You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
324 lines
12 KiB
324 lines
12 KiB
""" 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 select_adaptive_pool2d
|
|
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.global_pool = global_pool
|
|
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.classifier = nn.Conv2d(in_chs, 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 = global_pool
|
|
del self.classifier
|
|
if num_classes:
|
|
self.classifier = nn.Conv2d(self.num_features, num_classes, kernel_size=1, bias=True)
|
|
else:
|
|
self.classifier = None
|
|
|
|
def forward_features(self, x, pool=True):
|
|
x = self.features(x)
|
|
if pool:
|
|
x = select_adaptive_pool2d(x, pool_type=self.global_pool)
|
|
return x
|
|
|
|
def forward(self, x):
|
|
x = self.forward_features(x)
|
|
if self.drop_rate > 0.:
|
|
x = F.dropout(x, p=self.drop_rate, training=self.training)
|
|
out = self.classifier(x)
|
|
return out.view(out.size(0), -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
|
|
|
|
|
|
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
|