ResNet

Use cases

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.

MXNet ResNet-v1 ==> ONNX ResNet-v1 [18, 34, 50, 101, 152]

MXNet ResNet-v2 ==> ONNX ResNet-v2 [18, 34, 50, 101, 152]

Caffe2 ResNet-50 ==> ONNX ResNet [50-caffe2]

ONNX ResNet50-v1 ==> Quantized ONNX ResNet50-v1

Model

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.

Run ResNet-50 in browser - implemented by ONNX.js with ResNet50-Caffe2 release 1.2

Details of Resnet-50 quantization & evaluation - implemented by Intel® Low Precision Optimization Tool (LPOT)

• ResNet V1:

Note the latency depends on the test hardware and the displayed results are tested by Intel® Xeon® Platinum 8280 Processor.
Compared with the fp32 model, we get an accuracy drop of 0.2% and a performance improvement of 1.9X after quantization.

• ResNet V2:

Inference

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 preferably happen at preprocessing.

The following code shows how to preprocess a NCHW tensor:

import numpy

def preprocess(img_data):
    mean_vec = np.array([0.485, 0.456, 0.406])
    stddev_vec = np.array([0.229, 0.224, 0.225])
    norm_img_data = np.zeros(img_data.shape).astype('float32')
    for i in range(img_data.shape[0]):
         # for each pixel in each channel, divide the value by 255 to get value between [0, 1] and then normalize
        norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i]
    return norm_img_data

Check imagenet_preprocess.py for additional sample code.

Output

The model outputs image scores for each of the 1000 classes of ImageNet.

Postprocessing

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.

Dataset

Dataset used for train and validation: ImageNet (ILSVRC2012). Check imagenet_prep for guidelines on preparing the dataset.

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.

References

• ResNetv1
  Deep 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.

• ResNetv2
  Identity 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.

• MXNet, Gluon model zoo, GluonCV

Contributors

• ankkhedia (Amazon AI)
• abhinavs95 (Amazon AI)

License

Apache 2.0