Batch size
Also known as: batch
A batch is a group of examples processed together before the model updates its weights. The batch size is how many examples are in that group.
- Small batch: frequent, noisy updates (fast, less stable).
- Large batch: fewer, smoother updates (slower, more stable, needs more memory).
It’s one of the most-tuned knobs in a training run, with real effects on both speed and final quality.
Why it matters locally: bigger batches need more VRAM. When a tool says "out of memory", reducing batch size is often the first fix.