Batch Size

Number of patches processed together during each training step.

Batch size controls how many patches are processed simultaneously before the model weights are updated. The control differs between the two branches, because the StructN2V branch works on larger patches that consume more memory.

N2V branch

• A dropdown with options 1, 2, 4, 8, 16, 32, 64, 128

• Default: 128

• The small 64×64 patches let many fit in memory at once

StructN2V branch

• A numeric field accepting values from 1 to 256

• Default: 24

• The larger 128×128 patches need a smaller batch to fit in memory

Smaller batches

• Use less GPU memory

• Can lead to noisier training gradients

Larger batches

• More stable training gradients

• Require more GPU memory

Recommendations

• Keep the preset default (128 for N2V, 24 for StructN2V)

• Reduce it if you see out-of-memory errors, especially when running on CPU or a small GPU