Merge pull request #23 from cms-ml/fix_custom_signature_bug

Fix model preparation for custom signatures (#22)
cms-ml · Oct 29, 2024 · c99b79d · c99b79d
2 parents 439a2e6 + 6b49b8e
commit c99b79d
Show file tree

Hide file tree

Showing 2 changed files with 58 additions and 45 deletions.
diff --git a/cmsml/scripts/compile_tf_graph.py b/cmsml/scripts/compile_tf_graph.py
@@ -71,7 +71,7 @@ def compile_tf_graph(
     for bs in sorted(set(map(int, batch_sizes))):
         # create a fully defined signature, filling leading None's in shapes with the batch size
         specs = {}
-        for key, spec in model.signatures["serving_default"].structured_input_signature[1].items():
+        for key, spec in model.signatures[input_serving_key].structured_input_signature[1].items():
             # ignore inputs without undefined axes
             if None not in spec.shape:
                 continue

diff --git a/tests/test_compile_tf_graph.py b/tests/test_compile_tf_graph.py
@@ -58,59 +58,72 @@ def test_compile_tf_graph_static_preparation(self):
         model = self.create_test_model(tf)
 
         with tmp_dir(create=False) as model_path, tmp_dir(create=False) as static_saved_model_path:
-            tf.saved_model.save(model, model_path)
+            spec = [
+                tf.TensorSpec(shape=(None, 2), dtype=tf.float32, name="inputs"),
+                tf.TensorSpec(shape=(None, 3), dtype=tf.float32, name="inputs_1"),
+                tf.TensorSpec(shape=(None, 10), dtype=tf.float32, name="inputs_2"),
+            ]
+
+            conc_func = tf.function(model.call).get_concrete_function(spec)
+            signatures = {
+                tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY: conc_func,
+                "custom_signature": conc_func,
+            }
 
-            # throw error if compilation happens with illegal batch size
-            with self.assertRaises(ValueError):
+            tf.saved_model.save(model, model_path, signatures=signatures)
+
+            for signature in signatures:
+                # throw error if compilation happens with illegal batch size
+                with self.assertRaises(ValueError):
+                    compile_tf_graph(
+                        model_path=model_path,
+                        output_path=static_saved_model_path,
+                        batch_sizes=[-1],
+                        input_serving_key=signature,
+                        output_serving_key=None,
+                        compile_prefix=None,
+                        compile_class=None,
+                    )
+
+                batch_sizes = [1, 2]
                 compile_tf_graph(
                     model_path=model_path,
                     output_path=static_saved_model_path,
-                    batch_sizes=[-1],
-                    input_serving_key=tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY,
+                    batch_sizes=batch_sizes,
+                    input_serving_key=signature,
                     output_serving_key=None,
                     compile_prefix=None,
                     compile_class=None,
                 )
 
-            batch_sizes = [1, 2]
-            compile_tf_graph(
-                model_path=model_path,
-                output_path=static_saved_model_path,
-                batch_sizes=batch_sizes,
-                input_serving_key=tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY,
-                output_serving_key=None,
-                compile_prefix=None,
-                compile_class=None,
-            )
-
-            # load model
-            loaded_static_model = cmsml.tensorflow.load_model(static_saved_model_path)
-
-        # check input shape
-        for batch_size in batch_sizes:
-            # first entry is empty, second contains inputs tuple(tensorspecs)
-            key = f"serving_default_bs{batch_size}"
-            model_static_inputs = loaded_static_model.signatures[key].structured_input_signature[1]
-
-            expected_model_static_inputs = {
-                f"inputs_bs{batch_size}": tf.TensorSpec(
-                    shape=(batch_size, 2),
-                    dtype=tf.float32,
-                    name=f"inputs_bs{batch_size}",
-                ),
-                f"inputs_1_bs{batch_size}": tf.TensorSpec(
-                    shape=(batch_size, 3),
-                    dtype=tf.float32,
-                    name=f"inputs_1_bs{batch_size}",
-                ),
-                f"inputs_2_bs{batch_size}": tf.TensorSpec(
-                    shape=(batch_size, 10),
-                    dtype=tf.float32,
-                    name=f"inputs_2_bs{batch_size}",
-                ),
-            }
-
-            self.assertDictEqual(model_static_inputs, expected_model_static_inputs)
+                # load model
+                loaded_static_model = cmsml.tensorflow.load_model(static_saved_model_path)
+
+                # check input shape
+                for batch_size in batch_sizes:
+                    # first entry is empty, second contains inputs tuple(tensorspecs)
+                    key = f"{signature}_bs{batch_size}"
+                    model_static_inputs = loaded_static_model.signatures[key].structured_input_signature[1]
+
+                    expected_model_static_inputs = {
+                        f"inputs_bs{batch_size}": tf.TensorSpec(
+                            shape=(batch_size, 2),
+                            dtype=tf.float32,
+                            name=f"inputs_bs{batch_size}",
+                        ),
+                        f"inputs_1_bs{batch_size}": tf.TensorSpec(
+                            shape=(batch_size, 3),
+                            dtype=tf.float32,
+                            name=f"inputs_1_bs{batch_size}",
+                        ),
+                        f"inputs_2_bs{batch_size}": tf.TensorSpec(
+                            shape=(batch_size, 10),
+                            dtype=tf.float32,
+                            name=f"inputs_2_bs{batch_size}",
+                        ),
+                    }
+
+                    self.assertDictEqual(model_static_inputs, expected_model_static_inputs)
 
     def test_compile_tf_graph_static_aot_compilation(self):
         from cmsml.scripts.compile_tf_graph import compile_tf_graph