0
$\begingroup$

I am trying to label code snippets and I base on this article: https://arxiv.org/pdf/1906.01032.pdf

My dataset is just code snippets (tokenized as ascii characters) and 500 different labels from StackOverflow. Currently I have around 1,600,000 samples after filtering these with negative votes and less than 10 characters

This is my current implementation of network architecture in TensorFlow:

def build_cnn(config, hparams) -> tf.keras.Model:
    params = config["params"]
    inputs = tf.keras.layers.Input((1024, ))
    x = tf.keras.layers.Embedding(input_dim=1024, output_dim=16)(inputs)

    conv_outputs = []
    for filters, kernel in [
        (128, 2),
        (192, 3),
        (256, 4),
        (512, 5)
    ]:
        data = tf.keras.layers.Conv1D(filters=filters, kernel_size=kernel, padding="valid")(x)
        data = tf.keras.layers.BatchNormalization()(data)
        data = tf.keras.layers.Conv1D(filters=filters, kernel_size=kernel, padding="valid")(data)
        data = tf.keras.layers.Lambda(lambda a: tf.reduce_sum(a, axis=1))(data)
        conv_outputs.append(data)

    # Concatenation: output size is sum of all convolution filters
    concat_output = tf.keras.layers.Concatenate()(conv_outputs)
    concat_output = tf.keras.layers.BatchNormalization()(concat_output)

    concat_output = tf.keras.layers.Dense(480, activation="relu")(concat_output)
    concat_output = tf.keras.layers.BatchNormalization()(concat_output)

    concat_output = tf.keras.layers.Dense(480, activation="relu")(concat_output)
    concat_output = tf.keras.layers.BatchNormalization()(concat_output)

    outputs = tf.keras.layers.Dense(500, activation="sigmoid")(concat_output)

    return tf.keras.Model(inputs=inputs, outputs=outputs)

As my metric I use F1Score from tensorflow_addons:

model.compile(
    loss=tf.keras.losses.BinaryCrossentropy(),
    optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
    metrics=[
        F1Score(num_classes=500, average="micro", threshold=0.5, name="f1_micro"),
        F1Score(num_classes=500, average="macro", threshold=0.5, name="f1_macro"),
    ]
)

After 5 epochs results look like this (later it wasn't really improving): enter image description here

My first idea was to apply class weights but it didn't help. Next idea was an experiment with only 10 tags (still multilabel) and this is the result:

{
    "step": 12,
    "loss": 0.2092941254377365,
    "f1_micro": 0.6354409456253052,
    "f1_macro": 0.5389766097068787,
    "val_loss": 0.25258970260620117,
    "val_f1_micro": 0.547073483467102,
    "val_f1_macro": 0.4620901644229889
}

Last thing: after I saw some results with evaluation I saw that most of the outputs are empty or only 1 class, but using smaller threshold didn't increase the metric

Any idea how to improve the model for bigger number of tags? is it the metric?

$\endgroup$

0

You must log in to answer this question.

Browse other questions tagged .