Understanding config precedence between model and layer in config_from_keras_model()

[rf-ml]

I'm trying to understand that if I specify this:

hls_config_q = hls4ml.utils.config_from_keras_model(qmodel, granularity='name', backend='VivadoAccelerator')
hls_config_q['Model']['ReuseFactor'] = 1000
hls_config_q['Model']['Strategy'] = 'Resource'
hls_config_q['LayerName']['output_softmax']['Strategy'] = 'Stable'

does it mean that every layer will have a ReuseFactor of 1000?

I'm confused because if I then printout the hls_config_q I can see that while Model has taken on those values other layers of the network haven't e.g. they have a ReuseFactor of 1:

Model
  Precision:         fixed<16,6>
  ReuseFactor:       1000
  Strategy:          Resource
  BramFactor:        1000000000
  TraceOutput:       False
LayerName
  input_1
    Trace:           False
    Precision
      result:        fixed<16,6>
  fused_convbn_0
    Trace:           False
    Precision
      result:        fixed<16,6>
      weight:        fixed<6,1>
      bias:          fixed<6,1>
      accum:         fixed<16,6>
    ReuseFactor:     1
    ParallelizationFactor:1
    ConvImplementation:LineBuffer
  fused_convbn_0_linear
    Trace:           False
    Precision
      result:        fixed<16,6>
      table:         fixed<18,8>
    ReuseFactor:     1
    TableSize:       1024
  conv_act_0
    Trace:           False
    Precision
      result:        ufixed<6,0,RND_CONV,SAT>
      table:         fixed<18,8>
    ReuseFactor:     1
    TableSize:       1024
  pool_0
    Trace:           False
    Precision
      result:        fixed<16,6>
      accum:         fixed<16,6>
    ReuseFactor:     1
    ConvImplementation:LineBuffer
.....

So under the hood does the value used by Model level take precedent and applied to each level OR if not what exactly is the purpose of specifying Reuse at the model level?

Thanks

[bo3z]

You are right; if specified for the model and left unspecified for individual layers, the value 1000 will apply to the individual layers. You can verify this by running the hls4ml model conversion and inspecting the file parameters.h in your output directory. You will see the layer configuration that actually is used when compiling the model.

[rf-ml]

Thank you