Merge pull request #430 from dtrawins/openvino_demo

updated ensemble pipeline with OpenVINO component

examples/models/openvino_imagenet_ensemble/input_images.txt

@@ -0,0 +1,3 @@
+dog.jpeg 248
+zebra.jpeg 340
+pelican.jpeg 144

examples/models/openvino_imagenet_ensemble/resources/combiner/ImageNetCombiner.py

@@ -1,10 +1,11 @@
 import logging
+import numpy as np
 logger = logging.getLogger(__name__)
 
 class ImageNetCombiner(object):
 
     def aggregate(self, Xs, features_names):
         print("ImageNet Combiner aggregate called")
-        logger.info(Xs)
-        return (Xs[0]+Xs[1])/2.0
+        logger.debug(Xs)
+        return (np.reshape(Xs[0],(1,-1)) + np.reshape(Xs[1], (1,-1)))/2.0
 

examples/models/openvino_imagenet_ensemble/resources/combiner/Makefile

@@ -1,4 +1,4 @@
 
 build:
-	s2i build -E environment_grpc . seldonio/seldon-core-s2i-python36:0.4-SNAPSHOT seldonio/imagenet_combiner:0.1
+	s2i build -E environment_grpc . seldon_openvino_base:latest seldonio/imagenet_combiner:0.1
 

examples/models/openvino_imagenet_ensemble/resources/combiner/README.md

@@ -0,0 +1,8 @@
+# Combiner components for two models' results
+
+
+### Building
+```bash
+s2i build -E environment_grpc . seldon_openvino_base:latest seldonio/imagenet_combiner:0.1
+
+```

examples/models/openvino_imagenet_ensemble/resources/model/Prediction.py

@@ -0,0 +1,100 @@
+import numpy as np
+import logging
+import datetime
+import os
+import sys
+from urllib.parse import urlparse
+from google.cloud import storage
+from openvino.inference_engine import IENetwork, IEPlugin
+
+
+def get_logger(name):
+    logger = logging.getLogger(name)
+    log_formatter = logging.Formatter("%(asctime)s - %(name)s - "
+                                      "%(levelname)s - %(message)s")
+    logger.setLevel('DEBUG')
+
+    console_handler = logging.StreamHandler()
+    console_handler.setFormatter(log_formatter)
+    logger.addHandler(console_handler)
+
+    return logger
+
+logger = get_logger(__name__)
+
+
+def gs_download_file(path):
+    if path is None:
+        return None
+    parsed_path = urlparse(path)
+    bucket_name = parsed_path.netloc
+    file_path = parsed_path.path[1:]
+    gs_client = storage.Client()
+    bucket = gs_client.get_bucket(bucket_name)
+    blob = bucket.blob(file_path)
+    tmp_path = os.path.join('/tmp', file_path.split(os.sep)[-1])
+    blob.download_to_filename(tmp_path)
+    return tmp_path
+
+
+def s3_download_file(path):
+    if path is None:
+        return None
+    s3_endpoint = os.getenv('S3_ENDPOINT')
+    s3_client = boto3.client('s3', endpoint_url=s3_endpoint)
+    parsed_path = urlparse(path)
+    bucket_name = parsed_path.netloc
+    file_path = parsed_path.path[1:]
+    tmp_path = os.path.join('/tmp', file_path.split(os.sep)[-1])
+    s3_transfer = boto3.s3.transfer.S3Transfer(s3_client)
+    s3_transfer.download_file(bucket_name, file_path, tmp_path)
+    return tmp_path
+
+
+def GetLocalPath(requested_path):
+    parsed_path = urlparse(requested_path)
+    if parsed_path.scheme == '':
+        return requested_path
+    elif parsed_path.scheme == 'gs':
+        return gs_download_file(path=requested_path)
+    elif parsed_path.scheme == 's3':
+        return s3_download_file(path=requested_path)
+
+
+class Prediction(object):
+    def __init__(self):
+        try:
+            xml_path = os.environ["XML_PATH"]
+            bin_path = os.environ["BIN_PATH"]
+
+        except KeyError:
+            print("Please set the environment variables XML_PATH, BIN_PATH")
+            sys.exit(1)
+
+        xml_local_path = GetLocalPath(xml_path)
+        bin_local_path = GetLocalPath(bin_path)
+        print('path object', xml_local_path)
+
+        CPU_EXTENSION = os.getenv('CPU_EXTENSION', "/usr/local/lib/libcpu_extension.so")
+
+        plugin = IEPlugin(device='CPU', plugin_dirs=None)
+        if CPU_EXTENSION:
+            plugin.add_cpu_extension(CPU_EXTENSION)
+        net = IENetwork(model=xml_local_path, weights=bin_local_path)
+        self.input_blob = next(iter(net.inputs))
+        self.out_blob = next(iter(net.outputs))
+        self.batch_size = net.inputs[self.input_blob].shape[0]
+        self.inputs = net.inputs
+        self.outputs = net.outputs
+        self.exec_net = plugin.load(network=net, num_requests=self.batch_size)
+
+
+    def predict(self,X,feature_names):
+        start_time = datetime.datetime.now()
+        results = self.exec_net.infer(inputs={self.input_blob: X})
+        predictions = results[self.out_blob]
+        end_time = datetime.datetime.now()
+        duration = (end_time - start_time).total_seconds() * 1000
+        logger.debug("Processing time: {:.2f} ms".format(duration))
+        return predictions.astype(np.float64)
+

examples/models/openvino_imagenet_ensemble/resources/model/README.md

@@ -0,0 +1,41 @@
+# OpenVINO prediction component
+
+Model configuration is implemented using environment variables:
+
+`XML_PATH`  - s3, gs or local path to xml file in OpenVINO model server
+
+`BIN_PATH` - s3, gs or local path to bin file in OpenVINO model server
+
+When using GCS make sure you added also GOOGLE_APPLICATION_CREDENTIALS variable and mounted correspondent token file.
+
+In case is S3 or Minio storage add appropriate environment variables with the credentials.
+
+Component is executing inference operation. Processing time is included in the components debug logs.
+
+Model input and output tensors are determined automatically. There is assumed only one input tensor and output tensor.
+
+### Building example:
+
+```bash
+s2i build -E environment_grpc . seldon_openvino_base:latest seldon-openvino-prediction:0.1
+```
+The base image `seldon_openvino_base:latest` should be created according to this [procedure](../../../../../wrappers/s2i/python_openvino)
+
+
+### Local testing example:
+
+```bash
+docker run -it -v $GOOGLE_APPLICATION_CREDENTIALS:/etc/gcp.json -e GOOGLE_APPLICATION_CREDENTIALS=/etc/gcp.json \
+ -e XML_PATH=gs://inference-eu/models_zoo/resnet_V1_50/resnet_V1_50.xml \
+ -e BIN_PATH=gs://inference-eu/models_zoo/resnet_V1_50/resnet_V1_50.bin  
+starting microservice
+2019-02-05 11:13:32,045 - seldon_core.microservice:main:261 - INFO:  Starting microservice.py:main
+2019-02-05 11:13:32,047 - seldon_core.microservice:main:292 - INFO:  Annotations: {}
+path object /tmp/resnet_V1_50.xml
+  net = IENetwork(model=xml_local_path, weights=bin_local_path)
+2019-02-05 11:14:19,870 - seldon_core.microservice:main:354 - INFO:  Starting servers
+2019-02-05 11:14:19,906 - seldon_core.microservice:grpc_prediction_server:333 - INFO:  GRPC microservice Running on port 5000
+```
+
+
+

examples/models/openvino_imagenet_ensemble/resources/model/environment_grpc

@@ -0,0 +1,5 @@
+MODEL_NAME=Prediction
+API_TYPE=GRPC
+SERVICE_TYPE=MODEL
+PERSISTENCE=0
+

examples/models/openvino_imagenet_ensemble/resources/model/requirements.txt

@@ -0,0 +1,2 @@
+google-cloud-storage==1.13.0
+boto3==1.9.34

examples/models/openvino_imagenet_ensemble/resources/transformer/ImageNetTransformer.py

@@ -1,11 +1,10 @@
 import numpy as np
-from keras.applications.imagenet_utils import preprocess_input, decode_predictions
-from keras.preprocessing import image
 from seldon_core.proto import prediction_pb2
 import tensorflow as tf
 import logging
-import sys
-import io
+import datetime
+import cv2
+import os
 
 logger = logging.getLogger(__name__)
 
@@ -14,26 +13,48 @@ def __init__(self, metrics_ok=True):
         print("Init called")
         f = open('imagenet_classes.json')
         self.cnames = eval(f.read())
-
+        self.size = os.getenv('SIZE', 224)
+        self.dtype = os.getenv('DTYPE', 'float')
+        self.classes = os.getenv('CLASSES', 1000)
+
+    def crop_resize(self, img,cropx,cropy):
+        y,x,c = img.shape
+        if y < cropy:
+            img = cv2.resize(img, (x, cropy))
+            y = cropy
+        if x < cropx:
+            img = cv2.resize(img, (cropx,y))
+            x = cropx
+        startx = x//2-(cropx//2)
+        starty = y//2-(cropy//2)
+        return img[starty:starty+cropy,startx:startx+cropx,:]
+
     def transform_input_grpc(self, request):
-        logger.debug("Transform called")
-        b = io.BytesIO(request.binData)
-        img = image.load_img(b, target_size=(227, 227))
-        X = image.img_to_array(img)
-        X = np.expand_dims(X, axis=0)
-        X = preprocess_input(X)
-        X = X.transpose((0,3,1,2))
+        logger.info("Transform called")
+        start_time = datetime.datetime.now()
+        X = np.frombuffer(request.binData, dtype=np.uint8)
+        X = cv2.imdecode(X, cv2.IMREAD_COLOR)  # BGR format
+        X = self.crop_resize(X, self.size, self.size)
+        X = X.astype(self.dtype)
+        X = X.transpose(2,0,1).reshape(1,3,self.size,self.size)
+        logger.info("Shape: %s; Dtype: %s; Min: %s; Max: %s",X.shape,X.dtype,np.amin(X),np.amax(X))
+        jpeg_time = datetime.datetime.now()
+        jpeg_duration = (jpeg_time - start_time).total_seconds() * 1000
+        logger.info("jpeg preprocessing: %s ms", jpeg_duration)
         datadef = prediction_pb2.DefaultData(
             names = 'x',
             tftensor = tf.make_tensor_proto(X)
         )
+        end_time = datetime.datetime.now()
+        duration = (end_time - start_time).total_seconds() * 1000
+        logger.info("Total transformation: %s ms", duration)
         request = prediction_pb2.SeldonMessage(data = datadef)
         return request
 
     def transform_output_grpc(self, request):
-        logger.debug("Transform output called")
+        logger.info("Transform output called")
         result = tf.make_ndarray(request.data.tftensor)
-        result = result.reshape(1,1000)
+        result = result.reshape(1,self.classes)
 
         single_result = result[[0],...]
         ma = np.argmax(single_result)

examples/models/openvino_imagenet_ensemble/resources/transformer/Makefile

@@ -1,3 +1,3 @@
 
 build:
-	s2i build -E environment_grpc . seldonio/seldon-core-s2i-python36:0.5-SNAPSHOT seldonio/imagenet_transformer:0.1
+	s2i build -E environment_grpc . seldon_openvino_base:latest seldonio/imagenet_transformer:0.1

examples/models/openvino_imagenet_ensemble/resources/transformer/README.md

@@ -0,0 +1,24 @@
+## Transformer component
+
+Exemplary implementation of data transformation tasks.
+
+Input transformation function accepts as input the binary representation of jpeg content.
+It performs the following operations:
+- convert compressed jpeg content to numpy array (BGR format)
+- crop/resize the image to the square shape set in the environment variable `SIZE` (by default 224)
+- transpose the data to NCWH
+
+
+Output transformation function is consuming the imagenet classification models.
+It is converting the array including probability for each imagenet classes into the class name.
+It is returning 'human readable' string with most likely class name.
+The function is using `CLASSES` environment variable to define the expected number of classes in the output. 
+Depending on the model it could be 1000 (the default value) or 1001.
+
+
+### Building example:
+```bash
+s2i build -E environment_grpc . seldon_openvino_base:latest seldonio/imagenet_transformer:0.1
+```
+
+The base image `seldon_openvino_base:latest` should be created according to this [procedure](../../../../../wrappers/s2i/python_openvino)

examples/models/openvino_imagenet_ensemble/resources/transformer/environment_grpc

@@ -2,3 +2,4 @@ MODEL_NAME=ImageNetTransformer
 API_TYPE=GRPC
 SERVICE_TYPE=TRANSFORMER
 PERSISTENCE=0
+

examples/models/openvino_imagenet_ensemble/resources/transformer/requirements.txt

@@ -1,3 +0,0 @@
-numpy>=1.8.2
-keras
-pillow