Signed-off-by: mengniwa mengni.wang@intel.com

GoogleNet

Compared with the fp32 GoogleNet, int8 GoogleNet's Top-1 accuracy drop ratio is 0.07%, Top-5 accuracy drop ratio is 0.02% and performance improvement is 1.27x.

Note

The performance depends on the test hardware. Performance data here is collected with Intel® Xeon® Platinum 8280 Processor, 1s 4c per instance, CentOS Linux 8.3, data batch size is 1.

Description

GoogLeNet is the name of a convolutional neural network for classification,
which competed in the ImageNet Large Scale Visual Recognition Challenge in 2014.

Differences:

• not training with the relighting data-augmentation;
• not training with the scale or aspect-ratio data-augmentation;
• uses "xavier" to initialize the weights instead of "gaussian";

Dataset

ILSVRC2014

Source

Caffe BVLC GoogLeNet ==> Caffe2 GoogLeNet ==> ONNX GoogLeNet

Model input and output

Input

data_0: float[1, 3, 224, 224]

Output

prob_0: float[1, 1000]

Pre-processing steps

Necessary Imports

import imageio
from PIL import Image

Obtain and pre-process image

def get_image(path):
'''
    Using path to image, return the RGB load image
'''
    img = imageio.imread(path, pilmode='RGB')
    return img

# Pre-processing function for ImageNet models using numpy
def preprocess(img):
    '''
    Preprocessing required on the images for inference with mxnet gluon
    The function takes loaded image and returns processed tensor
    '''
    img = np.array(Image.fromarray(img).resize((224, 224))).astype(np.float32)
    img[:, :, 0] -= 123.68
    img[:, :, 1] -= 116.779
    img[:, :, 2] -= 103.939
    img[:,:,[0,1,2]] = img[:,:,[2,1,0]]
    img = img.transpose((2, 0, 1))
    img = np.expand_dims(img, axis=0)

    return img

Post-processing steps

def predict(path):
    # based on : https://mxnet.apache.org/versions/1.0.0/tutorials/python/predict_image.html
    img = get_image(path)
    img = preprocess(img)
    mod.forward(Batch([mx.nd.array(img)]))
    # Take softmax to generate probabilities
    prob = mod.get_outputs()[0].asnumpy()
    prob = np.squeeze(prob)
    a = np.argsort(prob)[::-1]
    return a

Sample test data

random generated sample test data:

• test_data_set_0
• test_data_set_1
• test_data_set_2
• test_data_set_3
• test_data_set_4
• test_data_set_5

Results/accuracy on test set

This bundled model obtains a top-1 accuracy 68.7% (31.3% error) and
a top-5 accuracy 88.9% (11.1% error) on the validation set, using
just the center crop. (Using the average of 10 crops,
(4 + 1 center) * 2 mirror, should obtain a bit higher accuracy.)

Quantization

GoogleNet-int8 and GoogleNet-qdq are obtained by quantizing fp32 GoogleNet model. We use Intel® Neural Compressor with onnxruntime backend to perform quantization. View the instructions to understand how to use Intel® Neural Compressor for quantization.

Environment

onnx: 1.9.0
onnxruntime: 1.8.0

Prepare model

wget https://github.com/onnx/models/raw/main/vision/classification/inception_and_googlenet/googlenet/model/googlenet-12.onnx

Model quantize

Make sure to specify the appropriate dataset path in the configuration file.

bash run_tuning.sh --input_model=path/to/model \  # model path as *.onnx
                   --config=googlenet.yaml \
                   --data_path=/path/to/imagenet \
                   --label_path=/path/to/imagenet/label \
                   --output_model=path/to/save

References

• Going deeper with convolutions

• Intel® Neural Compressor

Contributors

• mengniwang95 (Intel)
• airMeng (Intel)
• ftian1 (Intel)
• hshen14 (Intel)

License

BSD-3