Add --outputs_as_nchw option to transpose output to from nhwc to nchw (#1979)

* add output_as_nchw

Signed-off-by: Deyu Huang <[email protected]>

* fix node replace logic

Signed-off-by: Deyu Huang <[email protected]>

* add tests for outputs as nchw

Signed-off-by: Deyu Huang <[email protected]>

* add it into function and doc

Signed-off-by: Deyu Huang <[email protected]>

* fix output_names_with_port range

Signed-off-by: Deyu Huang <[email protected]>

* fix the input_as_nchw description

Signed-off-by: Deyu Huang <[email protected]>

* change tests name

Signed-off-by: Deyu Huang <[email protected]>

Author: hwangdeyu

README.md

@@ -292,8 +292,8 @@ import tf2onnx
 model_proto, external_tensor_storage = tf2onnx.convert.from_keras(model,
                 input_signature=None, opset=None, custom_ops=None,
                 custom_op_handlers=None, custom_rewriter=None,
-                inputs_as_nchw=None, extra_opset=None shape_override=None,
-                target=None, large_model=False, output_path=None)
+                inputs_as_nchw=None, outputs_as_nchw=None, extra_opset=None,
+                shape_override=None, target=None, large_model=False, output_path=None)
 
     Args:
         model: the tf.keras model we want to convert
@@ -307,7 +307,8 @@ model_proto, external_tensor_storage = tf2onnx.convert.from_keras(model,
         custom_rewriter: list of custom graph rewriters
         extra_opset: list of extra opset's, for example the opset's used by custom ops
         shape_override: dict with inputs that override the shapes given by tensorflow
-        inputs_as_nchw: transpose inputs in list from nchw to nhwc
+        inputs_as_nchw: transpose inputs in list from nhwc to nchw
+        outputs_as_nchw: transpose outputs in list from nhwc to nchw
         large_model: use the ONNX external tensor storage format
         output_path: save model to output_path
 
@@ -323,8 +324,8 @@ import tf2onnx
 
 model_proto, external_tensor_storage = tf2onnx.convert.from_function(function,
                 input_signature=None, opset=None, custom_ops=None,
-                custom_op_handlers=None, custom_rewriter=None,
-                inputs_as_nchw=None, extra_opset=None, shape_override=None,
+                custom_op_handlers=None, custom_rewriter=None, inputs_as_nchw=None,
+                outputs_as_nchw=None, extra_opset=None, shape_override=None,
                 target=None, large_model=False, output_path=None)
 
     Args:
@@ -339,7 +340,8 @@ model_proto, external_tensor_storage = tf2onnx.convert.from_function(function,
         custom_rewriter: list of custom graph rewriters
         extra_opset: list of extra opset's, for example the opset's used by custom ops
         shape_override: dict with inputs that override the shapes given by tensorflow
-        inputs_as_nchw: transpose inputs in list from nchw to nhwc
+        inputs_as_nchw: transpose inputs in list from nhwc to nchw
+        outputs_as_nchw: transpose outputs in list from nhwc to nchw
         large_model: use the ONNX external tensor storage format
         output_path: save model to output_path
 
@@ -354,7 +356,7 @@ import tf2onnx
 model_proto, external_tensor_storage = tf2onnx.convert.from_graph_def(graph_def,
                 name=None, input_names=None, output_names=None, opset=None,
                 custom_ops=None, custom_op_handlers=None, custom_rewriter=None, 
-                inputs_as_nchw=None, extra_opset=None,
+                inputs_as_nchw=None, outputs_as_nchw=None, extra_opset=None,
                 shape_override=None, target=None, large_model=False,
                 output_path=None)
 
@@ -369,7 +371,8 @@ model_proto, external_tensor_storage = tf2onnx.convert.from_graph_def(graph_def,
         custom_rewriter: list of custom graph rewriters
         extra_opset: list of extra opset's, for example the opset's used by custom ops
         shape_override: dict with inputs that override the shapes given by tensorflow
-        inputs_as_nchw: transpose inputs in list from nchw to nhwc
+        inputs_as_nchw: transpose inputs in list from nhwc to nchw
+        outputs_as_nchw: transpose outputs in list from nhwc to nchw
         large_model: use the ONNX external tensor storage format
         output_path: save model to output_path
 
@@ -383,8 +386,8 @@ import tf2onnx
 
 model_proto, external_tensor_storage = tf2onnx.convert.from_tflite(tflite_path,
                 input_names=None, output_names=None, opset=None, custom_ops=None, custom_op_handlers=None,
-                custom_rewriter=None, inputs_as_nchw=None, extra_opset=None, shape_override=None, target=None,
-                large_model=False, output_path=None):
+                custom_rewriter=None, inputs_as_nchw=None, outputs_as_nchw=None, extra_opset=None,
+                shape_override=None, target=None, large_model=False, output_path=None):
 
     Args:
         tflite_path: the tflite model file full path
@@ -396,7 +399,8 @@ model_proto, external_tensor_storage = tf2onnx.convert.from_tflite(tflite_path,
             runtime can still open the model. Type is a dictionary `{op name: domain}`.
         custom_op_handlers: dictionary of custom ops handlers
         custom_rewriter: list of custom graph rewriters
-        inputs_as_nchw: transpose inputs in list from nchw to nhwc
+        inputs_as_nchw: transpose inputs in list from nhwc to nchw
+        outputs_as_nchw: transpose outputs in list from nhwc to nchw
         extra_opset: list of extra opset's, for example the opset's used by custom ops
         shape_override: dict with inputs that override the shapes given by tensorflow
         target: list of workarounds applied to help certain platforms

tests/backend_test_base.py

@@ -20,6 +20,7 @@
 import onnx
 from common import get_test_config
 from tfjs_runner import run_tfjs
+from tf2onnx import constants
 from tf2onnx import utils
 from tf2onnx.tfonnx import process_tf_graph
 from tf2onnx import optimizer
@@ -366,6 +367,7 @@ def run_test_case(self, func, feed_dict, input_names_with_port, output_names_wit
         graph_def_path = os.path.join(self.test_data_directory, self._testMethodName + "_after_tf_optimize.pb")
         utils.save_protobuf(graph_def_path, graph_def)
         self.logger.debug("created file  %s", graph_def_path)
+        tfl_process_args = process_args.copy()
 
         if test_tfjs:
             tfjs_path = self.convert_to_tfjs(graph_def_path, output_names_with_port)
@@ -395,6 +397,10 @@ def run_test_case(self, func, feed_dict, input_names_with_port, output_names_wit
                     g = optimizer.optimize_graph(g, catch_errors=False)
                 actual = self.run_backend(g, output_names_with_port, onnx_feed_dict, large_model,
                                           use_custom_ops=use_custom_ops)
+            if 'outputs_as_nchw' in tfl_process_args:
+                for output_name in tfl_process_args['outputs_as_nchw']:
+                    i = output_names_with_port.index(output_name)
+                    actual[i] = np.transpose(actual[i], constants.NCHW_TO_NHWC)
 
             self.assert_results_equal(expected, actual, rtol, atol, mtol, check_value, check_shape, check_dtype)
             self.assert_shapes_correct(g, self.config.allow_missing_shapes, not self.config.skip_onnx_checker)
@@ -410,12 +416,14 @@ def run_test_case(self, func, feed_dict, input_names_with_port, output_names_wit
             if run_tfl_consistency_test:
                 self.assert_results_equal(expected, tfl_res, rtol, atol, mtol, check_value, check_shape, check_dtype)
 
-            tfl_process_args = process_args.copy()
             if 'inputs_as_nchw' in tfl_process_args:
                 nchw_inps_with_port = tfl_process_args['inputs_as_nchw']
                 tfl_process_args['inputs_as_nchw'] = [i.split(':')[0] for i in nchw_inps_with_port]
             input_names_without_port = [inp.split(':')[0] for inp in feed_dict.keys()]
-
+            if 'outputs_as_nchw' in tfl_process_args:
+                nchw_outps_with_port = tfl_process_args['outputs_as_nchw']
+                tfl_process_args['outputs_as_nchw'] = [i.split(':')[0] for i in nchw_outps_with_port]
+                output_names_with_port = [i.split(':')[0] for i in nchw_outps_with_port]
             g = process_tf_graph(None, opset=self.config.opset,
                                  input_names=input_names_without_port,
                                  output_names=tfl_outputs,
@@ -427,6 +435,10 @@ def run_test_case(self, func, feed_dict, input_names_with_port, output_names_wit
             onnx_feed_dict_without_port = {k.split(':')[0]: v for k, v in onnx_feed_dict.items()}
             onnx_tfl_res = self.run_backend(g, tfl_outputs, onnx_feed_dict_without_port,
                                             postfix="_from_tflite", use_custom_ops=use_custom_ops)
+            if 'outputs_as_nchw' in tfl_process_args:
+                for output_name in tfl_process_args['outputs_as_nchw']:
+                    i = output_names_with_port.index(output_name)
+                    onnx_tfl_res[i] = np.transpose(onnx_tfl_res[i], constants.NCHW_TO_NHWC)
 
             self.assert_results_equal(tfl_res, onnx_tfl_res, rtol, atol, mtol, check_value, check_shape, check_dtype)
             self.assert_shapes_correct(g, self.config.allow_missing_shapes, not self.config.skip_onnx_checker)
@@ -456,6 +468,10 @@ def run_test_case(self, func, feed_dict, input_names_with_port, output_names_wit
             g = optimizer.optimize_graph(g)
             onnx_tfjs_res = self.run_backend(g, None, onnx_feed_dict, large_model,
                                              postfix="_from_tfjs", use_custom_ops=use_custom_ops)
+            if 'outputs_as_nchw' in tfl_process_args:
+                for output_name in tfl_process_args['outputs_as_nchw']:
+                    i = output_names_with_port.index(output_name)
+                    onnx_tfjs_res[i] = np.transpose(onnx_tfjs_res[i], constants.NCHW_TO_NHWC)
 
             self.assert_results_equal(tfjs_res, onnx_tfjs_res, rtol, atol, mtol, check_value, check_shape,
                                       check_dtype=False)

tests/test_backend.py

@@ -712,7 +712,7 @@ def func(x):
                            graph_validator=lambda g: (check_op_count(g, "RandomUniform", 0) and
                                                       check_op_count(g, "RandomUniformLike", 0)))
 
-    def test_conv2d_with_input_transpose(self):
+    def test_inputs_as_nchw_arg(self):
         x_shape = [2, 32, 32, 3]
         kernel_shape = [3, 3, 3, 3]
         x_val = make_xval(x_shape)
@@ -725,6 +725,17 @@ def func(x):
                             process_args={"inputs_as_nchw": [_INPUT]},
                             onnx_feed_dict={_INPUT: x_val_for_onnx})
 
+    def test_outputs_as_nchw_arg(self):
+        x_shape = [2, 32, 32, 3]
+        kernel_shape = [3, 3, 3, 3]
+        x_val = make_xval(x_shape)
+        def func(x):
+            kernel = tf.constant(make_xval(kernel_shape), dtype=tf.float32, name='kernel')
+            conv = tf.nn.conv2d(x, kernel, strides=[1, 1, 1, 1], padding="SAME")
+            return tf.identity(conv, name=_TFOUTPUT)
+        self._run_test_case(func, [_OUTPUT], {_INPUT: x_val}, rtol=1e-05,
+                            process_args={"outputs_as_nchw": [_OUTPUT]})
+
     @skip_tflite("TFlite adds ops that obscure pattern")
     @check_tf_min_version("1.15")
     def test_conv1d_dilations_rewriter(self):