chaiNNer-org#305 Add support for MoSR

README.md

@@ -106,6 +106,7 @@ Spandrel currently supports a limited amount of network architectures. If the ar
 - [DRCT](https://github.com/ming053l/DRCT)
 - [PLKSR](https://github.com/dslisleedh/PLKSR) and [RealPLKSR](https://github.com/muslll/neosr/blob/master/neosr/archs/realplksr_arch.py) | [Models](https://drive.google.com/drive/u/1/folders/1lIkZ00y9cRQpLU9qmCIB2XtS-2ZoqKq8)
 - [SeemoRe](https://github.com/eduardzamfir/seemoredetails) | [Models](https://drive.google.com/drive/folders/15jtvcS4jL_6QqEwaRodEN8FBrqVPrO2u?usp=share_link)
+- [MoSR](https://github.com/umzi2/MoSR) | [Models](https://drive.google.com/drive/u/0/folders/1HPy7M4Zzq8oxhdsQ2cnfqy73klmQWp_r)
 
 #### Face Restoration
 

libs/spandrel/spandrel/__helpers/main_registry.py

@@ -27,6 +27,7 @@
     LaMa,
     MixDehazeNet,
     MMRealSR,
+    MoSR,
     NAFNet,
     OmniSR,
     RealCUGAN,
@@ -86,4 +87,5 @@
     ArchSupport.from_architecture(RetinexFormer.RetinexFormerArch()),
     ArchSupport.from_architecture(HVICIDNet.HVICIDNetArch()),
     ArchSupport.from_architecture(SeemoRe.SeemoReArch()),
+    ArchSupport.from_architecture(MoSR.MoSRArch()),
 )

libs/spandrel/spandrel/architectures/MoSR/__arch/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 umzi
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

libs/spandrel/spandrel/architectures/MoSR/__arch/mosr_arch.py

@@ -0,0 +1,201 @@
+import torch
+from torch import nn
+from torch.nn.init import trunc_normal_
+
+from spandrel.architectures.__arch_helpers.dysample import DySample
+from spandrel.util import store_hyperparameters
+from spandrel.util.timm.__drop import DropPath
+
+
+class GPS(nn.Module):
+    """Geo ensemble PixelShuffle"""
+
+    def __init__(
+        self,
+        dim,
+        scale,
+        out_ch=3,
+        # Own parameters
+        kernel_size: int = 3,
+    ):
+        super().__init__()
+        self.in_to_k = nn.Conv2d(
+            dim, scale * scale * out_ch * 8, kernel_size, 1, kernel_size // 2
+        )
+        self.ps = nn.PixelShuffle(scale)
+
+    def forward(self, x):
+        rgb = self._geo_ensemble(x)
+        rgb = self.ps(rgb)
+        return rgb
+
+    def _geo_ensemble(self, x):
+        x = self.in_to_k(x)
+        x = x.reshape(x.shape[0], 8, -1, x.shape[-2], x.shape[-1])
+        x = x.mean(dim=1)
+        return x
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(dim))
+        self.bias = nn.Parameter(torch.zeros(dim))
+        self.eps = eps
+
+    def forward(self, x):
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        return self.weight[:, None, None] * x + self.bias[:, None, None]
+
+
+class ConvBlock(nn.Module):
+    r"""https://github.com/joshyZhou/AST/blob/main/model.py#L22"""
+
+    def __init__(self, in_channel: int, out_channel: int, strides: int = 1):
+        super().__init__()
+        self.strides = strides
+        self.in_channel = in_channel
+        self.out_channel = out_channel
+        self.block = nn.Sequential(
+            nn.Conv2d(
+                in_channel, out_channel, kernel_size=3, stride=strides, padding=1
+            ),
+            nn.Mish(),
+            nn.Conv2d(
+                out_channel, out_channel, kernel_size=3, stride=strides, padding=1
+            ),
+            nn.Mish(),
+        )
+        self.conv11 = nn.Conv2d(
+            in_channel, out_channel, kernel_size=1, stride=strides, padding=0
+        )
+
+    def forward(self, x):
+        out1 = self.block(x)
+        out2 = self.conv11(x)
+        out = out1 + out2
+        return out
+
+
+class GatedCNNBlock(nn.Module):
+    r"""
+    modernized mambaout main unit
+    https://github.com/yuweihao/MambaOut/blob/main/models/mambaout.py#L119
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        expansion_ratio: float = 8 / 3,
+        conv_ratio: float = 1.0,
+        kernel_size: int = 7,
+        drop_path: float = 0.5,
+    ):
+        super().__init__()
+        self.norm = LayerNorm(dim)
+        hidden = int(expansion_ratio * dim)
+        self.fc1 = nn.Conv2d(dim, hidden * 2, 3, 1, 1)
+
+        self.act = nn.Mish()
+        conv_channels = int(conv_ratio * dim)
+        self.split_indices = [hidden, hidden - conv_channels, conv_channels]
+
+        self.conv = nn.Conv2d(
+            conv_channels,
+            conv_channels,
+            kernel_size,
+            1,
+            kernel_size // 2,
+            groups=conv_channels,
+        )
+        self.fc2 = nn.Conv2d(hidden, dim, 3, 1, 1)
+        self.drop_path = (
+            DropPath(drop_path)
+            if drop_path > 0.0 or not self.training
+            else nn.Identity()
+        )
+        self.apply(self._init_weights)
+
+    @staticmethod
+    def _init_weights(m):
+        if isinstance(m, nn.Conv2d | nn.Linear):
+            trunc_normal_(m.weight, std=0.02)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        shortcut = x
+        x = self.norm(x)
+        g, i, c = torch.split(self.fc1(x), self.split_indices, dim=1)
+        c = self.conv(c)
+        x = self.act(self.fc2(self.act(g) * torch.cat((i, c), dim=1)))
+        x = self.drop_path(x)
+        return x + (shortcut - 0.5)
+
+
+@store_hyperparameters()
+class MoSR(nn.Module):
+    """Mamba Out Super-Resolution"""
+
+    hyperparameters = {}
+
+    def __init__(
+        self,
+        *,
+        in_ch: int = 3,
+        out_ch: int = 3,
+        upscale: int = 4,
+        n_block: int = 24,
+        dim: int = 64,
+        upsampler: str = "ps",  # "ps" "dys" "gps"
+        drop_path: float = 0.0,
+        kernel_size: int = 7,
+        expansion_ratio: float = 1.5,
+        conv_ratio: float = 1.0,
+    ):
+        super().__init__()
+        if upsampler in ["ps", "gps"]:
+            out_ch = in_ch
+        dp_rates = [x.item() for x in torch.linspace(0, drop_path, n_block)]
+        self.gblocks = nn.Sequential(
+            *[nn.Conv2d(in_ch, dim, 3, 1, 1)]
+            + [
+                GatedCNNBlock(
+                    dim=dim,
+                    expansion_ratio=expansion_ratio,
+                    kernel_size=kernel_size,
+                    conv_ratio=conv_ratio,
+                    drop_path=dp_rates[index],
+                )
+                for index in range(n_block)
+            ]
+            + [
+                nn.Conv2d(dim, dim * 2, 3, 1, 1),
+                nn.Mish(),
+                nn.Conv2d(dim * 2, dim, 3, 1, 1),
+                nn.Mish(),
+                nn.Conv2d(dim, dim, 1, 1),
+            ]
+        )
+
+        self.shortcut = ConvBlock(in_ch, dim)
+
+        if upsampler == "ps":
+            self.upsampler = nn.Sequential(
+                nn.Conv2d(dim, out_ch * (upscale**2), 3, 1, 1), nn.PixelShuffle(upscale)
+            )
+        elif upsampler == "gps":
+            self.upsampler = GPS(dim, upscale, out_ch)
+        elif upsampler == "dys":
+            self.upsampler = DySample(dim, out_ch, upscale)
+        else:
+            raise ValueError(
+                f'upsampler: {upsampler} not supported, choose one of these options: \
+                ["ps", "gps", "dys"]'
+            )
+
+    def forward(self, x):
+        x = self.gblocks(x) + (self.shortcut(x) - 0.5)
+        return self.upsampler(x)

libs/spandrel/spandrel/architectures/MoSR/__init__.py

@@ -0,0 +1,97 @@
+import math
+
+from typing_extensions import override
+
+from spandrel.util import KeyCondition, get_seq_len
+
+from ...__helpers.model_descriptor import Architecture, ImageModelDescriptor, StateDict
+from .__arch.mosr_arch import MoSR
+
+
+class MoSRArch(Architecture[MoSR]):
+    def __init__(self) -> None:
+        super().__init__(
+            id="MoSR",
+            detect=KeyCondition.has_all(
+                "gblocks.0.weight",
+                "gblocks.0.bias",
+                "gblocks.1.norm.weight",
+                "gblocks.1.norm.bias",
+                "gblocks.1.fc1.weight",
+                "gblocks.1.fc1.bias",
+                "gblocks.1.conv.weight",
+                "gblocks.1.conv.bias",
+                "gblocks.1.fc2.weight",
+                "gblocks.1.fc2.bias",
+            ),
+        )
+
+    @override
+    def load(self, state_dict: StateDict) -> ImageModelDescriptor[MoSR]:
+        # default values
+        in_ch = 3
+        out_ch = 3
+        upscale = 4
+        n_block = 24
+        dim = 64
+        upsampler = "ps"  # "ps" "dys", "gps"
+        drop_path = 0.0
+        kernel_size = 7
+        expansion_ratio = 1.5
+        conv_ratio = 1.0
+
+        n_block = get_seq_len(state_dict, "gblocks") - 6
+        in_ch = state_dict["gblocks.0.weight"].shape[1]
+        dim = state_dict["gblocks.0.weight"].shape[0]
+
+        # Calculate expansion ratio and convolution ratio
+        expansion_ratio = (
+            state_dict["gblocks.1.fc1.weight"].shape[0]
+            / state_dict["gblocks.1.fc1.weight"].shape[1]
+        ) / 2
+        conv_ratio = state_dict["gblocks.1.conv.weight"].shape[0] / dim
+        kernel_size = state_dict["gblocks.1.conv.weight"].shape[2]
+        # Determine upsampler type and calculate upscale
+        if "upsampler.init_pos" in state_dict:
+            upsampler = "dys"
+            out_ch = state_dict["upsampler.end_conv.weight"].shape[0]
+            upscale = math.isqrt(state_dict["upsampler.offset.weight"].shape[0] // 8)
+        elif "upsampler.in_to_k.weight" in state_dict:
+            upsampler = "gps"
+            out_ch = in_ch
+            upscale = math.isqrt(
+                state_dict["upsampler.in_to_k.weight"].shape[0] // 8 // out_ch
+            )
+        else:
+            upsampler = "ps"
+            out_ch = in_ch
+            upscale = math.isqrt(state_dict["upsampler.0.weight"].shape[0] // out_ch)
+
+        model = MoSR(
+            in_ch=in_ch,
+            out_ch=out_ch,
+            upscale=upscale,
+            n_block=n_block,
+            dim=dim,
+            upsampler=upsampler,
+            drop_path=drop_path,
+            kernel_size=kernel_size,
+            expansion_ratio=expansion_ratio,
+            conv_ratio=conv_ratio,
+        )
+
+        return ImageModelDescriptor(
+            model,
+            state_dict,
+            architecture=self,
+            purpose="SR",
+            tags=[],
+            supports_half=True,
+            supports_bfloat16=True,
+            scale=upscale,
+            input_channels=in_ch,
+            output_channels=out_ch,
+        )
+
+
+__all__ = ["MoSRArch", "MoSR"]

tests/__snapshots__/test_MoSR.ambr

@@ -0,0 +1,37 @@
+# serializer version: 1
+# name: test_2x_nomos2_mosr
+  ImageModelDescriptor(
+    architecture=MoSRArch(
+      id='MoSR',
+      name='MoSR',
+    ),
+    input_channels=3,
+    output_channels=3,
+    purpose='SR',
+    scale=2,
+    size_requirements=SizeRequirements(minimum=0, multiple_of=1, square=False),
+    supports_bfloat16=True,
+    supports_half=True,
+    tags=list([
+    ]),
+    tiling=<ModelTiling.SUPPORTED: 1>,
+  )
+# ---
+# name: test_4x_nomos2_mosr
+  ImageModelDescriptor(
+    architecture=MoSRArch(
+      id='MoSR',
+      name='MoSR',
+    ),
+    input_channels=3,
+    output_channels=3,
+    purpose='SR',
+    scale=4,
+    size_requirements=SizeRequirements(minimum=0, multiple_of=1, square=False),
+    supports_bfloat16=True,
+    supports_half=True,
+    tags=list([
+    ]),
+    tiling=<ModelTiling.SUPPORTED: 1>,
+  )
+# ---