Merge pull request #1151 from onnx/tom/OutputFrozenGraph

Added output_frozen_graph flag and marked --fold_const as deprecated

Author: TomWildenhain-Microsoft

README.md

@@ -145,8 +145,10 @@ python -m tf2onnx.convert
     [--target TARGET]
     [--custom-ops list-of-custom-ops]
     [--fold_const]
+    [--large_model]
     [--continue_on_error]
     [--verbose]
+    [--output_frozen_graph]
 ```
 
 ### Parameters
@@ -199,13 +201,17 @@ Only valid with parameter `--saved_model`. If a model contains a list of concret
 
 Only valid with parameter `--saved_model`. When set, creates a zip file containing the ONNX protobuf model and large tensor values stored externally. This allows for converting models that exceed the 2 GB protobuf limit.
 
+#### --output_frozen_graph
+
+Saves the frozen tensorflow graph to file.
+
 #### --target
 
 Some models require special handling to run on some runtimes. In particular, the model may use unsupported data types. Workarounds are activated with ```--target TARGET```. Currently supported values are listed on this [wiki](https://github.com/onnx/tensorflow-onnx/wiki/target). If your model will be run on Windows ML, you should specify the appropriate target value.
 
 #### --fold_const
 
-When set, TensorFlow fold_constants transformation is applied before conversion. This benefits features including Transpose optimization (e.g. Transpose operations introduced during tf-graph-to-onnx-graph conversion will be removed), and RNN unit conversion (for example LSTM). Older TensorFlow version might run into issues with this option depending on the model.
+Deprecated. Constant folding is always enabled.
 
 ### <a name="summarize_graph"></a>Tool to get Graph Inputs & Outputs
 

tests/test_convert.py

@@ -10,13 +10,18 @@
 from tf2onnx import convert
 from common import check_tf_min_version
 
-def run_test_case(args):
+def run_test_case(args, paths_to_check=None):
     """ run case and clean up """
+    if paths_to_check is None:
+        paths_to_check = [args[-1]]
     sys.argv = args
     convert.main()
-    ret = os.path.exists(args[-1])
-    if ret:
-        os.remove(args[-1])
+    ret = True
+    for p in paths_to_check:
+        if os.path.exists(p):
+            os.remove(p)
+        else:
+            ret = False
     return ret
 
 
@@ -33,6 +38,20 @@ def test_convert_saved_model(self):
                                        '--output',
                                        'converted_saved_model.onnx']))
 
+    def test_convert_output_frozen_graph(self):
+        """ convert saved model """
+        self.assertTrue(run_test_case(['',
+                                       '--saved-model',
+                                       'tests/models/regression/saved_model',
+                                       '--tag',
+                                       'serve',
+                                       '--output',
+                                       'converted_saved_model.onnx',
+                                       '--output_frozen_graph',
+                                       'frozen_graph.pb'
+                                      ],
+                                      paths_to_check=['converted_saved_model.onnx', 'frozen_graph.pb']))
+
     @check_tf_min_version("2.2")
     def test_convert_large_model(self):
         """ convert saved model to onnx large model format """

tf2onnx/convert.py

@@ -68,7 +68,8 @@ def get_args():
     parser.add_argument("--continue_on_error", help="continue_on_error", action="store_true")
     parser.add_argument("--verbose", "-v", help="verbose output, option is additive", action="count")
     parser.add_argument("--debug", help="debug mode", action="store_true")
-    parser.add_argument("--fold_const", help="enable tf constant_folding transformation before conversion",
+    parser.add_argument("--output_frozen_graph", help="output frozen tf graph to file")
+    parser.add_argument("--fold_const", help="Deprecated. Constant folding is always enabled.",
                         action="store_true")
     # experimental
     parser.add_argument("--inputs-as-nchw", help="transpose inputs as from nhwc to nchw")
@@ -148,6 +149,8 @@ def main():
         const_node_values = None
         if args.large_model:
             const_node_values = compress_graph_def(graph_def)
+        if args.output_frozen_graph:
+            utils.save_protobuf(args.output_frozen_graph, graph_def)
         tf.import_graph_def(graph_def, name='')
     with tf_loader.tf_session(graph=tf_graph):
         g = process_tf_graph(tf_graph,