How to train a regressor with 3D X_train?

[wnanao]

Why does the process start training another model after all my epochs are trained? Is it because some of my parameters are not set correctly? In my understanding, the model training phase is generally over after all epochs have been trained, but the output area prints out the structural information and training information of model{n} and starts training another model. What part of the knowledge do I need to know or how to modify the code?
The following is my model definition code and output information：

regressor = ResNetRegressor(n_epochs=1500, batch_size=128, verbose=True)
# regressor = InceptionTimeRegressor(n_epochs=2, batch_size=128, verbose=True)
regressor.fit(X_nor, y_nor)

I found that the problem occurs in my input data, this does not occur when my training data X_train and y_train are 2D and 1D respectively. But when my X_train and y_trian are three-dimensional and two-dimensional respectively, like [n_instances, n_dimensions, series_length] and [n_instances, series_length]. The fit() function will train models of series_length length.
For example, the following code will train five times, 20 epochs each time:

    X_train = np.random.rand(100, 12, 5)
    y_train = np.random.rand(100, 5)
    clf = ResNetRegressor(n_epochs=20, verbose=True)
    print("shape of X_train:", X_train.shape)
    print("shape of y_train:", y_train.shape)
    clf.fit(X_train, y_train)

[wnanao]

@fkiraly Can you give me some helps?

[fkiraly]

Imo this is expected behaviour: if I look into the estimator, ResNetRegressor cannot deal with multivariate output y (in your case, y_train), because the capability:multioutput tag is False.

So, what the estimator will instead do, is train one instance of the model per column of y. In your case that's 5 models time 20 epochs, so 100 epochs overall.

Now, given the neural network architecture under the hood of ResNetRegressor, there is no inherent restriction of adapting it for multioutput y (you'd need to take care of the output layer being formatted right), it's just that it hasn't been done by the implementers, so it trains multiple models by default.

If you would like to give it a go, you'd be very welcome to contribute this extension!