BatchNormalization's argument seems incorrect - Githubissues

yashkant / enas-quantized-nets

Efficient Neural Architecture Search coupled with Quantized CNNs to search for resource efficient and accurate architectures.

4 stars 2 forks source link

BatchNormalization's argument seems incorrect #1

Closed Albelt closed 5 years ago

Albelt commented 5 years ago

Hi: Thanks for your work, here's my question

In pnas/model.py/parse_action(), action's input "x" is "channel_last", however the followed "BatchNormalization" function's argument axis is "1", I think the correct argument is "3" because the keras documentation writes as this:
```
axis: Integer, the axis that should be normalized
        (typically the features axis).
        For instance, after a `Conv2D` layer with
        `data_format="channels_first"`,
        set `axis=1` in `BatchNormalization`.
```
the axis should be the last axis, isn't that? Or maybe my lib's version is different from yours? Here is my development environment:
```
linux 4.19.13
python3.6
tensorflow 1.8.0
keras 2.2.4
```

yashkant commented 5 years ago

Hi Albelt,

You seem to have opened this issue in a wrong repository.

The concerned repository is: https://github.com/yashkant/PNAS-Binarized-Neural-Networks

Thanks for pointing this out. Yes, it is wrong, the axis=3 is the correct argument for it.

If you wish to correct it for others as well, please send a PR here I'll be glad to merge it. :)