[Feature Request] Re-enable multi-output logs

[innat]

If a model is configured for multi-output, in keras 2, we can get the log message for all corresponding outputs. But in keras 3, it is removed and only shows the summarized log message. Unlike mult-class classification, returning log message for mult-output sounds more natural.

priority_pred = layers.Dense(1, name="priority")(x)
department_pred = layers.Dense(num_departments, name="department")(x)

model = keras.Model(
    inputs=[title_input, body_input, tags_input],
    outputs={"priority": priority_pred, "department": department_pred},
)

model.compile(
    optimizer=keras.optimizers.RMSprop(1e-3),
    loss={
        "priority": keras.losses.BinaryCrossentropy(from_logits=True),
        "department": keras.losses.CategoricalCrossentropy(from_logits=True),
    },
    loss_weights={"priority": 1.0, "department": 0.2},
)

model.fit

in keras 2

Epoch 1/2
 40/40 ━ 1s 12ms/step - loss: ?, priority_loss: ?, department_loss: ?

in keras 3

Epoch 1/2
 40/40 ━ 1s 12ms/step - loss: ?

Posted as a bug #19159, but it turned out that it was intentionally removed.

[james77777778]

+1 for this

I have also filed the issue at #18993

[arthurflor23]

I'm having the same restriction using GAN models, which have several loss values...

[fchollet]

The recommendation is to pass metrics= with a dict mapping different outputs to the metric you want to monitor. If the metric you need is the same as the loss, pass that. e.g.

model.compile(
    optimizer=keras.optimizers.RMSprop(1e-3),
    loss={
        "priority": keras.losses.BinaryCrossentropy(from_logits=True),
        "department": keras.losses.CategoricalCrossentropy(from_logits=True),
    },
    metrics={
        "priority": [keras.metrics.BinaryCrossentropy(from_logits=True)],
        "department": [keras.metrics.CategoricalCrossentropy(from_logits=True)],
    },
    loss_weights={"priority": 1.0, "department": 0.2},
)

[innat]

@fchollet
Just to be clear, in keras 2, we can get

Epoch 1/2
 40/40 ━ 1s 12ms/step - loss: ?, priority_loss: ?, department_loss: ?

but in keras 3

Epoch 1/2
 40/40 ━ 1s 12ms/step - loss: ?

So, it's not about the metrics but loss scores of a multi output model. And it seems missing in the log in keras 3. We found this unexpected, coz sometimes we prefer to monitor target loss score using callback to reduce lr or saving checkpoint, etc. As this ticket is closed, I guess this is fixed but I've just tried a sample code with keras-nightly (3.2.0.dev2024040203), it's same as before.

model.compile(
    optimizer=keras.optimizers.RMSprop(1e-3),
    loss={
        "priority": keras.losses.BinaryCrossentropy(from_logits=True),
        "department": keras.losses.CategoricalCrossentropy(from_logits=True),
    },
    loss_weights={"priority": 0.1, "department": 0.2},
)

Epoch 1/2
40/40 ━━━━━━━━━━━━━━━━━━━━ 5s 48ms/step - loss: 0.6244
Epoch 2/2
40/40 ━━━━━━━━━━━━━━━━━━━━ 2s 44ms/step - loss: 0.6285