Is the maxpool and avgpool interchanged here? about pretrained-models.pytorch

[Cadene]

@Noplz @shicai You were right, I made a mistake. I fixed it in this commit 40f4f7e

@aussetg A more refactored version is available in Tensorflow https://github.com/taehoonlee/tensornets/blob/master/tensornets/nasnets.py
However, it was easier for me to create CellStem0, CellStem1, Cell0, Cell1 and ReductionCell0 regarding my will to port the pretrained parameters from tensorflow after my deadline (15 nov).
By the way I mainly used tensorboard.

[Cadene]

@rkaplan @Noplz @aussetg Done ! 247b037

[shicai]

@Noplz I think you are right. it should be max pooling for comb_iter_1_left and average pooling for comb_iter_2_left .

[aussetg]

I also don't understand why there are Cell1 and Cell0, the original paper only includes one type of Normal Cell, and one type of Reduction Cell

[aussetg]

@Cadene How do you port weights from TensorFlow?

Also I don't think the TF weights are sota, they don't implement path wise dropout like in the paper ( probably because they can't with TensorFlow as the graph then changes at every iteration. But then I wonder what framework the Google team used 😏 )

[Noplz]

@Cadene great job anyway, this repo is really helpful :) 👍

[rkaplan]

Congrats on finishing your deadline :) just wanted to say thank you for working on this NASNet port, looking forward to trying it out!