Vector output not possible for multitarget classification

[kaueltzen]

Hi, I noticed this issue while implementing multitarget classification in #245 .

At the moment, training a multitarget classification MODNetModel with a single output vector targets like this: [[["cl_target_0","cl_target_1"]]]
will only train on the first target and discard the others.

First, in the build_model method, the output shape is set to (None, num_classes[group[prop_idx][0]]) instead of (None, sum([v for k, v in num_classes.items() if k in group[prop_idx]])).

modnet/modnet/models/vanilla.py

Lines 196 to 209 in 7ee7d8b

	n = num_classes[group[prop_idx][0]]
	name = group[prop_idx][0]
	if n >= 2:
	out = tf.keras.layers.Dense(
	n,
	activation="sigmoid" if multi_label else "softmax",
	name=name,
	)(previous_layer)
	else:
	out = tf.keras.layers.Dense(
	len(group[prop_idx]),
	activation=out_act,
	name=name,
	)(previous_layer)

In the fit method, only the first target is converted to y.

modnet/modnet/models/vanilla.py

Lines 347 to 360 in 7ee7d8b

	for prop in self.targets_groups:
	if self.num_classes[prop[0]] >= 2: # Classification
	targ = prop[0]
	if self.multi_label:
	y_inner = np.stack(training_data.df_targets[targ].values)
	if loss is None:
	loss = "binary_crossentropy"
	else:
	y_inner = tf.keras.utils.to_categorical(
	training_data.df_targets[targ].values,
	num_classes=self.num_classes[targ],
	)
	if loss is None:
	loss = "categorical_crossentropy"

I'd either correct this or add in the documentation that multitarget classification should be done with separate output layers.