# Copyright (c) Alibaba, Inc. and its affiliates.
import torch.nn as nn
import torch.nn.functional as F
[docs]class GeMPooling(nn.Module):
"""GemPooling used for image retrival
p = 1, avgpooling
p > 1 : increases the contrast of the pooled feature map and focuses on the salient features of the image
p = infinite : spatial max-pooling layer
"""
[docs] def __init__(self, p=3, eps=1e-6):
super(GeMPooling, self).__init__()
# self.p = nn.Parameter(torch.ones(1)*p)
self.p = p
self.eps = eps
[docs] def forward(self, x):
return self.gem(x, p=self.p, eps=self.eps)
[docs] def gem(self, x, p=3, eps=1e-6):
return F.avg_pool2d(x.clamp(min=eps).pow(p),
(x.size(-2), x.size(-1))).pow(1. / p)
[docs]class MultiPooling(nn.Module):
"""Pooling layers for features from multiple depth.
"""
POOL_PARAMS = {
'resnet50': [
dict(kernel_size=10, stride=10, padding=4),
dict(kernel_size=16, stride=8, padding=0),
dict(kernel_size=13, stride=5, padding=0),
dict(kernel_size=8, stride=3, padding=0),
dict(kernel_size=6, stride=1, padding=0)
]
}
POOL_SIZES = {'resnet50': [12, 6, 4, 3, 2]}
POOL_DIMS = {'resnet50': [9216, 9216, 8192, 9216, 8192]}
[docs] def __init__(self,
pool_type='adaptive',
in_indices=(0, ),
backbone='resnet50'):
super(MultiPooling, self).__init__()
assert pool_type in ['adaptive', 'specified']
if pool_type == 'adaptive':
self.pools = nn.ModuleList([
nn.AdaptiveAvgPool2d(self.POOL_SIZES[backbone][l])
for l in in_indices
])
else:
self.pools = nn.ModuleList([
nn.AvgPool2d(**self.POOL_PARAMS[backbone][l])
for l in in_indices
])
[docs] def forward(self, x):
assert isinstance(x, (list, tuple))
return [p(xx) for p, xx in zip(self.pools, x)]
[docs]class MultiAvgPooling(nn.Module):
"""Pooling layers for features from multiple depth.
"""
POOL_PARAMS = {
'resnet50': [
dict(kernel_size=10, stride=10, padding=4),
dict(kernel_size=16, stride=8, padding=0),
dict(kernel_size=13, stride=5, padding=0),
dict(kernel_size=8, stride=3, padding=0),
dict(kernel_size=7, stride=1, padding=0)
]
}
# POOL_SIZES = {'resnet50': [12, 6, 4, 3, 2]}
POOL_SIZES = {'resnet50': [12, 6, 4, 3, 1]}
POOL_DIMS = {'resnet50': [9216, 9216, 8192, 9216, 2048]}
[docs] def __init__(self,
pool_type='adaptive',
in_indices=(0, ),
backbone='resnet50'):
super(MultiAvgPooling, self).__init__()
assert pool_type in ['adaptive', 'specified']
if pool_type == 'adaptive':
self.pools = nn.ModuleList([
nn.AdaptiveAvgPool2d(self.POOL_SIZES[backbone][l])
for l in in_indices
])
else:
self.pools = nn.ModuleList([
nn.AvgPool2d(**self.POOL_PARAMS[backbone][l])
for l in in_indices
])
[docs] def forward(self, x):
assert isinstance(x, (list, tuple))
return [p(xx) for p, xx in zip(self.pools, x)]