How to finetune alpha and gamma?

[mobassir94]

for my given data problem(nlp task) if i use focal_loss(gamma=1.5,alpha = .2) then the result i get like this :
Train for 2000 steps, validate for 32 steps Epoch 1/4 1999/2000 [============================>.] - ETA: 0s - loss: 0.0436 - auc: 0.9015 ROC-AUC - epoch: 1 - score: 0.875631 2000/2000 [==============================] - 1737s 869ms/step - loss: 0.0436 - auc: 0.9015 - val_loss: 0.0453 - val_auc: 0.8754 Epoch 2/4 1999/2000 [============================>.] - ETA: 0s - loss: 0.0241 - auc: 0.9758 ROC-AUC - epoch: 2 - score: 0.903044 2000/2000 [==============================] - 1610s 805ms/step - loss: 0.0241 - auc: 0.9758 - val_loss: 0.0360 - val_auc: 0.9030 Epoch 3/4 1999/2000 [============================>.] - ETA: 0s - loss: 0.0215 - auc: 0.9812 ROC-AUC - epoch: 3 - score: 0.911900 2000/2000 [==============================] - 1612s 806ms/step - loss: 0.0215 - auc: 0.9812 - val_loss: 0.0352 - val_auc: 0.9118 Epoch 4/4 1999/2000 [============================>.] - ETA: 0s - loss: 0.0191 - auc: 0.9852 ROC-AUC - epoch: 4 - score: 0.909566 2000/2000 [==============================] - 1610s 805ms/step - loss: 0.0191 - auc: 0.9852 - val_loss: 0.0445 - val_auc: 0.9074 CPU times: user 4min 15s, sys: 23.7 s, total: 4min 39s Wall time: 1h 49min 31s

now when i tried focal_loss(gamma=2.0,alpha = .2) i get :

Train for 1896 steps, validate for 32 steps Epoch 1/3 1895/1896 [============================>.] - ETA: 0s - loss: 0.0107 - auc: 0.9842 ROC-AUC - epoch: 1 - score: 0.544101 1896/1896 [==============================] - 1644s 867ms/step - loss: 0.0107 - auc: 0.9842 - val_loss: 0.0860 - val_auc: 0.5423 Epoch 2/3 1895/1896 [============================>.] - ETA: 0s - loss: 0.0083 - auc: 0.9904 ROC-AUC - epoch: 2 - score: 0.573175 1896/1896 [==============================] - 1522s 803ms/step - loss: 0.0083 - auc: 0.9904 - val_loss: 0.0659 - val_auc: 0.5210 Epoch 3/3 1895/1896 [============================>.] - ETA: 0s - loss: 0.0070 - auc: 0.9946 ROC-AUC - epoch: 3 - score: 0.375477 1896/1896 [==============================] - 1522s 803ms/step - loss: 0.0070 - auc: 0.9945 - val_loss: 0.0396 - val_auc: 0.4966 CPU times: user 2min 54s, sys: 17.4 s, total: 3min 12s Wall time: 1h 18min 9s

terrible val_auc right?
please help me choose alpha and gamma for focal loss