inception-resnet-v2 | Implementation of Google 's Inception ResNet v2 | Machine Learning library

It achieves similar results.

I just received an e-mail confirming the error from Alex Alemi, Senior Research Scientist at Google and original publisher of the blog post regarding the release of the code for InceptionResnetV2. It seems that during internal experiments, the STEM blocks were switched and the release just kept like that.

Cite:

Dani Azemar,

It seems you're right. Not entirely sure what happened but the code is obviously the source of truth in the sense that the released checkpoint is for the code that is also released. When we were developing the architecture we did a whole slew of internal experiments and I imagine at some point the stems were switched. Not sure I have the time to dig deeper at the moment, but like I said, the released checkpoint is a checkpoint for the released code as you can verify yourself by running the evaluation pipeline. I agree with you that it seems like this is using the original Inception V1 stem. Best Regards,

Alex Alemi

I'll update this post with changes regarding this subject.

UPDATE: Christian Szegedy, also publisher of the original paper, just tweeted me:

The original experiments and model was created in DistBelief, a completely different framework pre-dating Tensorflow.

The TF version was added a year later and might have had discrepancies from the original model, however it was made sure to achieve similar results.

So, since it achieves similar results, your experiments would be roughly the same.