AndreaBasile97 / GNN-BRaTs2021

0 stars 0 forks source link

Trovare il miglior setting per il modello GAT o GraphSage-pool #4

Closed AndreaBasile97 closed 1 year ago

AndreaBasile97 commented 1 year ago

Fase di ricerca del miglior setting per GAT o GraphSage-pool.

Una volta trovato il setting migliore, salvare il modello e generare predizioni sul test set salvando le seguenti metriche:

Le metriche del test devono essere salvate.

AndreaBasile97 commented 1 year ago
-- DATASET PICKLE --
dataset pickle = /ext/tesi_BraTS2021/pickle_dataset/full_dataset_with_id_01.pickle
-- MODEL PATH --
model path = /ext/tesi_BraTS2021/saved_models/GraphSage 2 giugno/model_epoch_25_graphSage.pth 

-- TYPE MODEL --
model = GraphSage(
  (layers): ModuleList(
    (0): SAGEConv(
      (feat_drop): Dropout(p=0, inplace=False)
      (fc_pool): Linear(in_features=20, out_features=20, bias=True)
      (fc_neigh): Linear(in_features=20, out_features=256, bias=False)
      (fc_self): Linear(in_features=20, out_features=256, bias=True)
    )
    (1-5): 5 x SAGEConv(
      (feat_drop): Dropout(p=0, inplace=False)
      (fc_pool): Linear(in_features=256, out_features=256, bias=True)
      (fc_neigh): Linear(in_features=256, out_features=256, bias=False)
      (fc_self): Linear(in_features=256, out_features=256, bias=True)
    )
    (6): SAGEConv(
      (feat_drop): Dropout(p=0, inplace=False)
      (fc_pool): Linear(in_features=256, out_features=256, bias=True)
      (fc_neigh): Linear(in_features=256, out_features=4, bias=False)
      (fc_self): Linear(in_features=256, out_features=4, bias=True)
    )
  )
)

-- HYPERPARAMS --
patience = 10
lr = 0.0005
weight_decay = 0.0001
gamma = 0.98
val_dropout = 0
layer_sizes = [256, 256, 256, 256, 256, 256]

-- PARAMETERS --
in_feats = 20
n_classes = 4

-- DATE --
timestamp = 2023-06-05 21:31:46.554066
VALORI MEDI TEST:

dice_score_node_WT - 0.783774
dice_score_node_CT - 0.735940
dice_score_node_ET - 0.807926
dice_score_voxel_WT - 0.764561
dice_score_voxel_CT - 0.719248
dice_score_voxel_ET - 0.693549
hd95_voxel_WT - 21.407069
hd95_voxel_CT - 27.250875
hd95_voxel_ET - 20.351951