profiling.numerical() fails for multi-input models

[schaefes]

Description

profiling.numerical() fails when used with a multi-input hls4ml model. The same model can successfully run inference using hls_model.predict(), so the issue appears to be isolated to the profiling code.

Environment

• hls4ml branch: main
• hls4ml commit: 717a57a309ab85c855ae2985d821d6a643cb7bc6
• Multi-input Keras model converted with hls4ml

Working inference

The model accepts four inputs, which are passed as a list of NumPy arrays:

model_dict = {k: np.ascontiguousarray(v, dtype=np.float64)
              for k, v in model_dict.items()}

hls_inputs = []
for var in model.hls_jet_model.get_input_variables():
    hls_inputs.append(
        np.ascontiguousarray(model_dict[var.name], dtype=np.float64)
    )

with shapes:

basic_input             (4722, 16, 21)
constituent_fraction    (4722, 16)
jet_features            (4722, 2)
pt_mask                 (4722, 16)

Inference works correctly:

model.hls_jet_model.predict(hls_inputs)

Profiling failure

Calling:

wp, wph, ap, aph = profiling.numerical(
    model=model.jet_model,
    hls_model=model.hls_jet_model,
    X=hls_inputs,
)

raises:

Traceback (most recent call last):
...
  File ".../profiling.py", line 546, in numerical
    data = activations_hlsmodel(hls_model, X, fmt='summary', plot=plot)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../profiling.py", line 330, in activations_hlsmodel
    _, trace = model.trace(np.ascontiguousarray(X))
                           ^^^^^^^^^^^^^^^^^^^^^^^
ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 2 dimensions. The detected shape was (4, 4722) + inhomogeneous part.