How Relay ir Represents GRU graph?

[witiny]

I’m wonder how Relay ir Represents GRU? I read the _convert_gru function in keras frontend, I find the ht-1 is expressed as ： def _convert_gru(inexpr, keras_layer, etab): _check_data_format(keras_layer) if not isinstance(inexpr, list): buf = np.zeros((1, keras_layer.units), ‘float32’) h_tm1 = etab.new_const(buf) inexpr = [inexpr, h_tm1] in_data = inexpr[0] h_tm1_op = inexpr[1] while I find the ht-1 is replaced by etab.new_const(np.zeros()), and i didn’t find any edge between ht-1 and output in relay ir !!

[witiny]

def _convert_gru(inexpr, keras_layer, etab):

_check_data_format(keras_layer)

if not isinstance(inexpr, list):

    buf = np.zeros((1, keras_layer.units), 'float32')

    h_tm1 = etab.new_const(buf)

    inexpr = [inexpr, h_tm1]

in_data = inexpr[0]

h_tm1_op = inexpr[1]

[witiny]

@haichen @comaniac @junrushao

[witiny]

@Huyuwei ~~~~~~~~~~~~~

[witiny]