This PR addresses issue #85.

Example usage

from diart.inference import Benchmark, Parallelize
from diart import OnlineSpeakerDiarization, PipelineConfig

config = PipelineConfig()
benchmark = Benchmark("/wav/dir", "/rttm/dir")
p_benchmark = Parallelize(benchmark, num_workers=4)
if __name__ == "__main__":  # Needed for multiprocessing
    p_benchmark(OnlineSpeakerDiarization, config)

Changelog

• Add --num-workers argument to diart.benchmark
• Add diart.inference.Parallelize, a wrapper for Benchmark to replace sequential execution with multiprocessing
• Expose some new fine-grained methods in Benchmark so that Parallelize can reuse it
• diart.stream now uses rich progress bars
• Add diart.progress package with ProgressBar, RichProgressBar and TQDMProgressBar as adapters for each library
• Chronometer can now be aware of the progress bar used so that it can print reports with the correct formatting
• BasePipeline objects now must be able to communicate their associated configuration class (through the get_config_class() static method)
• PipelineConfig.from_namespace() is now PipelineConfig.from_dict() and receives an easily serializable configuration so that workers can instantiate their own pipelines (entire models cannot be sent to child processes)
  • This dictionary needs to be documented and formalized, maybe as a data class. Otherwise its use can be confusing
• Add parallelization example in README.md
• Models are now lazy. They only load weights when required, making them lighter for inter-process communication

Future improvements and limitations

• Optimizer is still not compatible with Parallelize because some progress bars break
• Replace tqdm with rich as progress bars in both Benchmark and Optimizer (when not running in parallel)
• Spawn segmentation and embedding models as services in separate processes so the GPU memory requirements go down from O(num_workers * model_size) to O(model_size)