Source code for holocron.models.classification.sknet

# Copyright (C) 2020-2022, François-Guillaume Fernandez.

# This program is licensed under the Apache License version 2.
# See LICENSE or go to <https://www.apache.org/licenses/LICENSE-2.0.txt> for full license details.

from typing import Any, Callable, Dict, List, Optional

import torch
import torch.nn as nn

from holocron.nn import GlobalAvgPool2d

from ..presets import IMAGENETTE
from ..utils import conv_sequence, load_pretrained_params
from .resnet import ResNet, _ResBlock

__all__ = ['SoftAttentionLayer', 'SKConv2d', 'SKBottleneck', 'sknet50', 'sknet101', 'sknet152']


default_cfgs: Dict[str, Dict[str, Any]] = {
    'sknet50': {
        **IMAGENETTE,
        'input_shape': (3, 224, 224),
        'url': 'https://github.com/frgfm/Holocron/releases/download/v0.1.3/sknet50_224-5d2160f2.pth',
    },
    'sknet101': {
        **IMAGENETTE,
        'input_shape': (3, 224, 224),
        'url': None,
    },
    'sknet152': {
        **IMAGENETTE,
        'input_shape': (3, 224, 224),
        'url': None,
    },
}


class SoftAttentionLayer(nn.Sequential):

    def __init__(
        self,
        channels: int,
        sa_ratio: int = 16,
        out_multiplier: int = 1,
        act_layer: Optional[nn.Module] = None,
        norm_layer: Optional[Callable[[int], nn.Module]] = None,
        drop_layer: Optional[Callable[..., nn.Module]] = None
    ) -> None:
        super().__init__(
            GlobalAvgPool2d(flatten=False),
            *conv_sequence(channels, max(channels // sa_ratio, 32), act_layer, norm_layer, drop_layer,
                           kernel_size=1, stride=1, bias=(norm_layer is None)),
            *conv_sequence(max(channels // sa_ratio, 32), channels * out_multiplier,
                           nn.Sigmoid(), None, drop_layer, kernel_size=1, stride=1)
        )


class SKConv2d(nn.Module):

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        m: int = 2,
        sa_ratio: int = 16,
        act_layer: Optional[nn.Module] = None,
        norm_layer: Optional[Callable[[int], nn.Module]] = None,
        drop_layer: Optional[Callable[..., nn.Module]] = None,
        **kwargs: Any
    ) -> None:
        super().__init__()
        self.path_convs = nn.ModuleList([
            nn.Sequential(*conv_sequence(in_channels, out_channels, act_layer, norm_layer, drop_layer, kernel_size=3,
                                         bias=(norm_layer is None), dilation=idx + 1, padding=idx + 1, **kwargs))
            for idx in range(m)])
        self.sa = SoftAttentionLayer(out_channels, sa_ratio, m, act_layer, norm_layer, drop_layer)

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

        paths = torch.stack([path_conv(x) for path_conv in self.path_convs], dim=1)

        b, m, c = paths.shape[:3]
        z = self.sa(paths.sum(dim=1)).view(b, m, c, 1, 1)
        attention_factors = torch.softmax(z, dim=1)
        out = (attention_factors * paths).sum(dim=1)

        return out


class SKBottleneck(_ResBlock):

    expansion: int = 4

    def __init__(
        self,
        inplanes: int,
        planes: int,
        stride: int = 1,
        downsample: Optional[nn.Module] = None,
        groups: int = 32,
        base_width: int = 64,
        dilation: int = 1,
        act_layer: Optional[nn.Module] = None,
        norm_layer: Optional[Callable[[int], nn.Module]] = None,
        drop_layer: Optional[Callable[..., nn.Module]] = None,
        conv_layer: Optional[Callable[..., nn.Module]] = None,
        **kwargs: Any
    ) -> None:

        width = int(planes * (base_width / 64.)) * groups
        super().__init__(
            [*conv_sequence(inplanes, width, act_layer, norm_layer, drop_layer, conv_layer, kernel_size=1,
                            stride=1, bias=(norm_layer is None), **kwargs),
             SKConv2d(width, width, 2, 16, act_layer, norm_layer, drop_layer, groups=groups, stride=stride),
             *conv_sequence(width, planes * self.expansion, None, norm_layer, drop_layer, conv_layer, kernel_size=1,
                            stride=1, bias=(norm_layer is None), **kwargs)],
            downsample, act_layer)


def _sknet(
    arch: str,
    pretrained: bool,
    progress: bool,
    num_blocks: List[int],
    out_chans: List[int],
    **kwargs: Any,
) -> ResNet:

    # Build the model
    model = ResNet(SKBottleneck, num_blocks, out_chans, **kwargs)  # type: ignore[arg-type]
    model.default_cfg = default_cfgs[arch]  # type: ignore[assignment]
    # Load pretrained parameters
    if pretrained:
        load_pretrained_params(model, default_cfgs[arch]['url'], progress)

    return model


[docs] def sknet50(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: """SKNet-50 from `"Selective Kernel Networks" <https://arxiv.org/pdf/1903.06586.pdf>`_ Args: pretrained (bool): If True, returns a model pre-trained on ImageNet progress (bool): If True, displays a progress bar of the download to stderr Returns: torch.nn.Module: classification model """ return _sknet('sknet50', pretrained, progress, [3, 4, 6, 3], [64, 128, 256, 512], **kwargs)
[docs] def sknet101(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: """SKNet-101 from `"Selective Kernel Networks" <https://arxiv.org/pdf/1903.06586.pdf>`_ Args: pretrained (bool): If True, returns a model pre-trained on ImageNet progress (bool): If True, displays a progress bar of the download to stderr Returns: torch.nn.Module: classification model """ return _sknet('sknet101', pretrained, progress, [3, 4, 23, 3], [64, 128, 256, 512], **kwargs)
[docs] def sknet152(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: """SKNet-152 from `"Selective Kernel Networks" <https://arxiv.org/pdf/1903.06586.pdf>`_ Args: pretrained (bool): If True, returns a model pre-trained on ImageNet progress (bool): If True, displays a progress bar of the download to stderr Returns: torch.nn.Module: classification model """ return _sknet('sknet152', pretrained, progress, [3, 8, 86, 3], [64, 128, 256, 512], **kwargs)