Does PIC (personal image classifier) retrain the model?

Hi, I have a question about PIC web app.
I used the same training data (step 1), same model setting(step 2), and same testing data (step 3) to train a model four times, which mean what I only did was click the "Train Model" button four times and got four models. I found that the Loss of these four models were so different (0.0111--> 0.00509 --> 0.18294 --> 0.03581), and there were also different results in step 4 when different models with the same testing data. I am wonder if PIC web app retrain a model when I click the "Train Model" button, or the previous model will affect the training of the new model when I click the button ? (maybe something like transfer learning?)

Thank you

Each time you train it is a new model. The reason the losses are different is because during training the images are shown in a random order. You will likely have better convergence if you increase the number of epochs or the training fraction, at the risk of overtraining the model.

1 Like

I've got it, Thanks!