0

I am currently building a neural network model using Keras' Functional API and a dataset with multiple independent variables and a single categorical target variable using the following code.

input_layer = keras.Input(shape=(89,), name="input_layer")
dense_1 = keras.layers.Dense(50, name = 'dense_1')(input_layer)
dense_2 = keras.layers.Dense(50, name = 'dense_2')(dense_1)
classification_output_1 = keras.layers.Dense(31, activation = 'softmax', name = 'classification_output_1')(dense_2)
model = keras.Model(inputs = input_layer, outputs = [regression_output_1, classification_output_1])


model.compile(
    optimizer = "adam",
    loss = 'sparse_categorical_crossentropy'
)

model_1 = model.fit(
    X_train,
    y_train["Category_Target"],
    epochs = 10,
    batch_size = 50,
    verbose = 1
)

y_pred = model_1.predict(X_train)
y_pred = pd.DataFrame(y_pred[0])
y_pred.columns = [i for i in pd.get_dummies(y_train["Category_Target"]).columns]

As y_pred is in the form of a one-hot vector I presumed that using pd.get_dummies().columns would get me the class labels that I need. My question is, is this method reliable at all, and if not, is there any other way to get class labels for predicted values? The old method of getting labels for classes (i.e. predict_classes) is deprecated and I can't really find a reliable way to get them. I've also considered other methods such as:

y_pred.columns = list(y_train.drop_duplicates(by = "Category_Target")["Category_Target"])

Neither of the methods that I've tried seem to be a surefire way of labelling classes correctly

Daniel Coben
  • 15
  • 5

1 Answers1

1

Here is an example of a round trip going from a set of labels and an array of strings with those labels and back.

idx_to_labels = ['lions','tigers','bears']

target_strings = ['tigers','lions','lions','bears']

target_idx = tf.constant([labels.index(x) for x in target_labels], dtype=tf.int64)

target_one_hot = keras.utils.to_categorical(target_idx)

# this is the idiomatic way of going from one-hot to index
recovered_target_idx = tf.argmax(target_one_hot, axis=-1)

# if you have an ordered label list, this is a quick way to map a list with the integers back to strings
recovered_target_labels = [idx_to_labels[x] for x in  recovered_target_idx.numpy()]

If your target is one-hot encoded, then you need to use the CategoricalCrossentropy loss. SparseCrossentropy expects the targets to be indexes (in my example above, like the target_idx variable. If your evaluation results are poor, this may be the underlying cause.

https://keras.io/api/losses/probabilistic_losses/#categoricalcrossentropy-function

Mike Holcomb
  • 403
  • 3
  • 9