Grain 0.2.17
-
New features:
- Adds automated batching into shared memory when multiprocess prefetch is
used to save one data copy. - Updates TfMixtureIndexSampler to accept dynamic 1D/2D Tensor/Variable
proportions that could be updated during runtime. - Automatic Xprof subprocess trace collection into PyGrain workers. To enable,
set flaggrain_enable_multiprocess_worker_profiling=trueand add"profile_subprocesses" = Truein advanced profiler options.
- Adds automated batching into shared memory when multiprocess prefetch is
-
Breaking changes:
- Force index shuffle rounds to not be bigger than 1024
-
Bug fixes:
- Preserve and reconstruct tracebacks for exceptions thrown in worker processes across multiprocessing boundaries
- Fixes performance regression in traceback filtering when workers are
frequently exhausting small datasets. - Fixes silent string truncation in shared memory batching.