为什么 keras model 如果与 one-hot 标签和 categorical_crossentropy 和 softmax output 一起使用，则将所有预测为 1

Question

I have a simple tf.keras model:

inputs = keras.Input(shape=(9824,))
dense = layers.Dense(512, activation=keras.activations.relu, kernel_initializer=init)
x = dense(inputs)
x = layers.Dense(512, activation=keras.activations.relu)(x)
outputs = layers.Dense(3, activation=keras.activations.softmax)(x)
model = keras.Model(inputs=inputs, outputs=outputs)

When I compile it with sparse categorical crossentropy and actual labels it works as expected. But when I tried to one-hot encode the labels (with tf.keras.utils.to_categorical ) and use categorical_crossentropy (so I could use recall and precision as metrics during training), the model predicts everything as ones:

>>>print(predictions)
[[1. 1. 1.]
 [1. 1. 1.]
 [1. 1. 1.]
 ...
 [1. 1. 1.]
 [1. 1. 1.]
 [1. 1. 1.]]

If I understand correctly, softmax activation in the output layer should cause the output to be in range (0,1) and sum to 1. So, how is it possible that the class predictions are all 1? I was searching for an answer for hours but to no avail.

EDIT

Here is a minimalistic example .

I forgot to mention that I use scikeras package. Based on examples in the documentation , I assume the model is compiled implicitly. Here is the classifier constructor:

clf = KerasClassifier(
    model=keras_model_target,
    loss=SparseCategoricalCrossentropy(),
    name="model_target",
    optimizer=Adam(),
    init=GlorotUniform(),
    metrics=[SparseCategoricalAccuracy()],
    epochs=5,
    batch_size=128
)

I fit the model with

result = clf.fit(x_train, y_train)

and predict with:

predictions = clf.predict(x)

Answer 1

This was a bug in SciKeras that was fixed with the v0.3.1 release. Updating to the latest version should fix the issue.

As for the bug itself, it was due to how we were indexing numpy arrays, see this diff for details.