Issues with the batch_size setting

[Summer-68]

Due to insufficient GPU memory, I want to modify the batch_size parameter. Where in the code file should I modify the batch_size parameter?

[fuyuao000]

If you have already completed dataset preprocessing using the planner, the batch size can be modified in the DatasetXXX folder under nnUNet_preprocessed and in the nnUNetPlans.json file. However, for scenarios with limited GPU memory, the nnUNet planner generally sets a lower limit of batchsize=2. Forcibly reducing it to 1 may lead to a series of problems such as increased gradient estimation noise. From the perspective of nnUNet configuration process, I recommend that you reduce the patchsize rather than the batchsize in the json file, or introduce measures such as gradient checkpointing.