In tf.Keras , the function update_step() is used by an optimizer to update the variables at every iteration given the variable and the gradient. I want to know if this gradient is the full gradient of all the samples in the batch or the gradient for a sample in the batch?
thank you
When using the model.fit()
or the model.train_on_batch()
the optimizer’s update_step()
is called after the forward pass and the computation of the loss for a batch of data.
The gradients passed to the update_step()
function are the full gradients of all the samples in the batch (i.e batch gradient)
.
Thank you!