Add helper methods to load saved model from external files in model maker.

PiperOrigin-RevId: 480444918

mediapipe/model_maker/python/core/utils/model_util.py

@@ -26,6 +26,7 @@ from typing import Any, Callable, Dict, List, Optional, Text, Tuple, Union
 import numpy as np
 import tensorflow as tf
 
+# resources dependency
 from mediapipe.model_maker.python.core.data import dataset
 from mediapipe.model_maker.python.core.utils import quantization
 
@@ -33,6 +34,31 @@ DEFAULT_SCALE, DEFAULT_ZERO_POINT = 0, 0
 ESTIMITED_STEPS_PER_EPOCH = 1000
 
 
+def load_keras_model(model_path: str,
+                     compile_on_load: bool = False) -> tf.keras.Model:
+  """Loads a tensorflow Keras model from file and returns the Keras model.
+
+  Args:
+    model_path: Relative path to a directory containing model data, such as
+      <parent_path>/saved_model/.
+    compile_on_load: Whether the model should be compiled while loading. If
+      False, the model returned has to be compiled with the appropriate loss
+      function and custom metrics before running for inference on a test
+      dataset.
+
+  Returns:
+    A tensorflow Keras model.
+  """
+  # Extract the file path before mediapipe/ as the `base_dir`. By joining it
+  # with the `model_path` which defines the relative path under mediapipe/, it
+  # yields to the aboslution path of the model files directory.
+  cwd = os.path.dirname(__file__)
+  base_dir = cwd[:cwd.rfind('mediapipe')]
+  absolute_path = os.path.join(base_dir, model_path)
+  return tf.keras.models.load_model(
+      absolute_path, custom_objects={'tf': tf}, compile=compile_on_load)
+
+
 def get_steps_per_epoch(steps_per_epoch: Optional[int] = None,
                         batch_size: Optional[int] = None,
                         train_data: Optional[dataset.Dataset] = None) -> int:

mediapipe/model_maker/python/core/utils/model_util_test.py

@@ -15,7 +15,6 @@
 import os
 
 from absl.testing import parameterized
-import numpy as np
 import tensorflow as tf
 
 from mediapipe.model_maker.python.core.utils import model_util
@@ -25,6 +24,18 @@ from mediapipe.model_maker.python.core.utils import test_util
 
 class ModelUtilTest(tf.test.TestCase, parameterized.TestCase):
 
+  def test_load_model(self):
+    input_dim = 4
+    model = test_util.build_model(input_shape=[input_dim], num_classes=2)
+    saved_model_path = os.path.join(self.get_temp_dir(), 'saved_model')
+    model.save(saved_model_path)
+    loaded_model = model_util.load_keras_model(saved_model_path)
+
+    input_tensors = test_util.create_random_sample(size=[1, input_dim])
+    model_output = model.predict_on_batch(input_tensors)
+    loaded_model_output = loaded_model.predict_on_batch(input_tensors)
+    self.assertTrue((model_output == loaded_model_output).all())
+
   @parameterized.named_parameters(
       dict(
           testcase_name='input_only_steps_per_epoch',
@@ -124,9 +135,9 @@ class ModelUtilTest(tf.test.TestCase, parameterized.TestCase):
                    input_dim: int,
                    max_input_value: int = 1000,
                    atol: float = 1e-04):
-    np.random.seed(0)
-    random_input = np.random.uniform(
-        low=0, high=max_input_value, size=(1, input_dim)).astype(np.float32)
+    random_input = test_util.create_random_sample(
+        size=[1, input_dim], high=max_input_value)
+    random_input = tf.convert_to_tensor(random_input)
 
     self.assertTrue(
         test_util.is_same_output(

mediapipe/model_maker/python/core/utils/test_util.py

@@ -46,6 +46,24 @@ def create_dataset(data_size: int,
   return dataset
 
 
+def create_random_sample(size: Union[int, List[int]],
+                         low: float = 0,
+                         high: float = 1) -> np.ndarray:
+  """Creates and returns a random sample with floating point values.
+
+  Args:
+    size: Size of the output multi-dimensional array.
+    low: Lower boundary of the output values.
+    high: Higher boundary of the output values.
+
+  Returns:
+    1D array if the size is scalar. Otherwise, N-D array whose dimension equals
+    input size.
+  """
+  np.random.seed(0)
+  return np.random.uniform(low=low, high=high, size=size).astype(np.float32)
+
+
 def build_model(input_shape: List[int], num_classes: int) -> tf.keras.Model:
   """Builds a simple Keras model for test."""
   inputs = tf.keras.layers.Input(shape=input_shape)