Skip to content

[Major] - Plotting Update, SCU Standardization

Mohib Mustafa requested to merge dev_mm into main

  • Histogram plots for collected data includes strain rate.

  • Collected Data Plots are standardized and saved additionally in PDFs.

  • Data Plots are standardized and saved additionally in PDFs.

  • SCU prediction plots are standardized and saved additionally in PDFs.

  • SCU MSE error plots are standardized and saved additionally in PDFs.

  • SCU name changed to SCU_TH_D for the delta trained model.

  • SCUT name changed to SCU_TH_T for the total trained model.

  • Surrogates are stored in separate folders with unique names for tractability.

  • Data transfer modules are generated and updated inside the unique folders, this prevents unwanted IO runtime errors and allows parallel training of different surrogates on a single machine.

  • Cold start on existing surrogates results in warning and prevents data loss by overwriting.

  • Traced modules that can be used in cm3libraries are saved in the unique folders and are named cm3Model.pt, previously there was a risk of accidental call due to unambiguous naming.

  • Choice of loss optimizer now results in a new folder to prevent any pollution

  • libtorch c++ API implementations are improved:

    • Cold and warm starts are unified in single routine
    • Loss optimizers are unified in single routine
    • Print statements added to track hyper parameters
    • Device choice is automatic per available architecture
    • CPU and GPU modules are saved automatically per architecture
    • learning rate, momentum, weight decay and gradient clip can be changed without recompilation which results in better training control
    • Testing and training error logs are supplemented with epochs, this results in proper tracking and allows changing the batch size without manual alteration of plotting scripts

Merge request reports