f9b98760f445fc0219cfc9c4cada2b5f9d35ab1b,basic/model.py,Model,_build_loss,#Model#,219

Before Change

        if config.wy:
            if config.na:
                na = tf.reshape(self.na, [-1, 1])
                concat_y = tf.concat(1, [na, tf.reshape(self.wy, [-1, M * JX])])
                losses = tf.nn.softmax_cross_entropy_with_logits(self.concat_logits, tf.cast(concat_y, "float"))
            else:
                losses = tf.nn.softmax_cross_entropy_with_logits(
                    self.logits2, tf.cast(tf.reshape(self.wy, [-1, M * JX]), "float"))
            ce_loss = tf.reduce_mean(loss_mask * losses)
            tf.add_to_collection("losses", ce_loss)

        else:
            if config.na:
                na = tf.reshape(self.na, [-1, 1])

After Change

            damp_ratio = num_pos / num_neg
            dampened_losses = losses * (
                (tf.cast(self.x_mask, "float") - tf.cast(self.wy, "float")) * damp_ratio + tf.cast(self.wy, "float"))
            new_losses = tf.reduce_sum(dampened_losses, [1, 2])
            ce_loss = tf.reduce_mean(loss_mask * new_losses)
            
            if config.na:
                na = tf.reshape(self.na, [-1, 1])
                concat_y = tf.concat(1, [na, tf.reshape(self.wy, [-1, M * JX])])
                losses = tf.nn.softmax_cross_entropy_with_logits(
                    self.concat_logits, tf.cast(concat_y, "float") / tf.reduce_sum(tf.cast(self.wy, "float")))
            else:
                losses = tf.nn.softmax_cross_entropy_with_logits(
                    self.logits2, tf.cast(tf.reshape(self.wy, [-1, M * JX]), "float") / tf.reduce_sum(tf.cast(self.wy, "float")))
            ce_loss = tf.reduce_mean(loss_mask * losses)
            
            tf.add_to_collection("losses", ce_loss)

        else:
            if config.na:
                na = tf.reshape(self.na, [-1, 1])

Instances