Value error in multi-class text classification model with keras and LSTM layer

[spacyuj]

I'm trying to build a multi-class text classification model using a Keras Sequential API with an LSTM layer, but I'm encountering a ValueError: Unknown metric function: 'accuracy'. I'm using a custom metric for F1-score and the standard accuracy metric, but it seems there is an issue with how I'm compiling the model

Here is the code:

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Embedding, LSTM, Dense, Dropout
import numpy as np
from sklearn.model_selection import train_test_split

vocab_size = 10000
embedding_dim = 128
max_length = 100
num_classes = 5

X = np.random.randint(0, vocab_size, (1000, max_length))
y = np.random.randint(0, num_classes, 1000)

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

def f1_score_metric(y_true, y_pred):
    true_positives = tf.keras.backend.sum(tf.keras.backend.round(tf.keras.backend.clip(y_true * y_pred, 0, 1)))
    predicted_positives = tf.keras.backend.sum(tf.keras.backend.round(tf.keras.backend.clip(y_pred, 0, 1)))
    possible_positives = tf.keras.backend.sum(tf.keras.backend.round(tf.keras.backend.clip(y_true, 0, 1)))
    
    precision = true_positives / (predicted_positives + tf.keras.backend.epsilon())
    recall = true_positives / (possible_positives + tf.keras.backend.epsilon())
    
    f1_score_val = 2 * (precision * recall) / (precision + recall + tf.keras.backend.epsilon())
    return f1_score_val


model = Sequential([
    Embedding(vocab_size, embedding_dim, input_length=max_length),
    LSTM(64),
    Dropout(0.5),
    Dense(num_classes, activation='softmax')
])


model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy', f1_score_metric])

model.summary()

[Mariihmp]

Have to cast y_true to float to match y_pred and then one-hot encode the predicted classes, and then you can calculate tp, fp, and fn
Here is the f1_score_metric function:

def f1_score_metric(y_true, y_pred):

    y_true = tf.cast(y_true, tf.float32)
    y_pred_classes = tf.argmax(y_pred, axis=-1)
    
    y_pred_one_hot = tf.one_hot(y_pred_classes, depth=num_classes)
    
    true_positives = tf.reduce_sum(y_true * y_pred_one_hot, axis=0)
    false_positives = tf.reduce_sum(y_pred_one_hot * (1 - y_true), axis=0)
    false_negatives = tf.reduce_sum((1 - y_pred_one_hot) * y_true, axis=0)
    
    precision = true_positives / (true_positives + false_positives + tf.keras.backend.epsilon())
    recall = true_positives / (true_positives + false_negatives + tf.keras.backend.epsilon())
    
    f1_score_val = 2 * (precision * recall) / (precision + recall + tf.keras.backend.epsilon())
    
    return tf.reduce_mean(f1_score_val)

[spacyuj]

I tested this and it's fine. Thanks for your help