Predict an ONNX InferenceService

In [1]:

    

%%bash
echo "model name: $MODEL_NAME"
echo "service url: $SERVICE_HOSTNAME"
echo "cluster ip: $CLUSTER_IP"


    

    




model name: style-sample
service url: style-sample.default.example.com
cluster ip: 52.247.193.219

In [4]:

    

# if you dont see values above, see instructions in the readme and replace the values here!
import os
model_name = os.environ["MODEL_NAME"]
service_hostname = os.environ["SERVICE_HOSTNAME"]
cluster_ip = os.environ["CLUSTER_IP"]
predictor_url = "http://%s/v1/models/%s:predict" % (cluster_ip, model_name)
predictor_url


    

    
Out[4]:





'http://52.247.193.219/v1/models/style-sample:predict'

In [5]:

    

from PIL import Image
import numpy as np
import google.protobuf.json_format as json_format
import assets.onnx_ml_pb2 as onnx_ml_pb2
import assets.predict_pb2 as predict_pb2
import requests


    

In [6]:

    

# load & resize image
image = Image.open("image.jpg")
image = image.resize((244,244), Image.ANTIALIAS)
image


    

    
Out[6]:

In [7]:

    

# preprocess image data
norm_img_data = np.array(image).astype('float32')
norm_img_data = np.transpose(norm_img_data, [2, 0, 1])
norm_img_data = np.expand_dims(norm_img_data, axis=0)
np.shape(norm_img_data)


    

    
Out[7]:





(1, 3, 244, 244)

In [8]:

    

# Create request message to be sent to the predictor
import assets.onnx_ml_pb2 as onnx_ml_pb2
import assets.predict_pb2 as predict_pb2
import requests

input_tensor = onnx_ml_pb2.TensorProto()
input_tensor.dims.extend(norm_img_data.shape)
input_tensor.data_type = 1
input_tensor.raw_data = norm_img_data.tobytes()

request_message = predict_pb2.PredictRequest()
request_message.inputs["input1"].data_type = input_tensor.data_type
request_message.inputs["input1"].dims.extend(input_tensor.dims)
request_message.inputs["input1"].raw_data = input_tensor.raw_data

# write message data to JSON
message_data = json_format.MessageToJson(request_message)


    

In [9]:

    

# Call predictor
request_headers = {'Content-Type': 'application/json', 'Accept': 'application/json', 'Host': service_hostname}
response = requests.post(predictor_url, headers=request_headers, data=message_data)


    

In [10]:

    

# Parse response message
response_message = json_format.Parse(response.text, predict_pb2.PredictResponse())
output1 = np.frombuffer(response_message.outputs['output1'].raw_data, dtype=np.float32)
output1 = output1.reshape(3,244,244)


    

In [11]:

    

# postprocess
result = np.clip(output1, 0, 255)
result = result.transpose(1,2,0).astype("uint8")
img = Image.fromarray(result)
img


    

    
Out[11]: