Add ResNet50 to DeepImageFeaturizer (#57)

* resnet50 works

* clean up

* clearer comments

Author: sueann

python/sparkdl/transformers/keras_applications.py

@@ -16,7 +16,8 @@
 from abc import ABCMeta, abstractmethod
 
 import keras.backend as K
-from keras.applications import inception_v3, xception
+from keras.applications import inception_v3, xception, resnet50
+import numpy as np
 import tensorflow as tf
 
 from sparkdl.transformers.utils import (imageInputPlaceholder, InceptionV3Constants)
@@ -104,9 +105,41 @@ def inputShape(self):
     def _testKerasModel(self, include_top):
         return xception.Xception(weights="imagenet", include_top=include_top)
 
+class ResNet50Model(KerasApplicationModel):
+    def preprocess(self, inputImage):
+        return _imagenet_preprocess_input(inputImage, self.inputShape())
+
+    def model(self, preprocessed, featurize):
+        return resnet50.ResNet50(input_tensor=preprocessed, weights="imagenet",
+                                 include_top=(not featurize))
+
+    def inputShape(self):
+        return (224, 224)
+
+    def _testKerasModel(self, include_top):
+        return resnet50.ResNet50(weights="imagenet", include_top=include_top)
+
+def _imagenet_preprocess_input(x, input_shape):
+    """
+    For ResNet50, VGG models. For InceptionV3 and Xception it's okay to use the
+    keras version (e.g. InceptionV3.preprocess_input) as the code path they hit
+    works okay with tf.Tensor inputs. The following was translated to tf ops from
+    https://github.com/fchollet/keras/blob/fb4a0849cf4dc2965af86510f02ec46abab1a6a4/keras/applications/imagenet_utils.py#L52
+    It's a possibility to change the implementation in keras to look like the
+    following, but not doing it for now.
+    """
+    # 'RGB'->'BGR'
+    x = x[..., ::-1]
+    # Zero-center by mean pixel
+    mean = np.ones(input_shape + (3,), dtype=np.float32)
+    mean[..., 0] = 103.939
+    mean[..., 1] = 116.779
+    mean[..., 2] = 123.68
+    return x - mean
 
 KERAS_APPLICATION_MODELS = {
     "InceptionV3": InceptionV3Model,
-    "Xception": XceptionModel
+    "Xception": XceptionModel,
+    "ResNet50": ResNet50Model,
 }
 

python/sparkdl/transformers/named_image.py

@@ -29,7 +29,7 @@
 from sparkdl.transformers.tf_image import TFImageTransformer
 
 
-SUPPORTED_MODELS = ["InceptionV3", "Xception"]
+SUPPORTED_MODELS = ["InceptionV3", "Xception", "ResNet50"]
 
 
 class DeepImagePredictor(Transformer, HasInputCol, HasOutputCol):

python/tests/transformers/named_image_ResNet50_test.py

@@ -0,0 +1,21 @@
+# Copyright 2017 Databricks, Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+from .named_image_test import NamedImageTransformerBaseTestCase
+
+class NamedImageTransformerResNet50Test(NamedImageTransformerBaseTestCase):
+
+    __test__ = True
+    name = "ResNet50"

python/tests/transformers/named_image_test.py

@@ -14,6 +14,7 @@
 #
 
 import numpy as np
+from keras.applications import resnet50
 import tensorflow as tf
 
 from pyspark.ml import Pipeline
@@ -29,10 +30,26 @@
 from .image_utils import getSampleImageDF, getSampleImageList
 
 
-class GetKerasApplicationModelTestCase(SparkDLTestCase):
+class KerasApplicationModelTestCase(SparkDLTestCase):
     def test_getKerasApplicationModelError(self):
         self.assertRaises(ValueError, keras_apps.getKerasApplicationModel, "NotAModelABC")
 
+    def test_imagenet_preprocess_input(self):
+        # compare our tf implementation to the np implementation in keras
+        image = np.zeros((256, 256, 3))
+
+        sess = tf.Session()
+        with sess.as_default():
+            x = tf.placeholder(tf.float32, shape=[256, 256, 3])
+            processed = keras_apps._imagenet_preprocess_input(x, (256, 256)),
+            sparkdl_preprocessed_input = sess.run(processed, {x: image})
+
+        keras_preprocessed_input = resnet50.preprocess_input(np.expand_dims(image, axis=0))
+
+        # NOTE: precision errors occur for decimal > 5
+        np.testing.assert_array_almost_equal(sparkdl_preprocessed_input, keras_preprocessed_input,
+                                             decimal=5)
+
 
 class NamedImageTransformerBaseTestCase(SparkDLTestCase):
     """