preprocess.py

def preprocess_spatial_features(features, screen=True):
    """Embed categorical spatial features, log transform numeric features."""
    # ...
    # ...
    preprocess_ops = []
    for index, (feature_type, scale) in enumerate(feature_specs):
        layer = transposed[:, :, :, index]

        if feature_type == sc2_features.FeatureType.CATEGORICAL:
            # one-hot encode in channel dimension -> 1x1 convolution
            one_hot = tf.one_hot(
                layer,
                depth=scale,
                axis=-1,
                name="one_hot")

            embed = tf.layers.conv2d(
                inputs=one_hot,
                filters=1,
                kernel_size=[1, 1],
                strides=[1, 1],
                padding="SAME")

            preprocess_ops.append(embed)
        else:
            transform = tf.log(
                tf.cast(layer, tf.float32) + 1.,
                name="log")

            preprocess_ops.append(tf.expand_dims(transform, -1))

    preprocessed = tf.concat(preprocess_ops, -1)
    return preprocessed