Evaluating my model in new unseen dataset

Question

I have trained my model ( .fit() ) and satisfied with the performance on test split making prediction ( .predict() ). So I save my model to the disk ( .save('model.h5') ).

Now I'm given new unseen dataset and asked to evaluate my already saved model on this dataset for performance. I am required to not only report accuracy but stuff line precision/recall, confusion matrix etc...

I then loaded my saved model ( .load_model('model.h5') ).

Question: What the appropriate function that I should use to prepare report of the model performance on this new dataset. I should I use .predict() function or .evaluate() .

Answer 1

If you want to get loss/accuracy or whatever other metrics you had during training - you need .evaluate() method. If all you need is actual probabilities or regression values - you need .predict() method.

Answer 2

You can use sklearn's classification_report to generate all the relevant metrics.
Code:

preds = model.predict(x1)
y_pred = np.argmax(preds, axis = 1)
print(classification_report(y1, y_pred))

Output:

              precision    recall  f1-score   support

           0       0.71      0.08      0.15        59
           1       0.42      0.95      0.58        41

    accuracy                           0.44       100
   macro avg       0.57      0.52      0.37       100
weighted avg       0.59      0.44      0.33       100

You can see all available metrics here .

model.evaluate() :
This method will provide the values for loss and metrics you defined in your compile method which is generally not enough.