SqueezeNet 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. SqueezeNet models are highly efficient in terms of size and speed while providing good accuracies. This makes them ideal for platforms with strict constraints on size.
Description
SqueezeNet is a small CNN which achieves AlexNet level accuracy on ImageNet with 50x fewer parameters. SqueezeNet requires less communication across servers during distributed training, less bandwidth to export a new model from the cloud to an autonomous car and more feasible to deploy on FPGAs and other hardware with limited memory.
Model
Squeezenet 1.0 gives AlexNet level of accuracy with 50X fewer parameters. Run SqueezeNet 1.0 in browser - implemented by ONNX.js.
SqueezeNet 1.1 has 2.4x less computation and slightly fewer parameters than SqueezeNet 1.0, without sacrificing accuracy.
Compared with the fp32 SqueezeNet 1.0, int8 SqueezeNet 1.0's Top-1 accuracy drop ratio is 0.65%, Top-5 accuracy drop ratio is 0.14% and performance improvement is 1.31x.
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.
Inference
We used MXNet as framework with gluon APIs to perform inference for SqueezeNet 1.1. 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.
SqueezeNet 1.0 is converted from Caffe2 -> ONNX.
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 a jpeg image.
float[1, 3, 224, 224]
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 model on the validation set are mentioned above. The accuracies have been calculated on center cropped images with a maximum deviation of 1.2% (top-5 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 SqueezeNet.
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
SqueezeNet 1.0-int8 and SqueezeNet 1.0-qdq are obtained by quantizing fp32 SqueezeNet 1.0 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.
