Training the Image Embedder

[Zumbalamambo]

I would like to perform large scale image retrieval. How do I retrain the Image Embedder?

I have clustered the images into multiple folders by clustering based on the pose and trained the ImageClassifier thinking that I can use it as a model to extract the embedding. Unfortunately the loss is not reducing at all. Is there any specific way to train for image retrieval?

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[Zumbalamambo]

+1

[edgarriba]

@Zumbalamambo do you have already a pipeline or some code to take a look at?

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[Zumbalamambo]

+1 still a problem

[ethanwharris]

Hi @Zumbalamambo Could you provide some more details about the issue? If the loss isn't reducing for your classification data then it's possible that it's just not a very solvable problem. Have you been able to train effectively on this data outside of flash?

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.