pytorch-image-models/timm/models/dpn.py

""" PyTorch implementation of DualPathNetworks
Based on original MXNet implementation https://github.com/cypw/DPNs with
many ideas from another PyTorch implementation https://github.com/oyam/pytorch-DPNs.

This implementation is compatible with the pretrained weights
from cypw's MXNet implementation.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import torch
import torch.nn as nn
import torch.nn.functional as F
from collections import OrderedDict

from .registry import register_model
from .helpers import load_pretrained
from .adaptive_avgmax_pool import SelectAdaptivePool2d
from timm.data import IMAGENET_DPN_MEAN, IMAGENET_DPN_STD


__all__ = ['DPN']


def _cfg(url=''):
    return {
        'url': url, 'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': (7, 7),
        'crop_pct': 0.875, 'interpolation': 'bicubic',
        'mean': IMAGENET_DPN_MEAN, 'std': IMAGENET_DPN_STD,
        'first_conv': 'features.conv1_1.conv', 'classifier': 'classifier',
    }


default_cfgs = {
    'dpn68': _cfg(
        url='https://github.com/rwightman/pytorch-dpn-pretrained/releases/download/v0.1/dpn68-66bebafa7.pth'),
    'dpn68b': _cfg(
        url='https://github.com/rwightman/pytorch-dpn-pretrained/releases/download/v0.1/dpn68b_extra-84854c156.pth'),
    'dpn92': _cfg(
        url='https://github.com/rwightman/pytorch-dpn-pretrained/releases/download/v0.1/dpn92_extra-b040e4a9b.pth'),
    'dpn98': _cfg(
        url='https://github.com/rwightman/pytorch-dpn-pretrained/releases/download/v0.1/dpn98-5b90dec4d.pth'),
    'dpn131': _cfg(
        url='https://github.com/rwightman/pytorch-dpn-pretrained/releases/download/v0.1/dpn131-71dfe43e0.pth'),
    'dpn107': _cfg(
        url='https://github.com/rwightman/pytorch-dpn-pretrained/releases/download/v0.1/dpn107_extra-1ac7121e2.pth')
}


class CatBnAct(nn.Module):
    def __init__(self, in_chs, activation_fn=nn.ReLU(inplace=True)):
        super(CatBnAct, self).__init__()
        self.bn = nn.BatchNorm2d(in_chs, eps=0.001)
        self.act = activation_fn

    def forward(self, x):
        x = torch.cat(x, dim=1) if isinstance(x, tuple) else x
        return self.act(self.bn(x))


class BnActConv2d(nn.Module):
    def __init__(self, in_chs, out_chs, kernel_size, stride,
                 padding=0, groups=1, activation_fn=nn.ReLU(inplace=True)):
        super(BnActConv2d, self).__init__()
        self.bn = nn.BatchNorm2d(in_chs, eps=0.001)
        self.act = activation_fn
        self.conv = nn.Conv2d(in_chs, out_chs, kernel_size, stride, padding, groups=groups, bias=False)

    def forward(self, x):
        return self.conv(self.act(self.bn(x)))


class InputBlock(nn.Module):
    def __init__(self, num_init_features, kernel_size=7, in_chans=3,
                 padding=3, activation_fn=nn.ReLU(inplace=True)):
        super(InputBlock, self).__init__()
        self.conv = nn.Conv2d(
            in_chans, num_init_features, kernel_size=kernel_size, stride=2, padding=padding, bias=False)
        self.bn = nn.BatchNorm2d(num_init_features, eps=0.001)
        self.act = activation_fn
        self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)

    def forward(self, x):
        x = self.conv(x)
        x = self.bn(x)
        x = self.act(x)
        x = self.pool(x)
        return x


class DualPathBlock(nn.Module):
    def __init__(
            self, in_chs, num_1x1_a, num_3x3_b, num_1x1_c, inc, groups, block_type='normal', b=False):
        super(DualPathBlock, self).__init__()
        self.num_1x1_c = num_1x1_c
        self.inc = inc
        self.b = b
        if block_type == 'proj':
            self.key_stride = 1
            self.has_proj = True
        elif block_type == 'down':
            self.key_stride = 2
            self.has_proj = True
        else:
            assert block_type == 'normal'
            self.key_stride = 1
            self.has_proj = False

        if self.has_proj:
            # Using different member names here to allow easier parameter key matching for conversion
            if self.key_stride == 2:
                self.c1x1_w_s2 = BnActConv2d(
                    in_chs=in_chs, out_chs=num_1x1_c + 2 * inc, kernel_size=1, stride=2)
            else:
                self.c1x1_w_s1 = BnActConv2d(
                    in_chs=in_chs, out_chs=num_1x1_c + 2 * inc, kernel_size=1, stride=1)
        self.c1x1_a = BnActConv2d(in_chs=in_chs, out_chs=num_1x1_a, kernel_size=1, stride=1)
        self.c3x3_b = BnActConv2d(
            in_chs=num_1x1_a, out_chs=num_3x3_b, kernel_size=3,
            stride=self.key_stride, padding=1, groups=groups)
        if b:
            self.c1x1_c = CatBnAct(in_chs=num_3x3_b)
            self.c1x1_c1 = nn.Conv2d(num_3x3_b, num_1x1_c, kernel_size=1, bias=False)
            self.c1x1_c2 = nn.Conv2d(num_3x3_b, inc, kernel_size=1, bias=False)
        else:
            self.c1x1_c = BnActConv2d(in_chs=num_3x3_b, out_chs=num_1x1_c + inc, kernel_size=1, stride=1)

    def forward(self, x):
        x_in = torch.cat(x, dim=1) if isinstance(x, tuple) else x
        if self.has_proj:
            if self.key_stride == 2:
                x_s = self.c1x1_w_s2(x_in)
            else:
                x_s = self.c1x1_w_s1(x_in)
            x_s1 = x_s[:, :self.num_1x1_c, :, :]
            x_s2 = x_s[:, self.num_1x1_c:, :, :]
        else:
            x_s1 = x[0]
            x_s2 = x[1]
        x_in = self.c1x1_a(x_in)
        x_in = self.c3x3_b(x_in)
        if self.b:
            x_in = self.c1x1_c(x_in)
            out1 = self.c1x1_c1(x_in)
            out2 = self.c1x1_c2(x_in)
        else:
            x_in = self.c1x1_c(x_in)
            out1 = x_in[:, :self.num_1x1_c, :, :]
            out2 = x_in[:, self.num_1x1_c:, :, :]
        resid = x_s1 + out1
        dense = torch.cat([x_s2, out2], dim=1)
        return resid, dense


class DPN(nn.Module):
    def __init__(self, small=False, num_init_features=64, k_r=96, groups=32,
                 b=False, k_sec=(3, 4, 20, 3), inc_sec=(16, 32, 24, 128),
                 num_classes=1000, in_chans=3, drop_rate=0., global_pool='avg', fc_act=nn.ELU()):
        super(DPN, self).__init__()
        self.num_classes = num_classes
        self.drop_rate = drop_rate
        self.b = b
        bw_factor = 1 if small else 4

        blocks = OrderedDict()

        # conv1
        if small:
            blocks['conv1_1'] = InputBlock(
                num_init_features, in_chans=in_chans, kernel_size=3, padding=1)
        else:
            blocks['conv1_1'] = InputBlock(
                num_init_features, in_chans=in_chans, kernel_size=7, padding=3)

        # conv2
        bw = 64 * bw_factor
        inc = inc_sec[0]
        r = (k_r * bw) // (64 * bw_factor)
        blocks['conv2_1'] = DualPathBlock(num_init_features, r, r, bw, inc, groups, 'proj', b)
        in_chs = bw + 3 * inc
        for i in range(2, k_sec[0] + 1):
            blocks['conv2_' + str(i)] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'normal', b)
            in_chs += inc

        # conv3
        bw = 128 * bw_factor
        inc = inc_sec[1]
        r = (k_r * bw) // (64 * bw_factor)
        blocks['conv3_1'] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'down', b)
        in_chs = bw + 3 * inc
        for i in range(2, k_sec[1] + 1):
            blocks['conv3_' + str(i)] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'normal', b)
            in_chs += inc

        # conv4
        bw = 256 * bw_factor
        inc = inc_sec[2]
        r = (k_r * bw) // (64 * bw_factor)
        blocks['conv4_1'] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'down', b)
        in_chs = bw + 3 * inc
        for i in range(2, k_sec[2] + 1):
            blocks['conv4_' + str(i)] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'normal', b)
            in_chs += inc

        # conv5
        bw = 512 * bw_factor
        inc = inc_sec[3]
        r = (k_r * bw) // (64 * bw_factor)
        blocks['conv5_1'] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'down', b)
        in_chs = bw + 3 * inc
        for i in range(2, k_sec[3] + 1):
            blocks['conv5_' + str(i)] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'normal', b)
            in_chs += inc
        blocks['conv5_bn_ac'] = CatBnAct(in_chs, activation_fn=fc_act)
        self.num_features = in_chs
        self.features = nn.Sequential(blocks)

        # Using 1x1 conv for the FC layer to allow the extra pooling scheme
        self.global_pool = SelectAdaptivePool2d(pool_type=global_pool)
        self.classifier = nn.Conv2d(
            self.num_features * self.global_pool.feat_mult(), num_classes, kernel_size=1, bias=True)

    def get_classifier(self):
        return self.classifier

    def reset_classifier(self, num_classes, global_pool='avg'):
        self.num_classes = num_classes
        self.global_pool = SelectAdaptivePool2d(pool_type=global_pool)
        if num_classes:
            self.classifier = nn.Conv2d(
                self.num_features * self.global_pool.feat_mult(), num_classes, kernel_size=1, bias=True)
        else:
            self.classifier = None

    def forward_features(self, x):
        return self.features(x)

    def forward(self, x):
        x = self.forward_features(x)
        x = self.global_pool(x)
        if self.drop_rate > 0.:
            x = F.dropout(x, p=self.drop_rate, training=self.training)
        out = self.classifier(x)
        return out.flatten(1)


@register_model
def dpn68(pretrained=False, num_classes=1000, in_chans=3, **kwargs):
    default_cfg = default_cfgs['dpn68']
    model = DPN(
        small=True, num_init_features=10, k_r=128, groups=32,
        k_sec=(3, 4, 12, 3), inc_sec=(16, 32, 32, 64),
        num_classes=num_classes, in_chans=in_chans, **kwargs)
    model.default_cfg = default_cfg
    if pretrained:
        load_pretrained(model, default_cfg, num_classes, in_chans)
    return model


@register_model
def dpn68b(pretrained=False, num_classes=1000, in_chans=3, **kwargs):
    default_cfg = default_cfgs['dpn68b']
    model = DPN(
        small=True, num_init_features=10, k_r=128, groups=32,
        b=True, k_sec=(3, 4, 12, 3), inc_sec=(16, 32, 32, 64),
        num_classes=num_classes, in_chans=in_chans, **kwargs)
    model.default_cfg = default_cfg
    if pretrained:
        load_pretrained(model, default_cfg, num_classes, in_chans)
    return model


@register_model
def dpn92(pretrained=False, num_classes=1000, in_chans=3, **kwargs):
    default_cfg = default_cfgs['dpn92']
    model = DPN(
        num_init_features=64, k_r=96, groups=32,
        k_sec=(3, 4, 20, 3), inc_sec=(16, 32, 24, 128),
        num_classes=num_classes, in_chans=in_chans, **kwargs)
    model.default_cfg = default_cfg
    if pretrained:
        load_pretrained(model, default_cfg, num_classes, in_chans)
    return model


@register_model
def dpn98(pretrained=False, num_classes=1000, in_chans=3, **kwargs):
    default_cfg = default_cfgs['dpn98']
    model = DPN(
        num_init_features=96, k_r=160, groups=40,
        k_sec=(3, 6, 20, 3), inc_sec=(16, 32, 32, 128),
        num_classes=num_classes, in_chans=in_chans, **kwargs)
    model.default_cfg = default_cfg
    if pretrained:
        load_pretrained(model, default_cfg, num_classes, in_chans)
    return model


@register_model
def dpn131(pretrained=False, num_classes=1000, in_chans=3, **kwargs):
    default_cfg = default_cfgs['dpn131']
    model = DPN(
        num_init_features=128, k_r=160, groups=40,
        k_sec=(4, 8, 28, 3), inc_sec=(16, 32, 32, 128),
        num_classes=num_classes, in_chans=in_chans, **kwargs)
    model.default_cfg = default_cfg
    if pretrained:
        load_pretrained(model, default_cfg, num_classes, in_chans)
    return model


@register_model
def dpn107(pretrained=False, num_classes=1000, in_chans=3, **kwargs):
    default_cfg = default_cfgs['dpn107']
    model = DPN(
        num_init_features=128, k_r=200, groups=50,
        k_sec=(4, 8, 20, 3), inc_sec=(20, 64, 64, 128),
        num_classes=num_classes, in_chans=in_chans, **kwargs)
    model.default_cfg = default_cfg
    if pretrained:
        load_pretrained(model, default_cfg, num_classes, in_chans)
    return model