Support multiple outputs in pytorch parser (fastmachinelearning#1151)

* support multiple outputs in pytorch parser

* pre-commit

* [pre-commit.ci] auto fixes from pre-commit hooks

* fix missing input names in squeeze layers

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Author: JanFSchulte

hls4ml/converters/pytorch/reshape.py

@@ -38,6 +38,7 @@ def parse_squeeze_layer(operation, layer_name, input_names, input_shapes, node,
     layer = {}
     layer['class_name'] = 'Reshape'
     layer['name'] = layer_name
+    layer['inputs'] = input_names
 
     if len(node.args) > 1 or len(node.kwargs) > 0:  # 'dim' argument is specified
         output_shape = [i for i in input_shapes[0]]

hls4ml/converters/pytorch_to_hls.py

@@ -151,6 +151,7 @@ def parse_pytorch_model(config, verbose=True):
     inputs_map = {}
 
     input_layers = []
+    output_layers = []
 
     # Output shape tracking
     output_shapes = {}
@@ -399,12 +400,23 @@ def parse_pytorch_model(config, verbose=True):
     if len(input_layers) == 0:
         input_layers = None
 
-    return layer_list, input_layers
+    for layer in layer_list:
+        if layer['class_name'] == 'InputLayer':
+            continue
+        is_input = False
+        for lay in layer_list:
+            if 'inputs' not in lay.keys():
+                continue
+            if layer['name'] in lay['inputs']:
+                is_input = True
+        if not is_input:
+            output_layers.append(layer['name'])
+    return layer_list, input_layers, output_layers
 
 
 @requires('_torch')
 def pytorch_to_hls(config):
-    layer_list, input_layers = parse_pytorch_model(config)
+    layer_list, input_layers, output_layers = parse_pytorch_model(config)
     print('Creating HLS model')
-    hls_model = ModelGraph(config, layer_list, inputs=input_layers)
+    hls_model = ModelGraph(config, layer_list, inputs=input_layers, outputs=output_layers)
     return hls_model

hls4ml/utils/config.py

@@ -368,6 +368,7 @@ def config_from_pytorch_model(
     (
         layer_list,
         _,
+        _,
     ) = parse_pytorch_model(config, verbose=False)
 
     def make_layer_config(layer):