reconstruction/models/convnext.py

# taken from torchvision.models.convnext

from functools import partial
from typing import Any, Callable, List, Optional, Sequence

import torch
from torch import nn, Tensor
from torch.nn import functional as F

from torchvision.ops.misc import Conv2dNormActivation, Permute
from torchvision.ops.stochastic_depth import StochasticDepth
from torchvision.models._api import Weights, WeightsEnum



__all__ = [
    "ConvNeXt",
    "ConvNeXt_Tiny_Weights",
    "ConvNeXt_Small_Weights",
    "ConvNeXt_Base_Weights",
    "ConvNeXt_Large_Weights",
    "convnext_tiny",
    "convnext_small",
    "convnext_base",
    "convnext_large",
]


class LayerNorm2d(nn.LayerNorm):
    def forward(self, x: Tensor) -> Tensor:
        x = x.permute(0, 2, 3, 1)
        x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
        x = x.permute(0, 3, 1, 2)
        return x


class CNBlock(nn.Module):
    def __init__(
        self,
        dim,
        layer_scale: float,
        stochastic_depth_prob: float,
        norm_layer: Optional[Callable[..., nn.Module]] = None,
    ) -> None:
        super().__init__()
        if norm_layer is None:
            norm_layer = partial(nn.LayerNorm, eps=1e-6)

        self.block = nn.Sequential(
            nn.Conv2d(dim, dim, kernel_size=7, padding=3, groups=dim, bias=True),
            Permute([0, 2, 3, 1]),
            norm_layer(dim),
            nn.Linear(in_features=dim, out_features=4 * dim, bias=True),
            nn.GELU(),
            nn.Linear(in_features=4 * dim, out_features=dim, bias=True),
            Permute([0, 3, 1, 2]),
        )
        self.layer_scale = nn.Parameter(torch.ones(dim, 1, 1) * layer_scale)
        self.stochastic_depth = StochasticDepth(stochastic_depth_prob, "row")

    def forward(self, input: Tensor) -> Tensor:
        result = self.layer_scale * self.block(input)
        result = self.stochastic_depth(result)
        result += input
        return result


class CNBlockConfig:
    # Stores information listed at Section 3 of the ConvNeXt paper
    def __init__(
        self,
        input_channels: int,
        out_channels: Optional[int],
        num_layers: int,
    ) -> None:
        self.input_channels = input_channels
        self.out_channels = out_channels
        self.num_layers = num_layers

    def __repr__(self) -> str:
        s = self.__class__.__name__ + "("
        s += "input_channels={input_channels}"
        s += ", out_channels={out_channels}"
        s += ", num_layers={num_layers}"
        s += ")"
        return s.format(**self.__dict__)


class ConvNeXtBase(nn.Module):
    def __init__(
        self,
        block_setting: List[CNBlockConfig],
        stochastic_depth_prob: float = 0.0,
        layer_scale: float = 1e-6,
        block: Optional[Callable[..., nn.Module]] = None,
        norm_layer: Optional[Callable[..., nn.Module]] = None
    ) -> None:
        super().__init__()

        if not block_setting:
            raise ValueError("The block_setting should not be empty")
        elif not (isinstance(block_setting, Sequence) and all([isinstance(s, CNBlockConfig) for s in block_setting])):
            raise TypeError("The block_setting should be List[CNBlockConfig]")

        if block is None:
            block = CNBlock

        if norm_layer is None:
            norm_layer = partial(LayerNorm2d, eps=1e-6)

        layers: List[nn.Module] = []

        # Stem
        firstconv_output_channels = block_setting[0].input_channels
        layers.append(
            Conv2dNormActivation(
                3,
                firstconv_output_channels,
                kernel_size=4,
                stride=4,
                padding=0,
                norm_layer=norm_layer,
                activation_layer=None,
                bias=True,
            )
        )

        total_stage_blocks = sum(cnf.num_layers for cnf in block_setting)
        stage_block_id = 0
        for cnf in block_setting:
            # Bottlenecks
            stage: List[nn.Module] = []
            for _ in range(cnf.num_layers):
                # adjust stochastic depth probability based on the depth of the stage block
                sd_prob = stochastic_depth_prob * stage_block_id / (total_stage_blocks - 1.0)
                stage.append(block(cnf.input_channels, layer_scale, sd_prob))
                stage_block_id += 1
            layers.append(nn.Sequential(*stage))
            if cnf.out_channels is not None:
                # Downsampling
                layers.append(
                    nn.Sequential(
                        norm_layer(cnf.input_channels),
                        nn.Conv2d(cnf.input_channels, cnf.out_channels, kernel_size=2, stride=2),
                    )
                )

        self.features = nn.Sequential(*layers)

        for m in self.modules():
            if isinstance(m, (nn.Conv2d, nn.Linear)):
                nn.init.trunc_normal_(m.weight, std=0.02)
                if m.bias is not None:
                    nn.init.zeros_(m.bias)

    def _forward_impl(self, x: Tensor) -> Tensor:
        x = self.features(x)
        return x

    def forward(self, x: Tensor) -> Tensor:
        return self._forward_impl(x)



class ConvNeXtReverseBase(nn.Module):
    def __init__(
        self,
        block_setting: List[CNBlockConfig],
        stochastic_depth_prob: float = 0.0,
        layer_scale: float = 1e-6,
        block: Optional[Callable[..., nn.Module]] = None,
        norm_layer: Optional[Callable[..., nn.Module]] = None
    ) -> None:
        super().__init__()

        if not block_setting:
            raise ValueError("The block_setting should not be empty")
        elif not (isinstance(block_setting, Sequence) and all([isinstance(s, CNBlockConfig) for s in block_setting])):
            raise TypeError("The block_setting should be List[CNBlockConfig]")

        if block is None:
            block = CNBlock

        if norm_layer is None:
            norm_layer = partial(LayerNorm2d, eps=1e-6)

        layers: List[nn.Module] = []


        # add bottleneck which will be executed at the end, to process after the final larger upsampling
        stage: List[nn.Module] = []
        for _ in range(3):
            stage.append(block(1, layer_scale, 0.0))
        layers.append(nn.Sequential(*stage))


        # Stem
        firstconv_output_channels = block_setting[0].input_channels
        layers.append(
            # Convtr NormActivation(
            #     firstconv_output_channels,
            #     3,
            #     kernel_size=4,
            #     stride=4,
            #     padding=0,
            #     norm_layer=None,
            #     activation_layer=None,
            #     bias=True,
            # )
            nn.Sequential(
                nn.ConvTranspose2d(
                    firstconv_output_channels,
                    1,
                    kernel_size=4,
                    stride=4,
                    padding=0,
                    ),
            )
        )

        total_stage_blocks = sum(cnf.num_layers for cnf in block_setting)
        stage_block_id = 0
        for cnf in block_setting:
            # Bottlenecks
            stage: List[nn.Module] = []
            for _ in range(cnf.num_layers):
                # adjust stochastic depth probability based on the depth of the stage block
                sd_prob = stochastic_depth_prob * stage_block_id / (total_stage_blocks - 1.0)
                stage.append(block(cnf.input_channels, layer_scale, sd_prob))
                stage_block_id += 1
            layers.append(nn.Sequential(*stage))
            if cnf.out_channels is not None:
                # Upsampling
                layers.append(
                    nn.Sequential(
                        norm_layer(cnf.out_channels),
                        nn.ConvTranspose2d(cnf.out_channels, cnf.input_channels, kernel_size=2, stride=2),
                    )
                )

        self.features = nn.Sequential(*reversed(layers))

        for m in self.modules():
            if isinstance(m, (nn.Conv2d, nn.Linear)):
                nn.init.trunc_normal_(m.weight, std=0.02)
                if m.bias is not None:
                    nn.init.zeros_(m.bias)

    def _forward_impl(self, x: Tensor) -> Tensor:
        x = self.features(x)
        x = torch.tanh(x)
        return x

    def forward(self, x: Tensor) -> Tensor:
        return self._forward_impl(x)


def _convnext(
    block_setting: List[CNBlockConfig],
    stochastic_depth_prob: float,
    weights: Optional[WeightsEnum],
    reversed: bool,
    **kwargs: Any,
) -> ConvNeXtBase:

    if reversed:
        model = ConvNeXtReverseBase(block_setting, stochastic_depth_prob=stochastic_depth_prob, **kwargs)
    else:
        model = ConvNeXtBase(block_setting, stochastic_depth_prob=stochastic_depth_prob, **kwargs)

    if weights is not None:
        model.load_state_dict(weights.get_state_dict())

    return model


class ConvNeXt(nn.Module):
    def __init__(self, stochastic_depth_prob=0.0, reversed=False, **kwargs) -> None:
        super().__init__()


        block_setting = [
            CNBlockConfig(96, 192, 3),
            CNBlockConfig(192, 384, 3),
            CNBlockConfig(384, 768, 9),
            CNBlockConfig(768, None, 3),
        ]
        self.model = _convnext(block_setting, stochastic_depth_prob, None, reversed, **kwargs)

    def forward(self, x: Tensor) -> Tensor:
        x = self.model(x)
        return x
        


if __name__ == '__main__':
    model = ConvNeXt().to('cuda')
    test = model(torch.zeros((4, 3, 512, 256)).to('cuda'))

    pass