Hub: LRN for InceptionV1?

[swghosh]

I inspected the model being used by "https://tfhub.dev/google/imagenet/inception_v1/classification/1" (looked up the TF-slim implementation, since the SavedModel is ported from there), appears to me that the model is using Batch Normalization layer instead of LRN layer discussed in the original paper (InceptionV1 a.k.a. GoogLeNet)

[arnoegw]

That's correct.

[swghosh]

So like it's still an official implementation, right? Not much harm w/ using BatchNormalization instead of LRN.
I think so that w/ BatchNorm results had improved also. (over the original BVLC caffe weights)

[arnoegw]

I'm not sure what "official" would mean here... In any case, TF-Slim is a reputable source of CNNs and has been widely used. It provides a credible implementation of the Inception V1 architecture, albeit extended by batch normalization, as you say, which does improve training.

Is there a specific issue with TF Hub that you would like the TF Hub team to address?

[swghosh]

I think that's a reasonably good explanation. I was unsure about creating issues on the tf-slim repository as it not maintained anymore for which I queried it here. 😅

Thanks for the clarifications anyway!

[arnoegw]

Happy to help.

BTW, version /1 is outdated. I recommend using https://tfhub.dev/google/imagenet/inception_v1/classification/3 with TF1/hub.Module or https://tfhub.dev/google/imagenet/inception_v1/classification/4 with TF1 and the new APIs for TF2 SavedModels.

Our previous discussion applies equally.

[swghosh]

Yeah, am aware of that. I was using "/4", just moght have shared the wrong link.