Ensemble Models

[aditya0by0]

• Implementation for Add ensemble model #72

[aditya0by0]

Test example config :

class_path: chebai.models.ensemble.ChebiEnsemble
init_args:
  optimizer_kwargs:
    lr: 1e-3
  model_configs: {
    "chebi50": {
      "ckpt_path": "logs/chebi50_bce_unweighted/version_9/checkpoints/best_epoch=00_val_loss=0.3521_val_macro-f1=0.0381_val_micro-f1=0.1753.ckpt",
      "class_path": chebai.models.Electra,
      "labels_path": "data/chebi_v231/ChEBI50/processed/classes.json",
    },

    "chebi50_1": {
      "ckpt_path": "logs/chebi50_bce_unweighted/version_9/checkpoints/best_epoch=00_val_loss=0.3521_val_macro-f1=0.0381_val_micro-f1=0.1753.ckpt",
      "class_path": chebai.models.Electra,
      "labels_path": "data/chebi_v231/ChEBI50/processed/classes.json",
    },
  }

[aditya0by0]

Please find the latest config file eg... which is working for electra and gnn

class_path: chebai.ensemble.FullEnsembleWMV
init_args:
  data_processed_dir_main: "data/chebi_v231/ChEBI50/processed"
  operation_mode: "prediction"
  smiles_list_file_path: "data/chebi_v231/ChEBI50/processed/smiles.txt"
  _perform_validation_checks: False  # To avoid check for using same model with same model configs
  model_configs: {
    "chebi50": {
      "model_ckpt_path": "logs\\chebi50_electra_231\\version_1\\checkpoints\\best_epoch=00_val_loss=0.6766_val_macro-f1=0.7273_val_micro-f1=0.0248.ckpt",
      "model_config_file_path": "configs\\model\\electra.yml",
      "data_config_file_path": "configs\\data\\chebi\\chebi50.yml",
      "model_labels_path": "data/chebi_v231/ChEBI50/processed/classes.json",
      "wrapper_class_path": chebai.ensemble.NNWrapper,
    },

    "chebi50_1": {
      "model_ckpt_path": "logs\\chebi50_electra_231\\version_1\\checkpoints\\best_epoch=00_val_loss=0.6766_val_macro-f1=0.7273_val_micro-f1=0.0248.ckpt",
      "model_config_file_path": "configs\\model\\electra.yml",
      "data_config_file_path": "configs\\data\\chebi\\chebi50.yml",
      "model_labels_path": "data/chebi_v231/ChEBI50/processed/classes.json",
      "wrapper_class_path": chebai.ensemble.NNWrapper,
    },

    "chebi50_graph": {
      "model_ckpt_path": "logs\\regated_gnn_v231\\version_0\\checkpoints\\best_epoch=00_val_loss=0.6786_val_macro-f1=0.6364_val_micro-f1=0.0321.ckpt",
      "model_config_file_path": "..\\python-chebai-graph\\configs\\model\\gnn_res_gated.yml",
      "data_config_file_path":  "..\\python-chebai-graph\\configs\\data\\chebi50_graph_properties.yml",
      "model_labels_path": "data/chebi_v231/ChEBI50/processed/classes.json",
      "wrapper_class_path": chebai.ensemble.GNNWrapper,
    },

    # "chemlog": {
    #   "model_labels_path": "data/chebi_v231/ChEBI50/processed/classes.json",
    #   "wrapper_class_path": chebai.ensemble.ChemLogWrapper,
    # },
  }

[aditya0by0]

@sfluegel05, I have added wrapper for chemlog. What I need from the method is some kind of prediction for each class, which can be passed to sigmoid function later on.

I am not sure how chemlog achieves this. Can you please guide me?

[sfluegel05]

I moved the ensemble logic to https://github.com/ChEB-AI/python-chebifier and the property calculation to the results folder

[sfluegel05]

see ChEB-AI/python-chebifier#1

[sfluegel05]

All right, then this should be mergable