"do something" when Optuna finds new best parameters

Question

I have a neural-network where I'm using Optuna to find some optimal hyper-parameters e.g batch-size etc.

I want to save the nets-parameters when Optuna finds a new best parameter-combination.

I have tried the following two approahces:

SCORE = 0
def objective(trial):
    BATCH_SIZE = trial.suggest_int("BATCH_SIZE",20,100)
    LEARNING_RATE = trial.suggest_float("LEARNING_RATE",0.05,1)
    DROPOUT = trial.suggest_float("DROPOUT",0.1,0.9)

    Y_SCORE,Y_VAL = train_NN(X,y,word_model,BATCH_SIZE,250,LEARNING_RATE,DROPOUT)
    y_val_pred = Y_SCORE.argmax(axis=1)
    labels = encode.inverse_transform(np.arange(6))
    a = classification_report(Y_VAL, y_val_pred,zero_division=0,target_names=labels,output_dict=True)
    score = a.get("macro avg").get("f1-score")
    if score>SCORE: #New best weights found - save the net-parameters
        SCORE = score
        torch.save(net,"../model_weights.pt")
    return score

which fails with UnboundLocalError: local variable 'SCORE' referenced before assignment but if I move SCORE=0 inside the function at the top, it resets at each trial.

The reason I want to save the weights right away, and not just run another training with study.best_params at the end, is that sometimes the random initialization of the weights has an impact and gives a higher score (although if the training is robust, it should not make a difference) - but that is not the point/the issue.

Answer 1

This answer is helpful to save the neural network's weights; We can use the callback function to save the model's checkpoint.