Understanding Imagenet training

Question

I am trying to start training Imagenet classification training using Tensorflow's inception model. I am a bit confused as I am not sure how to fully train the model.

Wherever I go everyone seems to be using the 1000 classes training list. However, while digging in tensorflows inception model files I found a list of labels containing 21000+ classes.

So I am a bit confused as to whether I should train using the 1000 classes that everyone seems to be using or should I train on the 21000 classes. I have the full imagenet data and I checked the images I downloaded and almost all 21000 classes are there (missing only 200 classes).

I really want to train against the full imagenet dataset as this should give more accuracy but I am not sure what the downfalls will be or how to properly prepare the data and how many steps are appropriate for that.

Answer 1

The problem with 1000 classes is the ILSVRC2012 challenge which is using a subset of the full imagenet with "only" 1000 classes. The advantage of that is that it is fixed while imagenet itself is always expanding and new classes and examples may be added.

Training on the full ImageNet set is a very long task and will require either a very long time or a long time and a lot of computing powers (talking tens of GPUs). Although it's not short, training ILSVRC2012 can be done in one day up to a few days depending on your hardware resources and complexity of the model being training.