Fix profiling for multi-input models

hls4ml/model/profiling.py

@@ -327,7 +327,10 @@ def activations_hlsmodel(model, X, fmt='summary', plot='boxplot'):
     elif fmt == 'summary':
         data = []
 
-    _, trace = model.trace(np.ascontiguousarray(X))
+    if isinstance(X, (list, tuple)):
+        _, trace = model.trace([np.ascontiguousarray(x) for x in X])
+    else:
+        _, trace = model.trace(np.ascontiguousarray(X))
 
     if len(trace) == 0:
         raise RuntimeError('ModelGraph must have tracing on for at least 1 layer (this can be set in its config)')
@@ -457,8 +460,11 @@ def numerical(model=None, hls_model=None, X=None, plot='boxplot'):
     Args:
         model (optional): Keras of PyTorch model. Defaults to None.
         hls_model (ModelGraph, optional): The ModelGraph to profile. Defaults to None.
-        X (ndarray, optional): Test data on which to evaluate the model to profile activations.
-            Must be formatted suitably for the ``model.predict(X)``. Defaults to None.
+        X (ndarray or list of ndarray, optional): Test data on which to evaluate the model to profile
+            activations. Must be formatted suitably for the ``model.predict(X)``. For models with
+            multiple inputs, pass a list holding one array per model input, ordered as
+            ``hls_model.get_input_variables()`` (i.e. the order of the inputs of the original model).
+            Defaults to None.
         plot (str, optional): The type of plot to produce. Options are: 'boxplot' (default), 'violinplot', 'histogram',
             'FacetGrid'. Defaults to 'boxplot'.
 
@@ -682,7 +688,9 @@ def compare(keras_model, hls_model, X, plot_type='dist_diff'):
     Args:
         keras_model: Original keras model.
         hls_model (ModelGraph): Converted ModelGraph, with "Trace:True" in the configuration file.
-        X (ndarray): Input tensor for the model.
+        X (ndarray or list of ndarray): Input tensor for the model. For models with multiple inputs,
+            pass a list holding one array per model input, ordered as
+            ``hls_model.get_input_variables()`` (i.e. the order of the inputs of the original model).
         plot_type (str, optional): Different methods to visualize the y_model and y_sim differences.
             Possible options include:
             - 'norm_diff':: square root of the sum of the squares of the differences between each output vectors.
@@ -696,7 +704,10 @@ def compare(keras_model, hls_model, X, plot_type='dist_diff'):
     # Take in output from both models
     # Note that each y is a dictionary with structure {"layer_name": flattened ouput array}
     ymodel = get_ymodel_keras(keras_model, X)
-    _, ysim = hls_model.trace(X)
+    if isinstance(X, (list, tuple)):
+        _, ysim = hls_model.trace([np.ascontiguousarray(x) for x in X])
+    else:
+        _, ysim = hls_model.trace(np.ascontiguousarray(X))
 
     print('Plotting difference...')
     f = plt.figure()

test/pytest/test_keras_v3_profiling.py

@@ -86,7 +86,6 @@ def test_keras_v3_numerical_profiling_conv_model():
 
 
 @pytest.mark.skipif(not __keras_profiling_enabled__, reason='Keras 3.0 or higher is required')
-@pytest.mark.skip(reason='convert_from_config needs update for Keras v3 model serialization format')
 def test_keras_v3_numerical_profiling_with_hls_model(test_case_id):
     """Test numerical profiling with both Keras v3 model and hls4ml model."""
     import hls4ml
@@ -101,8 +100,10 @@ def test_keras_v3_numerical_profiling_with_hls_model(test_case_id):
     # Generate test data
     X_test = np.random.rand(100, 8).astype(np.float32)
 
-    # Create hls4ml model
+    # Create hls4ml model, tracing every layer so that activations can be profiled
     config = hls4ml.utils.config_from_keras_model(model, granularity='name')
+    for layer in config['LayerName'].keys():
+        config['LayerName'][layer]['Trace'] = True
     hls_model = hls4ml.converters.convert_from_keras_model(
         model,
         hls_config=config,
@@ -121,6 +122,43 @@ def test_keras_v3_numerical_profiling_with_hls_model(test_case_id):
     assert aph is not None  # HLS model activations (after optimization)
 
 
+@pytest.mark.skipif(not __keras_profiling_enabled__, reason='Keras 3.0 or higher is required')
+def test_keras_v3_numerical_profiling_multiple_inputs(test_case_id):
+    """Test numerical profiling of a model with more than one input, of differing shapes."""
+    import hls4ml
+
+    input_1 = keras.Input(shape=(16, 21), name='basic_input')
+    input_2 = keras.Input(shape=(1,), name='jet_pt')
+    x = keras.layers.Flatten()(input_1)
+    x = keras.layers.Dense(8, activation='relu')(x)
+    x = keras.layers.Concatenate()([x, input_2])
+    outputs = keras.layers.Dense(4, activation='softmax')(x)
+    model = keras.Model(inputs=[input_1, input_2], outputs=outputs)
+    model.compile(optimizer='adam', loss='categorical_crossentropy')
+
+    # One array per model input, with different shapes
+    X_test = [np.random.rand(10, 16, 21).astype(np.float32), np.random.rand(10, 1).astype(np.float32)]
+
+    config = hls4ml.utils.config_from_keras_model(model, granularity='name', backend='Vivado')
+    for layer in config['LayerName'].keys():
+        config['LayerName'][layer]['Trace'] = True
+
+    hls_model = hls4ml.converters.convert_from_keras_model(
+        model,
+        hls_config=config,
+        output_dir=str(Path(__file__).parent / test_case_id),
+        backend='Vivado',
+    )
+    hls_model.compile()
+
+    wp, wph, ap, aph = numerical(model, hls_model=hls_model, X=X_test)
+
+    assert wp is not None  # Keras model weights (before optimization)
+    assert wph is not None  # HLS model weights (after optimization)
+    assert ap is not None  # Keras model activations (before optimization)
+    assert aph is not None  # HLS model activations (after optimization)
+
+
 @pytest.mark.skipif(not __keras_profiling_enabled__, reason='Keras 3.0 or higher is required')
 def test_keras_v3_numerical_profiling_batch_norm():
     """Test numerical profiling with Keras v3 model containing BatchNormalization."""