scripts/models/resnext.py

# Copyright (c) Facebook, Inc. and its affiliates.

import sys
from collections import OrderedDict
from functools import partial

import torch.nn as nn
from inplace_abn import ABN
from modules import IdentityResidualBlock, GlobalAvgPool2d

from .util import try_index


class ResNeXt(nn.Module):
    def __init__(
        self,
        structure,
        groups=64,
        norm_act=ABN,
        input_3x3=False,
        classes=0,
        dilation=1,
        base_channels=(128, 128, 256),
    ):
        """Pre-activation (identity mapping) ResNeXt model

        Parameters
        ----------
        structure : list of int
            Number of residual blocks in each of the four modules of the network.
        groups : int
            Number of groups in each ResNeXt block
        norm_act : callable
            Function to create normalization / activation Module.
        input_3x3 : bool
            If `True` use three `3x3` convolutions in the input module instead of a single `7x7` one.
        classes : int
            If not `0` also include global average pooling and a fully-connected layer with `classes` outputs at the end
            of the network.
        dilation : list of list of int or list of int or int
            List of dilation factors, or `1` to ignore dilation. For each module, if a single value is given it is
            used for all its blocks, otherwise this expects a value for each block.
        base_channels : list of int
            Channels in the blocks of the first residual module. Each following module will multiply these values by 2.
        """
        super(ResNeXt, self).__init__()
        self.structure = structure

        if len(structure) != 4:
            raise ValueError("Expected a structure with four values")
        if dilation != 1 and len(dilation) != 4:
            raise ValueError("If dilation is not 1 it must contain four values")

        # Initial layers
        if input_3x3:
            layers = [
                ("conv1", nn.Conv2d(3, 64, 3, stride=2, padding=1, bias=False)),
                ("bn1", norm_act(64)),
                ("conv2", nn.Conv2d(64, 64, 3, stride=1, padding=1, bias=False)),
                ("bn2", norm_act(64)),
                ("conv3", nn.Conv2d(64, 64, 3, stride=1, padding=1, bias=False)),
                ("pool", nn.MaxPool2d(3, stride=2, padding=1)),
            ]
        else:
            layers = [
                ("conv1", nn.Conv2d(3, 64, 7, stride=2, padding=3, bias=False)),
                ("pool", nn.MaxPool2d(3, stride=2, padding=1)),
            ]
        self.mod1 = nn.Sequential(OrderedDict(layers))

        # Groups of residual blocks
        in_channels = 64
        channels = base_channels
        for mod_id, num in enumerate(structure):
            # Create blocks for module
            blocks = []
            for block_id in range(num):
                s, d = self._stride_dilation(mod_id, block_id, dilation)
                blocks.append(
                    (
                        "block%d" % (block_id + 1),
                        IdentityResidualBlock(
                            in_channels,
                            channels,
                            stride=s,
                            norm_act=norm_act,
                            groups=groups,
                            dilation=d,
                        ),
                    )
                )

                # Update channels
                in_channels = channels[-1]

            # Create and add module
            self.add_module("mod%d" % (mod_id + 2), nn.Sequential(OrderedDict(blocks)))
            channels = [c * 2 for c in channels]

        # Pooling and predictor
        self.bn_out = norm_act(in_channels)
        if classes != 0:
            self.classifier = nn.Sequential(
                OrderedDict(
                    [
                        ("avg_pool", GlobalAvgPool2d()),
                        ("fc", nn.Linear(in_channels, classes)),
                    ]
                )
            )

    def forward(self, img):
        out = self.mod1(img)
        out = self.mod2(out)
        out = self.mod3(out)
        out = self.mod4(out)
        out = self.mod5(out)
        out = self.bn_out(out)

        if hasattr(self, "classifier"):
            out = self.classifier(out)

        return out

    @staticmethod
    def _stride_dilation(mod_id, block_id, dilation):
        if dilation == 1:
            s = 2 if mod_id > 0 and block_id == 0 else 1
            d = 1
        else:
            if dilation[mod_id] == 1:
                s = 2 if mod_id > 0 and block_id == 0 else 1
                d = 1
            else:
                s = 1
                d = try_index(dilation[mod_id], block_id)
        return s, d


_NETS = {
    "50": {"structure": [3, 4, 6, 3]},
    "101": {"structure": [3, 4, 23, 3]},
    "152": {"structure": [3, 8, 36, 3]},
}

__all__ = []
for name, params in _NETS.items():
    net_name = "net_resnext" + name
    setattr(sys.modules[__name__], net_name, partial(ResNeXt, **params))
    __all__.append(net_name)