gmh14 / IRLAS

PyTorch implementation of IRLAS on ImageNet

IRLAS: Inverse Reinforcement Learning for Architecture Search

PyTorch implementation of IRLAS.

If you use the code, please cite:

@inproceedings{guo2019irlas,
  title={Irlas: Inverse reinforcement learning for architecture search},
  author={Guo, Minghao and Zhong, Zhao and Wu, Wei and Lin, Dahua and Yan, Junjie},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={9021--9029},
  year={2019}
}

Requirements

• PyTorch 0.4.0

Data and Model Preparation

• Download the ImageNet validation set. Put the images to **your_dataset_path**/val and the label file to **your_dataset_path**/meta/val.txt. The label file could be downloaded from val.txt.
• Download the pretrained model IRLAS-ImageNet-mobile and move it to pretrained_models

Test Accuracy

Set --dataset_path=**your_dataset_path** in test.sh and execute,

bash test.sh

The printed lines should read:

[**current time**][validate.py][line: 133][INFO]   Total params: 9.96M
[**current time**][validate.py][line:  70][INFO] Test: [0/196]   Time 11.517 (11.517)    Loss 1.1165 (1.1165)    Prec@1 72.266 (72.266)  Prec@5 92.578 (92.578)
[**current time**][validate.py][line:  70][INFO] Test: [100/196] Time 0.205 (0.574)      Loss 1.1337 (1.0932)    Prec@1 73.438 (75.174)  Prec@5 91.016 (92.168)
[**current time**][validate.py][line:  72][ INFO]  * Prec@1 75.150 Prec@5 92.090

Test Inference Latency

To match the measurement of inference latency in real products, the model is converted to Caffe in caffe_prototxt. Note that all the BatchNorm & Scale layers have been merged to Convolution layers, which is widely used to speed up inference in real products.

Download libnvinfer.so.4 & _netrt.cpython-36m-x86_64-linux-gnu.so and put them under tools/netrt.

The inference latency is measured on TensorRT framework. The test tools are included in libnvinfer.so.4 and _netrt.cpython-36m-x86_64-linux-gnu.so. Unfortunately, these .so is compiled using SenseTime internal tools, which I do not have access to. However, if you happen to have all the required sources (you can check by executing ldd libnvinfer.so.4 & ldd _netrt.cpython-36m-x86_64-linux-gnu.so), you can test the influence latency of the model by

bash test_time.sh

The last printed lines should read:

IRLAS avg cost: 9.308ms

The above latency is measured on TiTan Xp with 16 batch size, 224x224 input size. The latency may have a difference of around $\pm0.5$ms due to the fluctuation of occupancy rate or the difference of platform. When measured on 1080Ti, the latency will increase around $1$ms.