I am using a loss function:

meanAbsoluteError = tf.keras.losses.MeanAbsoluteError()

And a model:

model.compile(optimizer=tf.keras.optimizers.Adam(0.001), loss=meanAbsoluteError)

Now I train it and get the history:

history = model.fit(train_features, train_labels, epochs=100, validation_data=(test_features, test_labels))

I can check how low the loss has become at the end:

print(history.history["val_loss"][-1])

Now I try to compute the loss for the test data manually, using the same data as passed in validation_data when training:

print(meanAbsoluteError(model(test_features), test_labels).numpy())

But this gives me a completely different number! In my opinion this should equal the loss I got from the previous expression. But that's not the case.

Why the two numbers are not equal, and which one is the correct loss?

I checked how to call the loss function here: https://www.tensorflow.org/api_docs/python/tf/keras/losses/Loss#call

I edited this to use history.history["val_loss"] instead of history.history["loss"] as suggested, but it doesn't fix it.

CodePudding user response：

Your problem seems to be caused by the batch size. The loss in history['val_loss'] is calculated on the batch. If your test_features has a different size than your training batch, it will give you a larger loss.

You should try:

print(meanAbsoluteError(model(test_features[:batch_size]), test_labels[:batch_size]).numpy())