Input Layer duplicate when loading a keras model

[alejpe15]

Hi all,

when calling:
hls_model = hls4ml.converters.keras_to_hls(cfg)

I am getting an exception:

Interpreting Model
Topology:
Layer name: waveforms, layer type: InputLayer, input shapes: [[None, 30, 30, 5]], output shape: [None, 30, 30, 5]
Interpreting Model
Topology:
Layer name: waveforms, layer type: InputLayer, input shapes: [[None, 30, 30, 5]], output shape: [None, 30, 30, 5]
Layer name: SingleCNN_block_wvf_conv_1_1, layer type: Conv2D, input shapes: [[None, 30, 30, 5]], output shape: [None, 30, 30, 32]
Layer name: SingleCNN_block_wvf_pool_1, layer type: MaxPooling2D, input shapes: [[None, 30, 30, 32]], output shape: [None, 15, 15, 32]
Layer name: SingleCNN_block_wvf_conv_2_1, layer type: Conv2D, input shapes: [[None, 15, 15, 32]], output shape: [None, 15, 15, 32]
Layer name: SingleCNN_block_wvf_pool_2, layer type: MaxPooling2D, input shapes: [[None, 15, 15, 32]], output shape: [None, 7, 7, 32]
Layer name: SingleCNN_block_wvf_conv_3_1, layer type: Conv2D, input shapes: [[None, 7, 7, 32]], output shape: [None, 7, 7, 64]
Layer name: SingleCNN_block_wvf_pool_3, layer type: MaxPooling2D, input shapes: [[None, 7, 7, 64]], output shape: [None, 3, 3, 64]
Layer name: SingleCNN_block_wvf_flatten, layer type: Reshape, input shapes: [[None, 3, 3, 64]], output shape: [None, 576]
Layer name: SingleCNN_block, layer type: LayerGroup, input shapes: [[None, 30, 30, 5]], output shape: [None, 576]
Layer name: fc_particletype_1, layer type: Dense, input shapes: [[None, 576]], output shape: [None, 256]
Layer name: fc_particletype_2, layer type: Dense, input shapes: [[None, 256]], output shape: [None, 64]
Layer name: particletype, layer type: Dense, input shapes: [[None, 64]], output shape: [None, 3]
Layer name: type, layer type: Softmax, input shapes: [[None, 3]], output shape: [None, 3]
Creating HLS model
[...]
Exception: Layer names must be unique: "waveforms" already found in the model graph.

If I get a summary of the keras model with:
model.summary()

Layer (type) Output Shape Param #

waveforms (InputLayer) [(None, 30, 30, 5)] 0

SingleCNN_block (Functiona (None, 576) 29216
l)

fc_particletype_1 (Dense) (None, 256) 147712

fc_particletype_2 (Dense) (None, 64) 16448

particletype (Dense) (None, 3) 195

type (Softmax) (None, 3) 0

=================================================================
Total params: 193571 (756.14 KB)
Trainable params: 193571 (756.14 KB)
Non-trainable params: 0 (0.00 Byte)

So, in principle the model seems to be defined well. But when loading it into hls4ml somehow the input layer seems to duplicate and rises the exception. I am loading the model from a directory including the .pb file.

Any feedback is welcomed.