Custom Lost Function in Tensorflow Keras has no Gradient - Stack Overflow

admin管理员组

文章数量:1398999

I'm trying to write a custom loss function to train a neural net with Keras/Tensorflow. However, when I train the model, I get the following error:

ValueError: No gradients provided for any variable: ['dense/kernel:0', 'dense/bias:0', 'dense_1/kernel:0', 'dense_1/bias:0', 'dense_2/kernel:0', 'dense_2/bias:0', 'dense_3/kernel:0', 'dense_3/bias:0', 'dense_4/kernel:0', 'dense_4/bias:0'].

Below is my loss function. The goal is to convert a tensor that has probabilities for coordinate boxes into a coordinate prediction and then compute the distance between the coordinates to use that as my loss.

def haversine_distance(lat1, lon1, lat2, lon2):

    R = 3958.8  # Radius of the Earth in miles
    

    # lat1, lon1, lat2, lon2 = map(tf.math.radians, [lat1, lon1, lat2, lon2])

    dlat = lat2 - lat1
    dlon = lon2 - lon1

    a = tf.sin(dlat / 2.0) ** 2 + tf.cos(lat1) * tf.cos(lat2) * tf.sin(dlon / 2.0) ** 2
    c = 2 * tf.atan2(tf.sqrt(a), tf.sqrt(1 - a))

    return R * c

def custom_loss(y_true, y_pred):

    # Convert box probabilities into coordinates
    true_coords = tf.map_fn(get_coordinates, tf.cast(y_true, dtype=tf.float32))  # (batch_size, 2)
    pred_coords = tf.map_fn(get_coordinates, tf.cast(y_pred, dtype=tf.float32))  # (batch_size, 2)

    lat1, lon1 = tf.unstack(true_coords, axis=1)
    lat2, lon2 = tf.unstack(pred_coords, axis=1)

    return haversine_distance(lat1, lon1, lat2, lon2)

def get_coordinates(prediction, box_thresh=0.03):
    total = 0
    lat_p = 0
    lon_p = 0
    for i in range(x_boxes*y_boxes):
        pred = tf.keras.backend.get_value(prediction[i])
        if pred > box_thresh:
            total = total + pred
    for i in range(x_boxes*y_boxes):
        pred = tf.keras.backend.get_value(prediction[i])
        if pred > box_thresh:
            lat_p += pred * centers[i][0] / total
            lon_p += pred * centers[i][1] / total
    result = [lat_p, lon_p]
        
    return tf.convert_to_tensor(result, dtype=tf.float32)

I think it has to do with the function not being differentiable, but I'm not sure why it wouldn't be. Thanks for the help!

本文标签： Custom Lost Function in Tensorflow Keras has no GradientStack Overflow