DEVOCEAN

본 예제는 Resnet50 모형을 MXNet으로 학습한 후, ONNX에서 Prediction하는 예제입니다.

주어진 코끼리를 ONNX에서 잘 예측 할 수 있을까요?

			## requirements
numpy==1.14.6
mxnet-cu90==1.3.1
Keras==2.2.4
tensorflow==1.3.0
onnx==1.3.0
onnxruntime==0.3.0

#####

MXNet to ONNX

			import mxnet as mx
import numpy as np
from mxnet.contrib import onnx as onnx_mxnet
import logging
logging.basicConfig(level=logging.INFO)

# Download pre-trained resnet model - json and params by running following code.
path='http://data.mxnet.io/models/imagenet/'
[mx.test_utils.download(path+'resnet/18-layers/resnet-18-0000.params'),
 mx.test_utils.download(path+'resnet/18-layers/resnet-18-symbol.json'),
 mx.test_utils.download(path+'synset.txt')]

# Downloaded input symbol and params files
sym = './resnet-18-symbol.json'
params = './resnet-18-0000.params'

# Standard Imagenet input - 3 channels, 224*224
input_shape = (1,3,224,224)

# Path of the output file
onnx_file = './mxnet_exported_resnet50.onnx'

# Invoke export model API. It returns path of the converted onnx model
converted_model_path = onnx_mxnet.export_model(sym, params, [input_shape], np.float32, onnx_file)

Inference using ONNX

			from keras.preprocessing import image
from keras.applications.resnet50 import preprocess_input, decode_predictions
# image preprocessing
img_path = 'elephant.jpg'   # make sure the image is in img_path
img_size = 224
img = image.load_img(img_path, target_size=(img_size, img_size))
x = image.img_to_array(img)
x = x.transpose(2,0,1) ## c, w, h
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)

x = x if isinstance(x, list) else [x]
feed = dict([(input.name, x[n]) for n, input in enumerate(sess.get_inputs())])

			import onnxruntime
sess = onnxruntime.InferenceSession(converted_model_path)
pred_onnx = sess.run(None, feed)

print('Predicted:', decode_predictions(pred_onnx[0], top=3)[0])

Predicted: [('n02504458', 'African_elephant', 0.6264173), ('n01871265', 'tusker', 0.22193906), ('n02504013', 'Indian_elephant', 0.14317688)]