Commits on Aug 26, 2024

1. Run Torchvision problems with Benchmark[Problem/Runner/Metric]; conso… …

   …lidate PyTorchCNN problems (facebook#2688)
   
   Summary:
   Pull Request resolved: facebook#2688
   
   Context: Consolidating on common abstractions will make it easier to add functionality to all classes. Also, this will make the code smaller and easier to navigate. This diff is substantially LOC-negative outside of tests.
   
   This PR:
   * Updates functionality in torchvision.py that replaces MNIST datasets with fakes when run in a test environment; the data sets are now realistic enough to be usable.
   * Merge the functionality in `pytorch_cnn.py` into `torchvision.py`; it was only ever used in order to support `torchvision.py`.
   * Remove `PyTorchCNNTorchvisionBenchmarkProblem` and its special serialization logic; it is replaced with `BenchmarkProblem`
   * Remove `PyTorchCNNTorchvisionRunner`, `PyTorchCNNBenchmarkProblem`, `PyTorchCNNMetric`, and `PyTorchCNNRunner`
   * Introduce `PyTorchCNNTorchvisionParamBasedProblem`. It does not need special serialization logic because it is a dataclass with datasets constructed in the `__post_init__`. When an instance is serialized, the data sets are not serialized; they are reconstructed when the instance is decoded. Using a dataclass here also allows for an automatic and more rigorous equality check.
   * Use `BenchmarkRunner`; as per D61483962, this means that this problem now has a ground truth, which won't change its behavior since it doesn't have noise added.
   
   Reviewed By: Balandat
   
   Differential Revision: D61414680

   

   esantorella authored and facebook-github-bot committed Aug 26, 2024
   Configuration menu

   • View commit details
   • Copy full SHA for 1c042f8
   • Browse repository at this point

   Copy the full SHA

   1c042f8 View commit details

   Browse the repository at this point in the history