Test a newly trained model, which weights/biases set does the mode use?

[Z_Huang]

I have a silly question. For example, I trained my model with 100 epochs. At epoch 88, the accuracy rate has reached the highest (92%) level. The 100th epoch only produced an 89% accuracy rate.

Once the train is finished with the 100 epochs (train/validation at the same time), the next step is to test the model. I was wondering which epoch weights/biases do the model use? Is it the 88th (92%) model or the 100th (89%) model?

When I save the model to my local device, which weights/biases set does the model used to store the model file locally?

[ptrblck]

@Z_Huang

Once the train is finished with the 100 epochs (train/validation at the same time), the next step is to test the model. I was wondering which epoch weights/biases do the model use? Is it the 88th (92%) model or the 100th (89%) model?

Assuming you are calculating the accuracy on the validation dataset, you could use early stopping and store the model giving the best validation accuracy.

@Z_Huang

When I save the model to my local device, which weights/biases set does the model used to store the model file locally?

Whatever the current state_dict returns, i.e. if you’ve continued training the model to 100 epochs and store the model.state_dict() afterwards, the current parameters trained for 100 epoch are stored.