Skip to content

LayerName configuration capitalization

Created by: jmitrevs

I noticed that functions accessing the configuration use layer.name.lower() as the key, for example in:

    def get_layer_config(self, layer):
        hls_config = self.config['HLSConfig']
        layer_config = {}

        type_config = hls_config.get('LayerType', {}).get(layer.class_name, None)
        if type_config is not None:
            layer_config.update(type_config)

        name_config = hls_config.get('LayerName', {}).get(layer.name.lower(), None)
        if name_config is not None:
            layer_config.update(name_config)

        return layer_config

But in config_from_keras_model when you use the granularity='name' feature, the name is used directly, without being made .lower(): https://github.com/fastmachinelearning/hls4ml/blob/master/hls4ml/utils/config.py#L299-L305. So the generated configuration does not pass the information correctly because it could well have capital letters in the keys. This needs to be made consistent. I think there are two reasonable options:

  1. Make the comparisons case-sensitive
  2. Make the name comparison case-insensitive, even if they are entered in mixed case in the configuration

One could add a .lower() in config_from_keras_model as well, but I think that's a worse option, because it would be hard to enforce in hand-generated configurations. For the second option, I think there could a key "fixing" step in the HLSConfig constructor to make the keys lower case.

What do people think?