VGG models perform image classification - they take images as input and classify the major object in the image into a set of pre-defined classes. They are trained on ImageNet dataset which contains images from 1000 classes.
VGG models provide very high accuracies but at the cost of increased model sizes. They are ideal for cases when high accuracy of classification is essential and there are limited constraints on model sizes.
Description
VGG presents the effect of the convolutional network depth on its accuracy in the large-scale image recognition setting. VGG networks have increased depth with very small (3 × 3) convolution filters, which showed a significant improvement on the prior-art configurations achieved by pushing the depth to 16–19 weight layers. The work secured the first and the second places in the localization and classification tracks respectively in ImageNet Challenge 2014. The representations from VGG generalize well to other datasets, where they achieve state-of-the-art results.
MXNet VGG ==> ONNX VGG [16, 16-bn, 19, 19-bn]
Caffe2 VGG-19 ==> ONNX VGG [19-caffe2]
ONNX vgg16 ==> Quantized ONNX vgg16
Model
The models below are variant of same network with different number of layers and use of batch normalization. VGG 16 and VGG 19 have 16 and 19 convolutional layers respectively. VGG 16_bn and VGG 19_bn have the same architecture as their original counterparts but with batch normalization applied after each convolutional layer, which leads to better convergence and slightly better accuracies.
Compared with the fp32 VGG 16, int8 VGG 16's Top-1 accuracy drop ratio is 0.06%, Top-5 accuracy drop ratio is 0.03% and performance improvement is 2.31x.
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.
We used MXNet as framework with gluon APIs to perform inference. View the notebook imagenet_inference to understand how to use above models for doing inference. Make sure to specify the appropriate model name in the notebook.
Input
All pre-trained models expect input images normalized in the same way, i.e. mini-batches of 3-channel RGB images of shape (N x 3 x H x W), where N is the batch size, and H and W are expected to be at least 224.
The inference was done using jpeg image.
Preprocessing
The images have to be loaded in to a range of [0, 1] and then normalized using mean = [0.485, 0.456, 0.406] and std = [0.229, 0.224, 0.225]. The transformation should preferrably happen at preprocessing. Check imagenet_preprocess.py for code.
The post-processing involves calculating the softmax probablility scores for each class and sorting them to report the most probable classes. Check imagenet_postprocess.py for code.
To do quick inference with the model, check out Model Server.
The accuracies obtained by the models on the validation set are mentioned above. The accuracies have been calculated on center cropped images with a maximum deviation of 0.4% (top-1 accuracy) from the paper.
Training
We used MXNet as framework with gluon APIs to perform training. View the training notebook to understand details for parameters and network for each of the above variants of VGG.
Validation
We used MXNet as framework with gluon APIs to perform validation. Use the notebook imagenet_validation to verify the accuracy of the model on the validation set. Make sure to specify the appropriate model name in the notebook.
Quantization
VGG 16-int8 and VGG 16-qdq are obtained by quantizing VGG 16-fp32 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.
Make sure to specify the appropriate dataset path in the configuration file.
bash
1bash run_tuning.sh --input_model=path/to/model \# model path as *.onnx2--config=vgg16.yaml \3--output_model=path/to/save
Model inference
We use onnxruntime to perform VGG 16-fp32 and VGG 16-int8 inference. View the notebook onnxrt_inference to understand how to use these 2 models for doing inference as well as which preprocess and postprocess we use.
