14f0a88cc84dbb29cb2f2eb6e502dc82a77b0714,finetune/network_modules.py,,sequence_labeler,#Any#Any#Any#Any#Any#Any#Any#,243

Before Change

            log_likelihood = tf.constant(0.)

        class_weights = kwargs.get("class_weights")
        if class_weights is not None:
            flat_loss = tf.reshape(log_likelihood, [-1, n_targets])
            flat_targets = tf.reshape(targets, [-1])
            one_hot_targets = tf.one_hot(flat_targets, depth=n_targets)
            log_likelihood = _apply_class_weight(flat_loss, one_hot_targets, class_weights)

        return {
            "logits": logits,
            "losses": -log_likelihood,
            "predict_params": {

After Change

        transition_params = tf.get_variable("Transition_matrix", shape=[n_targets, n_targets])

        class_weights = kwargs.get("class_weights")
        if class_weights is not None:
            logits = class_reweighting(class_weights)(logits)
        
        if train:
            log_likelihood, _ = crf_log_likelihood(
                logits, 
                targets,

Instances