update keras activation parsing, especially leaky relu (#1085)

* update keras activation parsing, especially leaky relu

* fix assert in pytest as in oneAPI branch

Author: jmitrevs

hls4ml/converters/keras/core.py

@@ -67,7 +67,8 @@ def parse_activation_layer(keras_layer, input_names, input_shapes, data_reader):
         layer['class_name'] = 'ELU'  # always use ELU type for elu, even if passed as activation
 
     if layer['class_name'] == 'LeakyReLU':
-        layer['activ_param'] = keras_layer['config'].get('alpha', 0.3)
+        # the name changes for version 3
+        layer['activ_param'] = keras_layer['config'].get('negative_slope', keras_layer['config'].get('alpha', 0.3))
     elif layer['class_name'] == 'ThresholdedReLU':
         layer['activ_param'] = keras_layer['config'].get('theta', 1.0)
     elif layer['class_name'] == 'ELU':
@@ -83,6 +84,10 @@ def parse_activation_layer(keras_layer, input_names, input_shapes, data_reader):
         layer['class_name'] = 'HardActivation'
     if layer['class_name'] == 'Softmax':
         layer['axis'] = keras_layer['config'].get('axis', -1)
+    if layer['class_name'] == 'Activation' and layer['activation'] == 'leaky_relu':
+        layer['class_name'] = 'LeakyReLU'
+        # The parameter name changes for API v3; the default is different than in LeakyReLU layer
+        layer['activ_param'] = keras_layer['config'].get('negative_slope', keras_layer['config'].get('alpha', 0.2))
 
     return layer, [shape for shape in input_shapes[0]]
 

hls4ml/converters/keras_to_hls.py

@@ -297,26 +297,18 @@ def parse_keras_model(model_arch, reader):
         layer_list.append(layer)
         if 'activation' in layer and layer['class_name'] not in activation_layers + recurrent_layers:  # + qkeras_layers:
             act_layer = {}
+            act_details = layer['activation']
             # Workaround for QKeras activations passed as an argument
-            if isinstance(layer['activation'], dict):
-                act_details = layer['activation']
+            if isinstance(act_details, dict):
                 act_layer['class_name'] = 'QActivation'
                 act_layer['config'] = {
                     'name': layer['name'] + '_' + act_details['class_name'],
                     'activation': act_details,
                 }
-                act_layer, output_shape = layer_handlers['QActivation'](act_layer, None, [output_shape], reader)
             else:
-                act_layer['name'] = layer['name'] + '_' + layer['activation']
-                act_layer['activation'] = layer['activation']
-                if 'activ_param' in layer:
-                    act_layer['activ_param'] = layer['activ_param']
-                    act_layer['class_name'] = layer['activation']
-                elif layer['activation'] == 'softmax':
-                    act_layer['class_name'] = 'Softmax'
-                    act_layer['axis'] = -1
-                else:
-                    act_layer['class_name'] = 'Activation'
+                act_layer['class_name'] = 'Activation'
+                act_layer['config'] = {'name': layer['name'] + '_' + act_details, 'activation': act_details}
+            act_layer, output_shape = layer_handlers[act_layer['class_name']](act_layer, None, [output_shape], reader)
             inputs_map[layer['name']] = act_layer['name']
             if output_layers is not None and layer['name'] in output_layers:
                 output_layers = [act_layer['name'] if name == layer['name'] else name for name in output_layers]

test/pytest/test_activations.py

@@ -19,6 +19,7 @@
     [
         (ReLU(), 'relu'),
         (LeakyReLU(alpha=1.5), 'leaky_relu'),
+        (Activation('leaky_relu'), 'leaky_relu_act'),
         (ThresholdedReLU(theta=0.75), 'threshold_relu'),
         (ELU(alpha=1.25), 'elu'),
         (Activation('selu'), 'selu'),