Compared with the fp32 ZFNet-512, int8 ZFNet-512's Top-1 accuracy drop ratio is 0.23%, Top-5 accuracy drop ratio is 0.10% and performance improvement is 1.78x.

Note

Different preprocess methods will lead to different accuracies, the accuracy in table depends on this specific preprocess method.

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.

ZFNet-512 is a deep convolutional networks for classification. This model's 4th layer has 512 maps instead of 1024 maps mentioned in the paper.

ILSVRC2013

Caffe2 ZFNet-512 ==> ONNX ZFNet-512

gpu_0/data_0: float[1, 3, 224, 224]

gpu_0/softmax_1: float[1, 1000]

random generated sampe 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

ZFNet-512-int8 and ZFNet-512-qdq are obtained by quantizing fp32 ZFNet-512 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.

onnx: 1.9.0 onnxruntime: 1.8.0

wget https://github.com/onnx/models/raw/main/vision/classification/zfnet-512/model/zfnet512-12.onnx

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=zfnet512.yaml \
                   --data_path=/path/to/imagenet \
                   --label_path=/path/to/imagenet/label \
                   --output_model=path/to/save

• Visualizing and Understanding Convolutional Networks

• Intel® Neural Compressor

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

MIT