How to grab gradients?

[patins1]

Hi,

During training, at the onEpoch event, I want to read the gradients for certain model parameters, but they are all zero already.
I realized, they become zero by a call to ParameterStore.updateAllParameters() during a training step (I debugged it down to a call of NDArrayEx.adamUpdate()). It surely makes sense to put the parameters to zero again after a training step, however it doesn't let me log the gradients for analysis reasons, is there a better approach to achieve that?

[zachgk]

I assume you are using EasyTrain so correct me if I am wrong, but it may not work entirely for you. There aren't any TrainingListener events that occur mid-batch right now. I recommend forking the whole EasyTrain class. Then, you can access the gradients inside trainBatch before they are cleared

[patins1]

Thanks, I found the event to get grab the gradient eventually: onTrainingBatch