Batch Size

Number of patches processed together during each training step.

Batch size controls how many patches are processed simultaneously before updating model weights.

Smaller batches (1-4)

• Use less GPU memory

• Can lead to noisier training

• Necessary for limited GPU memory

Larger batches (8-32)

• More stable training gradients

• Require more GPU memory

• Can train faster per epoch

Recommendations

• Start with 4-8 and adjust based on GPU memory

• Reduce if you see memory errors

• Larger batches work better with larger learning rates