Train CNNs using deeprankcore

[gcroci2]

We use here newly generated data (see issue #149), with no pssm (the files we have are wrong) and with the newest software version (15/05/2023), for trying out CNNs, always with the 100k pMHCI data (see issue #113).

• Experiments are saved in /data/pMHCI/trained_models/deeprankcore/experiments/
• HDF5 files are in /data/pMHCI/features_output_folder/deeprankcore/residue/230515
  • grids + graphs, 99854
  • 96 files, ~14 GB each, total 1.45 TB ~13h for generation
  • Only cl_peptide target added

Experiments

• exp_100k_cnn_bs64_0_230530

  • Using CnnClassification (in deeprankcore.neuralnets.cnn.model3d)
  • Bad results in terms of metrics (random classifier)
  • Run 26 epochs, total time ~14h (1 GPU, 16 workers)

• exp_100k_cnn_bs128_CnnClass4ConvKS3Lin128ChannExpand_230601

  • CnnClass4ConvKS3Lin128ChannExpand
  • AUC on shuffled data 0.73, waiting for a decision about what to do for the paper

[gcroci2]

We decided to use the original deeprank for the scientific paper. We will use deeprankcore with CNNs for the future framework paper. So for now, these experiments will stop. But we do know that deeprankcore works for CNNs as well.