Replace some functions in the forward method of Inception3 with Module object calls

[mitmul]

I tried to use torchvision.models._utils.IntermediateLayerGetter to extract features from the output of Mixed_7c block of Inception3 model, and realized that naively using this useful class for the model skips some operations such as F.max_pool2d, F.adaptive_avg_pool2d, and F.dropout which are applied in the forward method of the model class.

So, I think it's good to add those operations as nn.Module objects in the constructor, e.g., nn.MaxPool2d object for max pooling operation, etc., to enable IntermediateLayerGetter to perform desired forward computation with Inception3 model.

By this PR, how to extract feature maps from intermediate layer outputs of Inception3 will be changed as followings.

Before:

import torchvision.models as models
import torch
import torch.nn.functional as F

class Inception3_feature(torch.nn.Module):
    def __init__(self, net):
        super().__init__()
        self.net = net
    def forward(self, x):
        x = self.net.Conv2d_1a_3x3(x)
        x = self.net.Conv2d_2a_3x3(x)
        x = self.net.Conv2d_2b_3x3(x)
        x = F.max_pool2d(x, kernel_size=3, stride=2)
        x = self.net.Conv2d_3b_1x1(x)
        x = self.net.Conv2d_4a_3x3(x)
        x = F.max_pool2d(x, kernel_size=3, stride=2)
        x = self.net.Mixed_5b(x)
        x = self.net.Mixed_5c(x)
        x = self.net.Mixed_5d(x)
        x = self.net.Mixed_6a(x)
        x = self.net.Mixed_6b(x)
        x = self.net.Mixed_6c(x)
        x = self.net.Mixed_6d(x)
        x = self.net.Mixed_6e(x)
        x = self.net.Mixed_7a(x)
        x = self.net.Mixed_7b(x)
        x = self.net.Mixed_7c(x)
        return x

extractor = Inception3_feature(models.inception_v3(pretrained=True))
x = torch.rand((1, 3, 224, 224))

feature = extractor(x)

After:

import torchvision.models as models
import torch
from torchvision.models._utils import IntermediateLayerGetter

inception = models.inception_v3(pretrained=True, aux_logits=False)
extractor = IntermediateLayerGetter(inception, return_layers={'Mixed_7c': 'out'})
x = torch.rand((1, 3, 224, 224))

feature = extractor(x)

I would appreciate it if this PR could be acceptable.

[fmassa]

Thanks a lot!

For information, the current situation to extract intermediate features in PyTorch is not perfect and IntermediateLayerGetter is fairly fragile, specially if the order of the modules registered is not the same as the order they are used in the forward pass.

One proposal that we had was to use torchscript to potentially give us a more robust solution, but it is still under discussion pytorch/pytorch#21064