如何使用 tf.data.Dataset 训练多元回归 output？

Question

Only the first output parameter is learned to be properly estimated during training of a multi regression output.net. Second and subsequent parameters only seem to follow first parameter. It seems, that ground truth for second output parameter is not used during training. How do I shape tf.data.Dataset and input it into model.fit() function so second output parameter is trained?

    import tensorflow as tf
    import pandas as pd
    from tensorflow import keras
    from keras import layers
    
    #create dataset from csv
    file = pd.read_csv( 'minimalDataset.csv', skipinitialspace = True)
    input = file["input"].values
    output1 = file["output1"].values
    output2 = file["output2"].values
    dataset = tf.data.Dataset.from_tensor_slices((input, (output1, output2))).batch(4)
    
    #create multi output regression net
    input_layer = keras.Input(shape=(1,))
    x = layers.Dense(20, activation="relu")(input_layer)
    x = layers.Dense(60, activation="relu")(x)
    output_layer = layers.Dense(2)(x)
    model = keras.Model(input_layer, output_layer)
    model.compile(optimizer="adam", loss="mean_squared_error")
    
    #train model and make prediction (deliberately overfitting to illustrate problem)
    model.fit(dataset, epochs=500)
    
    prediction = model.predict(dataset) 

minimalDataset.csv and predictions:

input	output1	output2		prediction_output1	prediction_output2
0	-1	1		-0.989956	-0.989964
1	2	0		1.834444	1.845085
2	0	2		0.640249	0.596099
3	1	-1		0.621426	0.646796

If I create two independent dense final layers the second parameter is learned accurately but I get two losses:

output_layer = (layers.Dense(1)(x), layers.Dense(1)(x))

Note: I want to use tf.data.Dataset because I build a 20k image/csv with it and do per-element transformations as preprocessing.

Answer 1

tf.data.Dataset.from_tensor_slices() slices along the first dimension. Because of this the input and output tensors need to be transposed:

dataset = tf.data.Dataset.from_tensor_slices((tf.transpose(input), (tf.transpose([output1, output2])))).batch(4)