[WIP]Use TF hub supported models to do model testing

diagnostics/model_test/tf_hub/tf_hub_model_test.py

@@ -0,0 +1,83 @@
+import tensorflow as tf
+import tensorflow_hub as hub
+import numpy as np
+import ngraph_config
+import argparse
+
+url_dict = {}
+url_dict[
+    "mobilenetv2"] = "https://tfhub.dev/google/imagenet/mobilenet_v2_140_224/classification/2"
+url_dict[
+    "inceptionv3"] = "https://tfhub.dev/google/imagenet/inception_v3/classification/1"
+url_dict[
+    "mobilenetv1"] = "https://tfhub.dev/google/imagenet/mobilenet_v1_100_224/classification/1"
+url_dict[
+    "inception_resentv2"] = "https://tfhub.dev/google/imagenet/inception_resnet_v2/classification/1"
+url_dict[
+    "resnetv2"] = "https://tfhub.dev/google/imagenet/resnet_v2_101/classification/1"
+url_dict[
+    "inceptionv1"] = "https://tfhub.dev/google/imagenet/inception_v1/classification/1"
+url_dict[
+    "pnasnet"] = "https://tfhub.dev/google/imagenet/pnasnet_large/classification/2"
+url_dict[
+    "nasnet"] = "https://tfhub.dev/google/imagenet/nasnet_large/classification/1"
+
+
+def compare(model_name, batch_size, threshold=0.001):
+    print("Running {} on Ngraph".format(model_name))
+    module_url = url_dict[model_name]
+    # Running on NGraph
+    with tf.Graph().as_default():
+        module = hub.Module(module_url)
+        height, width = hub.get_expected_image_size(module)
+        np.random.seed(100)
+        random_input = np.random.random_sample([batch_size, height, width, 3])
+        logits_ngraph = module(random_input)
+
+        with tf.Session() as sess:
+            sess.run(tf.global_variables_initializer())
+            sess.run(tf.tables_initializer())
+
+            ngraph_result = sess.run(logits_ngraph)
+
+    ngraph_config.disable()
+    if ngraph_config.is_enabled():
+        raise Exception("Ngraph should be disabled")
+    else:
+        print("Running {} on CPU".format(model_name))
+
+    with tf.Graph().as_default():
+        module_cpu = hub.Module(module_url)
+        logits_cpu = module_cpu(random_input)
+
+        with tf.Session() as sess:
+            sess.run(tf.global_variables_initializer())
+            sess.run(tf.tables_initializer())
+            cpu_result = sess.run(logits_cpu)
+
+    l1_norm = np.linalg.norm((ngraph_result - cpu_result), 1)
+    l2_norm = np.linalg.norm((ngraph_result - cpu_result), 2)
+    inf_norm = np.linalg.norm((ngraph_result - cpu_result), np.inf)
+    try: 
+        assert l1_norm < threshold
+        print ("l1 norm for model {} passed".format(model_name))
+    except AssertionError:
+        print("l1 norm {} is greater than the threshold {} set".format(l1_norm,threshold))
+    try:
+        assert l2_norm < threshold
+        print ("l2 norm for model {} passed".format(model_name))
+    except AssertionError:
+        print("l2 norm {0} is greater than the threshold {} set".format(l2_norm,threshold))
+    try:
+        assert inf_norm < threshold
+        print ("inf norm for model {} passed\n".format(model_name))
+    except AssertionError:
+        print("inf norm {} is greater than the threshold {} set\n".format(inf_norm,threshold))
+
+if __name__ == "__main__":
+    classification_model = ["mobilenetv2", "inceptionv3", "mobilenetv1","inception_resentv2","resnetv2","inceptionv1","pnasnet","nasnet"]
+    for model in classification_model:
+        compare(model, 100)
+
+
+