DeepDriveMD / DeepDriveMD-pipeline

DeepDriveMD: Deep-Learning Driven Adaptive Molecular Simulations
MIT License
15 stars 9 forks source link

Dbg/integration #19

Closed braceal closed 3 years ago

braceal commented 3 years ago

This PR includes:

Bug Fixes:

  1. Fix: AAE sh MASTER_ADDR.
  2. Fix: Specify virtual_ prefix on virtual h5 files so the name doesn't collide with h5 files copied to node local ssd.
  3. Fix: Typing in AAE config.

New Features:

  1. Logging which h5 files are selected during training/agent stages.
  2. Start of Analysis module with utility functions for accessing DDMD data.
  3. Utility function get_total_iterations for getting the total number of DDMD iterations.
  4. Logic for retrain_freq. Model selection always runs and duplicates previous latest checkpoint if ML stage is skipped.
  5. Support for node local SSD for training and simulation.
  6. Support for training a different number of epochs during the first iteration.
  7. apply_analysis_fn for applying a function over MD task outputs in parallel.

Updates to test configurations to test the new features.