Skip to content

Latest commit

 

History

History
146 lines (96 loc) · 3.5 KB

README.md

File metadata and controls

146 lines (96 loc) · 3.5 KB
Raw

ONNX README

Usage

Currently we only implemented the IR -> ONNX part.

We'll take Keras DenseNet121 -> ONNX as an example.

Pre-request

Install TensorFlow

TensorFlow runs as the backend of both Keras and ONNX.

pip3 install tensorflow

See here for more infomation.

Install Keras

pip3 install keras

See here for more infomation.

Install ONNX frontend and backend

Here we use ONNX-TensorFlow to install ONNX and its TensorFlow backend.

pip install onnx-tf

Install latest MMdnn

As the IR -> ONNX part is not included in the lastest version of MMdnn, we need to install MMdnn master branch:

git clone https://github.com/Microsoft/MMdnn.git
pip3 install -e MMdnn/

Install some utilities

pip3 install pillow

Prepare Keras DenseNet121 model

The code below creates the DenseNet121 model (saved as densenet121.keras), and predicts the elephant picture (as below) with the model.

from keras.applications.densenet import DenseNet121
from keras.preprocessing import image
from keras.applications.densenet import preprocess_input, decode_predictions
import numpy as np

model = DenseNet121(include_top=True, weights='imagenet', input_tensor=None,
                    input_shape=None, pooling=None, classes=1000)

model.save('densenet121.keras')

img = image.load_img('elephant.jpg', target_size=(224, 224))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)

preds = model.predict(x)

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

Elephant

Here is the output:

Predicted: [('n02504458', 'African_elephant', 0.7482961), ('n01871265', 'tusker', 0.17631474), ('n02504013', 'Indian_elephant', 0.0753157)]

Convert DenseNet121 model from Keras to ONNX

Just run the one-step command below:

mmconvert -sf keras -iw densenet121.keras -df onnx -om densenet121.onnx

Here is the output:

IR network structure is saved as [77a4c18fc6254078ba4daca924eac3ab.json].
IR network structure is saved as [77a4c18fc6254078ba4daca924eac3ab.pb].
IR weights are saved as [77a4c18fc6254078ba4daca924eac3ab.npy].
Parse file [77a4c18fc6254078ba4daca924eac3ab.pb] with binary format successfully.
Target network code snippet is saved as [77a4c18fc6254078ba4daca924eac3ab.py].
ONNX model file is saved as [densenet121.onnx], generated by [77a4c18fc6254078ba4daca924eac3ab.py] and [77a4c18fc6254078ba4daca924eac3ab.npy].

Now you'll find the onnx model file densenet121.onnx in your current directory.

Load and run ONNX DenseNet121 model

import numpy as np
import onnx
from keras.preprocessing import image
from keras.applications.densenet import preprocess_input, decode_predictions
from onnx_tf.backend import prepare


model = onnx.load('densenet121.onnx')
tf_rep = prepare(model)

img = image.load_img('elephant.jpg', target_size=(224, 224))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)

predictions = tf_rep.run(x)
preds = predictions[0].reshape((1,1000))

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

Here is the output:

Predicted: [('n02504458', 'African_elephant', 0.7482961), ('n01871265', 'tusker', 0.17631474), ('n02504013', 'Indian_elephant', 0.0753157)]

Is that the same with Keras output?

Develop version

Ubuntu 16.04 with

  • Keras 2.1.6
  • onnx-tf 1.1.2
  • ONNX 1.2.1

@ 2018/06/09