notes

batch/epoch/iter

Batch:

One batch means a group of samples processed together in one forward/backward pass. If Batch size = 100 means this batch contains 100 images.

Epoch:

One epoch means a full pass through your entire training dataset

• if you have 1000 training images and batch size is 10, then batches per epoch = 100

iter:

One iteration is one update of the model’s weight. so it is also called “step”

iterations per epoch = total training samples / batch size
total iterations = epochs * ibatchsize

Iteration (or “step”)
One iteration is one weight update—that is:

1. take one batch of size B,
2. do a forward pass,
3. compute the loss,
4. do a backward pass,
5. update the weights.