The model is not working for different image classification problems, can you explain how to go about this issue?

[ab-bh]

No description provided.

[thibo73800]

Hard to say without more information about the exact problem and the dataset

[ab-bh]

I am using kaggle dataset for Food-101 which is similar in structure to the pickle files you used. The accurqcy score is juggling between 0.04 and 0.0 even after 1500 steps.

[ab-bh]

Link: https://www.kaggle.com/kmader/food41/data

the data i specifically downloaded was 32x32x3 as per the one you used for the traffic sign data.

[ab-bh]

Can you please tell the reason about this issue? what could be the steps that are to be taken to deal with it?

[thibo73800]

Hi,

Unfortunately, I don't have enough time to test this dataset with caps net. However, it seems to be a very complex dataset for the present architecture.

This architecture works well on simple Dataset with repetitive pattern (Thanks to the vector representation of the shape). But, it is not state of the art on other datasets like CIFAR (which is less difficult than your dataset).

There is a lot of issues to solve on a CapsuleNetowk to scale it to this kind of dataset. I don't have a solution to give you, I can just tell you what I would do to deal with this dataset, which is maybe not the solution:

I would start to apply a classic convolutional neural network on it and tune the hyperparameters to see how well I perform, I would apply data augmentation as well. Once I would have a good CNN architecture, I would start to change some component of the CNN step by step toward a CapsNet architecture. By doing so, I would observe and identify which parts of the CapsNet architecture doesn't work properly and then try to deduce why and how to fix the problems.

[shrybht]

Could you please walk me through the steps to train CapsNet on CoCo dataset?