[FIXED] How to combine multiple loss and accuracy during training in Keras?

Issue

With reference to the question, during training I am getting multiple losses and accuracies. For example,

Epoch 1/100
1382/1382 [==============================] - 694s 502ms/step - loss: 0.6798 - decoder_loss: 0.3399 - decoder_accuracy: 0.5770 - decoder_accuracy_1: 0.5770 - val_loss: 0.6606 - val_decoder_loss: 0.3373 - val_decoder_accuracy: 0.5783 - val_decoder_accuracy_1: 0.5783

It's hard to understand and what I want is only loss, accuracy, validation_loss and validation_accuracy. Is there any way to merge them.

Another thing is my log file is getting bigger and bigger as the network print the loss/accuracy after each step as:

1/1382 [..............................] - ETA: 13:59 - loss: 0.7226 - decoder_loss: 0.3613 - decoder_accuracy: 0.5536 - decoder_accuracy_1: 0.5536
2/1382 [..............................] - ETA: 10:23 - loss: 0.7204 - decoder_loss: 0.3602 - decoder_accuracy: 0.5881 - decoder_accuracy_1: 0.5881
3/1382 [..............................] - ETA: 8:57 - loss: 0.7151 - decoder_loss: 0.3576 - decoder_accuracy: 0.5821 - decoder_accuracy_1: 0.5821 

Can I reduce it to output only when the epoch ends instead of each step?

Solution

i assume you use the fit funciton of keras with a verbose of 1. E.g.

model.fit(Xtrain, Ytrain, batch_size = 32, epochs = 100, verbose=1)

You can change the verbose to 2 then you will get only one output per epoch. Alternatively you can turn the training progress output off if you set verbose to 0. (Source: Keras API) E.g.

model.fit(Xtrain, Ytrain, batch_size = 32, epochs = 5, verbose=2)

Should give you an output like:

I hope this solves your problem. If not you might want to share a little bit of your code so we can help you futher.

EDIT1: As for your question if you can influence your monitoring metrics: Yes you can, but I can only answer your question with assumptions because I don't have your code. If you use keras for your model you should have used the "compile" funciton of keras somewhere. E.g

model.compile(optimizer=keras.optimizers.RMSprop(),  # Optimizer
          # Loss function to minimize
          loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
          # List of metrics to monitor
          metrics=["accuracy"])

You can influence the metrics which are shown in your console if you change the metrics values to the values you want to see. (Source:Kearas API,Keras Metrics)

If you only want to see the loss, accuracy you can use the example above. The metrics val_... will be generated accordingly if you use a validation _split or validation_datat in your fit function. E.g.

history = model.fit(x_train, y_train,
                batch_size=64,
                epochs=3,
                # We pass some validation for
                # monitoring validation loss and metrics
                # at the end of each epoch
                validation_data=(x_val, y_val))

If build you two examples to visualize my point Example 1 shows accuracy, loss, val_acc and val_loss Example 2 shows accuracy, loss,mae, val_acc, val_loss and val_mae

For completness here the link to the colab notebook, where you can see my code:Colab for Examples

I hope this helps to solve your quesion completly. If not let me know.

Answered By - Fabian

Answer Checked By - Marilyn (PHPFixing Volunteer)