MyAdaptivePooling.py

import torch
from torch import nn

class MyAdaptiveMaxPool2d(nn.Module):
    def __init__(self, sz=None):
        super().__init__()
        self.p = nn.MaxPool2d((10, 10), padding=0) 
        # why (10, 10)? Because input image size is 299, \
        # if you use 224, this should be (7, 7)
        # if you want to know which number for other image size, 
        # put pdb.set_trace() at forward method and print x.size()
    
    def forward(self, x):
        return self.p(x)

class MyAdaptiveAvgPool2d(nn.Module):
    def __init__(self, sz=None):
        super().__init__()
        self.p = nn.AvgPool2d((10, 10), padding=0)

    def forward(self, x): 
        return self.p(x)

class AdaptiveConcatPool2d(nn.Module):
    def __init__(self, sz=None):
        super().__init__()
        sz = sz or (1,1)
        #self.ap = nn.AdaptiveAvgPool2d(sz)
        #self.mp = nn.AdaptiveMaxPool2d(sz)
        self.ap = MyAdaptiveAvgPool2d(sz) # <-- replace new layer of AdaptivePooling here
        self.mp = MyAdaptiveMaxPool2d(sz) # <-- replace new layer of AdaptivePooling here
    def forward(self, x): return torch.cat([self.mp(x), self.ap(x)], 1)