replaced mean on dimensions 2,3 by adaptive_avg_pooling2d

[talcs]

with destination of 1 (which is a Pytorch way to implement GlobalAvgPooling2D), to remove hardcoded dimension ordering

[talcs]

This pull request is following this issue: #1773

Since F.adaptive_avg_pool2d(x, 1) returns a tensor shaped as torch.Size([B, C, 1, 1]), I applied torch.squeeze on it, to get a torch.Size([B, C]) array

[codecov-io]

Codecov Report

❗ No coverage uploaded for pull request base (master@5c49d99). Click here to learn what that means.
The diff coverage is 0%.

@@           Coverage Diff            @@
##             master   #1838   +/-   ##
========================================
  Coverage          ?   0.43%           
========================================
  Files             ?      92           
  Lines             ?    7389           
  Branches          ?    1113           
========================================
  Hits              ?      32           
  Misses            ?    7349           
  Partials          ?       8

Impacted Files	Coverage Δ
torchvision/models/inception.py	0% <0%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 5c49d99...d6b5279. Read the comment docs.

[talcs]

I first used squeeze to turn the [B,C,1,1] shaped avg pooling result to [B,C], but when B=1, the shape will become [C], which is incorrect. Hence, in order to keep the batch size dimension even when B=1, I had to use reshape(x.shape[0], -1).

[fmassa]

Thanks!