Custom image encoder [P]

“`html

A user on Reddit asked about the feasibility of building a custom image encoder for their video frame classification task. They are looking to replace existing models like CLIP, SigLIP/SigLIP2, or DINO with a smaller and more efficient model tailored specifically for their dataset.

• The user is concerned about processing speed and deploying on small CPU-only devices.
• They aim to train a custom encoder using only a few million parameters and around four to five labels.
• This approach aims to improve both the embedding generation speed and the accuracy of their Transformer model.

The question revolves around whether it is viable to develop such a custom encoder, considering constraints like computational efficiency on resource-limited devices. This inquiry highlights the ongoing interest in optimizing AI models for specific use cases while maintaining performance standards.

“`

Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.

Please enable JavaScript in your browser to complete this form.

Name *

First

Last

Name Email

Email *

Submit