How I could initialize fixed weights for cnn each time when I start training?
You want every initialization of your network to be the same when you start training?
You can set random seeds for that.
Be careful though that many things are non-deterministic and so won’t behave exactly the same between runs. You can read more about this here.
I want to compare performance between 3 CNNs but I want to fix the weights to ensure that the changes in performance doesn’t comes from weights random initialization of each
The thing is that the random value of the weights might favor one architecture/training procedure over the others. So fixing one random initialization is not really a fair comparison.
To prevent weight initialization from interfering, you would need to run each model to a few random seeds and compare average and variance of their performances.