I would like to implement a metric in TensorFlow based on the combined results of two outputs.
My model takes in a 100-character string and returns two outputs (called flavour and form) based on this string. These outputs are both softmax probabilities that are compared with a one-hot encoded vectors (standard classification). The code for the model is:
inputs = Input(shape=(None,))
input_embeddings = Embedding(vocab_size, embedding_size, mask_zero=True)(inputs)
shared_lstm = Bidirectional(LSTM(units, return_sequences=True, dropout=0.2))(input_embeddings)
fl_lstm = Bidirectional(LSTM(units, dropout=0.2))(shared_lstm)
fl_dense = Dense(flavour_size, activation='softmax', name='flavour')(fl_lstm)
fo_lstm = Bidirectional(LSTM(units, dropout=0.2))(shared_lstm)
fo_dense = Dense(form_size, activation='softmax', name='form')(fo_lstm)
split_shared_model = Model(inputs=inputs, outputs=[fl_dense, fo_dense])
Here is a flow diagram of the model's architecture:
Currently, I am compiling and fitting as follows:
split_shared_model.compile(optimizer='adam', loss=CategoricalCrossentropy(),
metrics=['accuracy'])
split_shared_model.fit(X_train, [fl_train, fo_train],
batch_size=32,
epochs=10)
Each of the individual outputs (flavour and form) reach score around 96% accuracy on the test set. However, I would like to create a metric that combines these two predictions and assess them together. Something that might look a little bit like:
def combined_accuracy(fl_true, fo_true, fl_pred, fo_pred):
correct_fl = K.equal(fl_true, K.round(fl_pred))
correct_fo = K.equal(fo_true, K.round(fo_pred))
combined = tf.logical_and(correct_fl, correct_fo)
return K.mean(combined)
I've looked at the Keras documentation (https://keras.io/api/metrics/#creating-custom-metrics), but it seems as if the custom metric is applied to each output individually.
What can I try next?