ResNet 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. ResNet models provide very high accuracies with affordable model sizes. They are ideal for cases when high accuracy of classification is required.
Description
Deeper neural networks are more difficult to train. Residual learning framework ease the training of networks that are substantially deeper. The research explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. It also provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset the residual nets were evaluated with a depth of up to 152 layers — 8× deeper than VGG nets but still having lower complexity.
The model below are ResNet v1 and v2. ResNet models consists of residual blocks and came up to counter the effect of deteriorating accuracies with more layers due to network not learning the initial layers.
ResNet v2 uses pre-activation function whereas ResNet v1 uses post-activation for the residual blocks. The models below have 18, 34, 50, 101 and 152 layers for with ResNetv1 and ResNetv2 architecture.
Compared with the fp32 ResNet50, int8 ResNet50's Top-1 accuracy drop ratio is 0.27%, Top-5 accuracy drop ratio is 0.01% and performance improvement is 1.82x.
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 image needs to be preprocessed before fed to the network.
The first step is to extract a 224x224 crop from the center of the image. For this, the image is first scaled to a minimum size of 256x256, while keeping aspect ratio. That is, the shortest side of the image is resized to 256 and the other side is scaled accordingly to maintain the original aspect ratio. After that, the image is normalized with mean = 255*[0.485, 0.456, 0.406] and std = 255*[0.229, 0.224, 0.225]. Last step is to transpose it from HWC to CHW layout.
The described preprocessing steps can be represented with an ONNX model:
The post-processing involves calculating the softmax probability scores for each class. You can also sort them to report the most probable classes. Check imagenet_postprocess.py for code.
To do quick inference with the model, check out Model Server.
Caffe2 Version of ResNet50 uses the ImageNet dataset from 2015 -- ILSVRC2015.
Validation accuracy
The accuracies obtained by the models on the validation set are mentioned above. The validation was done using center cropping of images unlike the paper which uses ten-cropping. We expect an increase of 1-2% in accuracies using ten cropping and that would lead to accuracies similar to 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 ResNet.
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
ResNet50-int8 and ResNet50-qdq are obtained by quantizing ResNet50-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=resnet50_v1_5.yaml \3--output_model=path/to/save
Model inference
We use onnxruntime to perform Resnet50_fp32 and Resnet50_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.
References
ResNetv1Deep residual learning for image recognition
He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770-778. 2016.
ResNetv2Identity mappings in deep residual networks
He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
In European Conference on Computer Vision, pp. 630-645. Springer, Cham, 2016.
