Keras mentions that it provided models pretrained on ImageNet. However, it doesn't specify what they mean by "ImageNet" - like is it a certain subset of ImageNet of the complete set of images? I mean, I have a hard time imagining that they have used all the 14 million images for pretraining but maybe I am wrong.

I have read quite a lot of research papers about some of the proposed architectures and they are all trained on subsets of ImageNet specific to the ImageNet Large Scale Visual Recognition Challenge and this is explicitly mentioned in the papers. Of course, if these Keras weights are actually representing the whole ImageNet, there is no problem :-) But would like some sort of confirmation/affirmation.

Does anyone know more specifically what is mean by "ImageNet"? https://keras.io/api/applications/


The Keras link you provided itself provides links to the definition of each model. For example, ResNet-50 the relevant paper is located at: https://arxiv.org/abs/1512.03385

The paper explicitly states:

  1. Experiments

4.1. ImageNet Classification

We evaluate our method on the ImageNet 2012 classification dataset [36] that consists of 1000 classes. The models are trained on the 1.28 million training images, and evalu- ated on the 50k validation images. We also obtain a final result on the 100k test images, reported by the test server. We evaluate both top-1 and top-5 error rates.

Most models which enter the ImageNet competition use the same subset, subject to confirmation from each paper.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.