Fine-tuning and retraining of the model

[andergalvaan]

Hi,

Does anyone know of a method to retrain the model with some images and then test it with images of the same and different people? I mean, it is possible to keep the feature extraction that the model implements, but retraining the model to better detect some people? I have retrained the model with images of some people and it detects them more reliably through cosine similarity, but with images of people who do not participate in the retraining more false positives appear.

Thank you in advanced,
Ander

[Leo021017]

I have the same problem, do you solve it now? Thank you!

[andergalvaan]

Hi @Leo021017

I have tried not to train all layers. I have frozen some of them and the rest I have retrained them (partial fine-tuning). This approach works well in my scenario.

Best regards,
Ander