search

dchaley / deepcell-imaging

Tools & guidance to scale DeepCell imaging on Google Cloud Batch
8 stars 2 forks source link

Create benchmark that skips post-processing #172

Closed dchaley closed 6 months ago

dchaley commented 6 months ago

Follow up to #96 .

We don't just want to speed up the benchmark. The post-processing step is very time & memory intensive. This requires large machines with many cores to get enough RAM, but then we can't attach any GPUs (because of quota).

For now, let's just skip post-processing entirely and exercise the inference portion w/ more data sizes & GPUs.

dchaley commented 6 months ago

See also documentation for Mesmer application: https://deepcell.readthedocs.io/en/master/_modules/deepcell/applications/mesmer.html#Mesmer

This hard-codes the post-processing function, it's not available as a parameter to Mesmer.

However we can create our own Mesmer subclass application, which overrides the constructor to provide an empty postprocessing function. This avoids creating a deepcell fork with a custom Mesmer implementation.