GoogLeNet

[Atcold]

I was thinking that, for completeness, you may want to train a GoogLeNet as well. Here's the model.

[szagoruyko]

@Atcold thanks. It is unlikely that I will have spare GPUs to train it in the near future.

[Atcold]

Oh, OK, I'll train it myself then. Hopefully I get comparable results, which could broaden your comparison. Also, what's about NIN 0(?) and inception-v2 51.00(?)? You didn't include any additional note.

[szagoruyko]

I couldn't get specified performance for these models. either due to preprocessing or to class ordering probably.

[Atcold]

"class ordering probably"? What do you mean?

[szagoruyko]

there is classes.t7 which generated randomly sometimes and NIN authors used a different one from usually used in caffe. it is random for example in https://github.com/soumith/imagenet-multiGPU.torch. I just didn't have enough courage to fix that

[Atcold]

Hmm... Not sure I got it right. Anyhow, your inception-v2 I believe is actually GoogLeNet (without first layer factorisation).

[ducha-aiki]

Why not BVLC-GoogLeNet (inception 1) there? Is there some reason for it?

[szagoruyko]

@ducha-aiki no one cared to import it from caffe, there are better models trained in torch

[ducha-aiki]

@szagoruyko well, there is no any Inception 1 model in your evaluation, wherever they were trained :)

[szagoruyko]

@ducha-aiki I believe @Atcold should know better, he did an evaluation paper with these networks recently. Alfredo, can you help us? Is inception-v2 in my table actually a GoogLeNet

[Atcold]

Here I am, sorry for the delay, I was on vacation.

There is a bit of a confusion, with nomenclature. I'll try to shade some light. (Ref. http://arxiv.org/abs/1602.07261) Inception is the name of the whole Google's family networks, using parallel convolutions with different receptive fields.

• Inception-v1: GoogLeNet -> https://arxiv.org/abs/1409.4842
• Inception-v2: GoogLeNet + BN -> https://arxiv.org/abs/1502.03167
• Inception-v3 -> http://arxiv.org/abs/1512.00567
• Inception-v4 -> http://arxiv.org/abs/1602.07261

The only correct GoogLeNet model around, is this -> https://github.com/e-lab/imagenet-multiGPU.torch/blob/trueGoogLeNet/models/googlenetFixed_cudnn.lua which is not trained. All others models are not GoogLeNet, but some approximations.

In my analysis, I dismissed the Inception-v2 from this repo, because it was not a reliable reproduction of the original GoogLeNet (the input convolution should have been factorised, brining a substantial operation count saving).

I hope I have somehow answered. If not, please, feel free to bug me 😊