Use an old .tlt model to retrain it with a new dataset

Hello, I’m trying to train a .tlt model which I trained with 133 classes with a new dataset that has 221 classes. Do you know if it is possible to use that .tlt model with this new dataset with a different number of classes?

The model is an efficientnet B1 for image classification. I’m facing an error when I’m doing tao train, because in my dense layer I have 133 nodes (which is the expected number for the model). I think it should be 221 because that’s the new number of classes but I don’t know how to change it.

Is there any parameter to change this that I’m not considering?

Thank you in advance.

For classification, I can run the experiment you mentioned successfully.
Just set the pretrained model to your existing 133_classes.tlt.

pretrained_model_path: “<your 133 classes tlt model”>

Then run training with the new 221 classes dataset.

Ok thank you, I’ll try again and let you know. If the model is pruned could be a problem?

So, do you mean you use the pruned tlt model as the pretrained model?

More, can you share the error log?

Please share training spec as well. Thanks.

Right now I’m doing the train with the NGC model, so I can’t share it in this moment. I’ll share it with you in a few hours but It said something like “(133,” shape was expected but “(221,” was found, where 133 and 221 are the number of classes of the old and new model respectively.

And yes, the .tlt model I was trying to use was a pruned model.

Can you try to set “pretrained_model_path:” to the existing tlt model when you train for the first time?

More, please remove previous result folder before triggering new training.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.