensor: |
|
bases = torch.rand((B * S, D, R)).to(device) |
|
return F.normalize(bases, dim=1) |
|
|
|
def local_step( |
|
self, x: torch.Tensor, bases: torch.Tensor, coef: torch.Tensor |
|
) -> Tuple[torch.Tensor, torch.Tensor]: |
|
"""Update bases and coefficient.""" |
|
|
|
numerator = torch.bmm(x.transpose(1, 2), bases) |
|
|
|
denominator = coef.bmm(bases.transpose(1, 2).bmm(bases)) |
|
|
|
coef = coef * numerator / (denominator + 1e-6) |
|
|
|
numerator = torch.bmm(x, coef) |
|
|
|
denominator = bases.bmm(coef.transpose(1, 2).bmm(coef)) |
|
|
|
bases = bases * numerator / (denominator + 1e-6) |
|
return bases, coef |
|
|
|
def compute_coef( |
|
self, x: torch.Tensor, bases: torch.Tensor, coef: torch.Tensor |
|
) -> torch.Tensor: |
|
"""Compute coefficient.""" |
|
|
|
numerator = torch.bmm(x.transpose(1, 2), bases) |
|
|
|
denominator = coef.bmm(bases.transpose(1, 2).bmm(bases)) |
|
|
|
return coef * numerator / (denominator + 1e-6) |
|
|
|
def local_inference( |
|
self, x: torch.Tensor, bases: torch.Tensor |
|
) -> Tuple[torch.Tensor, torch.Tensor]: |
|
"""Local inference.""" |
|
|
|
coef = torch.bmm(x.transpose(1, 2), bases) |
|
coef = F.softmax(self.inv_t * coef, dim=-1) |
|
|
|
steps = self.train_steps if self.training else self.eval_steps |
|
for _ in range(steps): |
|
bases, coef = self.local_step(x, bases, coef) |
|
|
|
return bases, coef |
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor: |
|
"""Forward pass.""" |
|
B, C, H, W = x.shape |
|
|
|
|
|
D = C // self.S |
|
N = H * W |
|
x = x.view(B * self.S, D, N) |
|
|
|
|
|
bases = self._build_bases(B, self.S, D, self.R, device=x.device) |
|
bases, coef = self.local_inference(x, bases) |
|
|
|
coef = self.compute_coef(x, bases, coef) |
|
|
|
x = torch.bmm(bases, coef.transpose(1, 2)) |
|
|
|
x = x.view(B, C, H, W) |
|
|
|
bases = bases.view(B, self.S, D, self.R) |
|
|
|
return x |
|
|
|
|
|
class Hamburger(nn.Module): |
|
"""Hamburger Module.""" |
|
|
|
def __init__(self, ham_channels: int = 512): |
|
"""Hambuger Module. |
|
|
|
Args: |
|
ham_channels (int, optional): Number of channels in the hamburger module. Defaults to |
|
512. |
|
""" |
|
super().__init__() |
|
self.ham_in = ConvModule(ham_channels, ham_channels, 1) |
|
self.ham = NMF2D() |
|
self.ham_out = ConvModule(ham_channels, ham_channels, 1) |
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor: |
|
"""Forward pass.""" |
|
enjoy = self.ham_in(x) |
|
enjoy = F.relu(enjoy, inplace=False) |
|
enjoy = self.ham(enjoy) |
|
enjoy = self.ham_out(enjoy) |
|
ham = F.relu(x + enjoy, inplace=False) |
|
return ham |
|
|
|
|
|
class LightHamHead(nn.Module): |
|
"""Is Attention Better Than Matrix Decomposition? |
|
|
|
This head is the implementation of `HamNet <https://arxiv.org/abs/2109.04553>`. |
|
""" |
|
|
|
def __init__(self): |
|
"""Light hamburger decoder head.""" |
|
super().__init__() |
|
self.in_index = [0, 1, 2, 3] |
|
self.in_channels = [64, 128, 320, 512] |
|
self.out_channels = 64 |
|
self.ham_channels = 512 |
|
self.align_corners = False |
|
|
|
self.squeeze = ConvModule(sum(self.in_channels), self.ham_channels, 1) |
|
|
|
self.hamburger = Hamburger(self.ham_channels) |
|
|
|
self.align = ConvModule(self.ham_channels, self.out_channels, 1) |
|
|
|
self.linear_pred_uncertainty = nn.Sequential( |
|
ConvModule( |
|
in_channels=self.out_channels, |
|
out_channels=self.out_channels, |
|
kernel_size=3, |
|
padding=1, |
|
bias=False, |
|
), |
|
nn.Conv2d(in_channels=self.out_channels, out_channels=1, kernel_size=1), |
|
) |
|
|
|
self.out_conv = ConvModule(self.out_channels, self.out_channels, 3, padding=1, bias=False) |
|
self.ll_fusion = FeatureFusionBlock(self.out_channels, upsample=False) |
|
|
|
def forward(self, features: Dict[str, torch.Tensor]) -> Tuple[torch.Tensor, torch.Tensor]: |
|
"""Forward pass.""" |
|
inputs = [features["hl"][i] for i in self.in_index] |
|
|
|
inputs = [ |
|
F.interpolate( |
|
level, size=inputs[0].shape[2:], mode="bilinear", align_corners=self.align_corners |
|
) |
|
for level in inputs |
|
] |
|
|
|
inputs = torch.cat(inputs, dim=1) |
|
x = self.squeeze(inputs) |
|
|
|
x = self.hamburger(x) |
|
|
|
feats = self.align(x) |
|
|
|
assert "ll" in features, "Low-level features are required for this model" |
|
feats = F.interpolate(feats, scale_factor=2, mode="bilinear", align_corners=False) |
|
feats = self.out_conv(feats) |
|
feats = F.interpolate(feats, scale_factor=2, mode="bilinear", align_corners=False) |
|
feats = self.ll_fusion(feats, features["ll"].clone()) |
|
|
|
uncertainty = self.linear_pred_uncertainty(feats).squeeze(1) |
|
|
|
return feats, uncertainty |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
class DWConv(nn.Module): |
|
"""Depthwise convolution.""" |
|
|
|
def __init__(self, dim: int = 768): |
|
"""Depthwise convolution. |
|
|
|
Args: |
|
dim (int, optional): Number of features. Defaults to 768. |
|
""" |
|
super().__init__() |
|
self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim) |
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor: |
|
"""Forward pass.""" |
|
return self.dwconv(x) |
|
|
|
|
|
class Mlp(nn.Module): |
|
"""MLP module.""" |
|
|
|
def __init__( |
|
self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.0 |
|
): |
|
"""Initialize the MLP.""" |
|
super().__init__() |
|
out_features = out_features or in_features |
|
hidden_features = hidden_features or in_features |
|
self.fc1 = nn.Conv2d(in_features, hidden_features, 1) |
|
self.dwconv = DWConv(hidden_features) |
|
self.act = act_layer() |
|
self.fc2 = nn.Conv2d(hidden_features, out_features, 1) |
|
self.drop = nn.Dropout(drop) |
|
|
|
def forward(self, x): |
|
"""Forward pass.""" |
|
x = self.fc1(x) |
|
|
|
x = self.dwconv(x) |
|
x = self.act(x) |
|
x = self.drop(x) |
|
x = self.fc2(x) |
|
x = self.drop(x) |
|
|
|
return x |
|
|
|
|
|
class StemConv(nn.Module): |
|
"""Simple stem convolution module.""" |
|
|
|
def __init__(self, in_channels: int, out_channels: int): |
|
"""Simple stem convolution module. |
|
|
|
Args: |
|
in_channels (int): Input channels. |
|
out_channels (int): Output channels. |
|
""" |
|
super().__init__() |
|
|
|
self.proj = nn.Sequential( |
|
nn.Conv2d( |
|
in_channels, out_channels // 2, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1) |
|
), |
|
nn.BatchNorm2d(out_channels // 2), |
|
nn.GELU(), |
|
nn.Conv2d( |
|
out_channels // 2, out_channels, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1) |
|
), |
|
nn.BatchNorm2d(out_channels), |
|
) |
|
|
|
def forward(self, x): |
|
"""Forward pass.""" |
|
x = self.proj(x) |
|
_, _, H, W = x.size() |
|
x = x.flatten(2).transpose(1, 2) |
|
return x, H, W |
|
|
|
|
|
class AttentionModule(nn.Module): |
|
"""Attention module.""" |
|
|
|
def __init__(self, dim: int): |
|
"""Attention module. |
|
|
|
Args: |
|
dim (int): Number of features. |
|
""" |
|
super().__init__() |
|
self.conv0 = nn.Conv2d(dim, dim, 5, padding=2, groups=dim) |
|
self.conv0_1 = nn.Conv2d(dim, dim, (1, 7), padding=(0, 3), groups=dim) |
|
self.conv0_2 = nn.Conv2d(dim, dim, (7, 1), padding=(3, 0), groups=dim) |
|
|
|
self.conv1_1 = nn.Conv2d(dim, dim, (1, 11), padding=(0, 5), groups=dim) |
|
self.conv1_2 = nn.Conv2d(dim, dim, (11, 1), padding=(5, 0), groups=dim) |
|
|
|
self.conv2_1 = nn.Conv2d(dim, dim, (1, 21), padding=(0, 10), groups=dim) |
|
self.conv2_2 = nn.Conv2d(dim, dim, (21, 1), padding=(10, 0), groups=dim) |
|
self.conv3 = nn.Conv2d(dim, dim, 1) |
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor: |
|
"""Forward pass.""" |
|
u = x.clone() |
|
attn = self.conv0(x) |
|
|
|
attn_0 = self.conv0_1(attn) |
|
attn_0 = self.conv0_2(attn_0) |
|
|
|
attn_1 = self.conv1_1(attn) |
|
attn_1 = self.conv1_2(attn_1) |
|
|
|
attn_2 = self.conv2_1(attn) |
|
attn_2 = self.conv2_2(attn_2) |
|
attn = attn + attn_0 + attn_1 + attn_2 |
|
|
|
attn = self.conv3(attn) |
|
return attn * u |
|
|
|
|
|
class SpatialAttention(nn.Module): |
|
"""Spatial attention module.""" |
|
|
|
def __init__(self, dim: int): |
|
"""Spatial attention module. |
|
|
|
Args: |
|
dim (int): Number of features. |
|
""" |
|
super().__init__() |
|
self.d_model = dim |
|
self.proj_1 = nn.Conv2d(dim, dim, 1) |
|
self.activation = nn.GELU() |
|
self.spatial_gating_unit = AttentionModule(dim) |
|
self.proj_2 = nn.Conv2d(dim, dim, 1) |
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor: |
|
"""Forward pass.""" |
|
shorcut = x.clone() |
|
x = self.proj_1(x) |
|
x = self.activation(x) |
|
x = self.spatial_gating_unit(x) |
|
x = self.proj_2(x) |
|
x = x + shorcut |
|
return x |
|
|
|
|
|
class Block(nn.Module): |
|
"""MSCAN block.""" |
|
|
|
def __init__( |
|
self, dim: int, mlp_ratio: float = 4.0, drop: float = 0.0, act_layer: nn.Module = nn.GELU |
|
): |
|
"""MSCAN block. |
|
|
|
Args: |
|
dim (int): Number of features. |
|
mlp_ratio (float, optional): Ratio of the hidden features in the MLP. Defaults to 4.0. |
|
drop (float, optional): Dropout rate. Defaults to 0.0. |
|
act_layer (nn.Module, optional): Activation layer. Defaults to nn.GELU. |
|
""" |
|
super().__init__() |
|
self.norm1 = nn.BatchNorm2d(dim) |
|
self.attn = SpatialAttention(dim) |
|
self.drop_path = nn.Identity() |
|
self.norm2 = nn.BatchNorm2d(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 |
|
) |
|
layer_scale_init_value = 1e-2 |
|
self.layer_scale_1 = nn.Parameter( |
|
layer_scale_init_value * torch.ones((dim)), requires_grad=True |
|
) |
|
self.layer_scale_2 = nn.Parameter( |
|
layer_scale_init_value * torch.ones((dim)), requires_grad=True |
|
) |
|
|
|
def forward(self, x: torch.Tensor, H: int, W: int) -> torch.Tensor: |
|
"""Forward pass.""" |
|
B, N, C = x.shape |
|
x = x.permute(0, 2, 1).view(B, C, H, W) |
|
x = x + self.drop_path(self.layer_scale_1[..., None, None] * self.attn(self.norm1(x))) |
|
x = x + self.drop_path(self.layer_scale_2[..., None, None] * self.mlp(self.norm2(x))) |
|
return x.view(B, C, N).permute(0, 2, 1) |
|
|
|
|
|
class OverlapPatchEmbed(nn.Module): |
|
"""Image to Patch Embedding""" |
|
|
|
def __init__( |
|
self, patch_size: int = 7, stride: int = 4, in_chans: int = 3, embed_dim: int = 768 |
|
): |
|
"""Image to Patch Embedding. |
|
|
|
Args: |
|
patch_size (int, optional): Image patch size. Defaults to 7. |
|
stride (int, optional): Stride. Defaults to 4. |
|
in_chans (int, optional): Number of input channels. Defaults to 3. |
|
embed_dim (int, optional): Embedding dimension. Defaults to 768. |
|
""" |
|
super().__init__() |
|
patch_size = to_2tuple(patch_size) |
|
|
|
self.proj = nn.Conv2d( |
|
in_chans, |
|
embed_dim, |
|
kernel_size=patch_size, |
|
stride=stride, |
|
padding=(patch_size[0] // 2, patch_size[1] // 2), |
|
) |
|
self.norm = nn.BatchNorm2d(embed_dim) |
|
|
|
def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, int, int]: |
|
"""Forward pass.""" |
|
x = self.proj(x) |
|
_, _, H, W = x.shape |
|
x = self.norm(x) |
|
x = x.flatten(2).transpose(1, 2) |
|
return x, H, W |
|
|
|
|
|
class MSCAN(nn.Module): |
|
"""Multi-scale convolutional attention network.""" |
|
|
|
def __init__(self): |
|
"""Multi-scale convolutional attention network.""" |
|
super().__init__() |
|
self.in_channels = 3 |
|
self.embed_dims = [64, 128, 320, 512] |
|
self.mlp_ratios = [8, 8, 4, 4] |
|
self.drop_rate = 0.0 |
|
self.drop_path_rate = 0.1 |
|
self.depths = [3, 3, 12, 3] |
|
self.num_stages = 4 |
|
|
|
for i in range(self.num_stages): |
|
if i == 0: |
|
patch_embed = StemConv(3, self.embed_dims[0]) |
|
else: |
|
patch_embed = OverlapPatchEmbed( |
|
patch_size=7 if i == 0 else 3, |
|
stride=4 if i == 0 else 2, |
|
in_chans=self.in_chans if i == 0 else self.embed_dims[i - 1], |
|
embed_dim=self.embed_dims[i], |
|
) |
|
|
|
block = nn.ModuleList( |
|
[ |
|
Block( |
|
dim=self.embed_dims[i], |
|
mlp_ratio=self.mlp_ratios[i], |
|
drop=self.drop_rate, |
|
) |
|
for _ in range(self.depths[i]) |
|
] |
|
) |
|
norm = nn.LayerNorm(self.embed_dims[i]) |
|
|
|
setattr(self, f"patch_embed{i + 1}", patch_embed) |
|
setattr(self, f"block{i + 1}", block) |
|
setattr(self, f"norm{i + 1}", norm) |
|
|
|
def forward(self, data): |
|
"""Forward pass.""" |
|
|
|
x = data["image"][:, [2, 1, 0], :, :] * 255.0 |
|
B = x.shape[0] |
|
|
|
outs = [] |
|
for i in range(self.num_stages): |
|
patch_embed = getattr(self, f"patch_embed{i + 1}") |
|
block = getattr(self, f"block{i + 1}") |
|
norm = getattr(self, f"norm{i + 1}") |
|
x, H, W = patch_embed(x) |
|
for blk in block: |
|
x = blk(x, H, W) |
|
x = norm(x) |
|
x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous() |
|
outs.append(x) |
|
|
|
return {"features": outs} |
|
|