Skip to content

thesis_experiments.md

Latest commit

 

History

History
57 lines (31 loc) · 4.22 KB

thesis_experiments.md

File metadata and controls

57 lines (31 loc) · 4.22 KB

Configurations for all schizophrenia experiments reported in "Utilization of brain connectivity in classification and regression tasks in brain data" diploma thesis.

Schizophrenia discrimination

Standard ML models

  • Config: ml_config.py
  • Script: run_ml_experiment.py

Deep learning models

For all deep learning models, the hyperparameter space was first explored using the --single_select_fold option of run_experiment.py.

Dense model

  • Config: dense_config.py
  • Script: run_experiment.py with triangular_dense as model_type

Results are reported for this configuration:

[{'node_features': [<NodeFeatures.FC_MATRIX_ROW: 1>], 'correlation_type': [<CorrelationType.XI: functools.partial(<function xicorr at 0x000002142869AAF8>)>], 'batch_size': [4, 8, 16], 'num_hidden_features': [2, 4], 'num_sublayers': [1], 'dropout': [0.3], 'criterion': [<function cosine_loss at 0x000002142FEFF708>], 'optimizer': [<class 'torch.optim.adamw.AdamW'>], 'optimizer_kwargs': {'lr': [0.01], 'weight_decay': [0.0001]}, 'epochs': [50], 'scheduler': [<class 'torch.optim.lr_scheduler.LinearLR'>], 'scheduler_kwargs': {'start_factor': [1], 'end_factor': [0.01], 'total_iters': [50]}}]

Proposed model

  • Config: connectivity_dense_config.py
  • Script: run_experiment.py with connectivity_dense as model_type

Results are reported for this configuration:

[{'node_features': [<NodeFeatures.FC_MATRIX_ROW: 1>], 'correlation_type': [<CorrelationType.SPEARMAN: 'spearman'>], 'batch_size': [2], 'num_hidden_features': [45], 'num_sublayers': [2], 'dropout': [0.3], 'readout': ['max'], 'emb_dropout': [0.0, 0.1], 'emb_residual': [None, 'mean'], 'emb_init_weights': ['normal'], 'emb_val': [0.0], 'emb_std': [0.0001], 'criterion': [<function mse_loss at 0x0000029BB2773F78>], 'optimizer': [<class 'torch.optim.adamw.AdamW'>], 'optimizer_kwargs': {'lr': [0.005], 'weight_decay': [0.0001]}, 'epochs': [50], 'scheduler': [<class 'torch.optim.lr_scheduler.LinearLR'>], 'scheduler_kwargs': {'start_factor': [1], 'end_factor': [0.005], 'total_iters': [50]}}]

GAT model

  • Config: gat_config.py
  • Script: run_experiment.py with gat as model_type

Results are reported for this configuration:

[{'node_features': [<NodeFeatures.FC_MATRIX_ROW: 1>], 'correlation_type': [<CorrelationType.PEARSON: 'pearson'>, <CorrelationType.XI: <numpy.vectorize object at 0x00000251C3D814C8>>], 'batch_size': [16], 'num_hidden_features': [45], 'num_sublayers': [1, 2], 'graph_kwargs': {'thresholding_function': [<ThresholdingFunction.SUBJECT_VALUES: 2>], 'threshold_type': [<DataThresholdingType.FIXED_THRESHOLD: 2>], 'threshold': [-100.0]}, 'criterion': [<function mse_loss at 0x00000251B0253F78>], 'optimizer': [<class 'torch.optim.adamw.AdamW'>], 'optimizer_kwargs': {'lr': [0.003], 'weight_decay': [0.0001]}, 'epochs': [30], 'scheduler': [<class 'torch.optim.lr_scheduler.LinearLR'>], 'scheduler_kwargs': {'start_factor': [1], 'end_factor': [0.01], 'total_iters': [50]}}]

GIN model

  • Config: gin_config.py
  • Script: run_experiment.py with gin as model_type

Results are reported for this configuration:

[{'node_features': [<NodeFeatures.ONE_HOT_REGION: 2>, <NodeFeatures.FC_MATRIX_ROW: 1>], 'correlation_type': [<CorrelationType.PEARSON: 'pearson'>], 'batch_size': [32], 'num_hidden_features': [45], 'num_sublayers': [2], 'dropout': [0.5], 'eps': [0.0], 'graph_kwargs': {'thresholding_function': [<ThresholdingFunction.SUBJECT_VALUES: 2>], 'threshold_by_absolute_value': [False], 'return_absolute_value': [False], 'threshold_type': [<DataThresholdingType.KNN: 1>], 'threshold': [5, 8, 10]}, 'criterion': [<function mse_loss at 0x0000019E119B3F78>], 'optimizer': [<class 'torch.optim.adamw.AdamW'>], 'optimizer_kwargs': {'lr': [0.003], 'weight_decay': [0.0001]}, 'epochs': [100], 'scheduler': [<class 'torch.optim.lr_scheduler.LinearLR'>], 'scheduler_kwargs': {'start_factor': [1], 'end_factor': [0.01], 'total_iters': [50]}}]

Sparsification experiment

  • Config: gin_config.py
  • Script: run_experiment.py with gin as model_type, and --single_select_fold option along with num_select_folds 10 and num_assess_folds 100

The configuration is too long for printing. As such it is available in sparsification_config.py (still it needs to be run as gin_config.py).