Pytorch实现PNASNet(Progressive Neural Architecture Search)
Xu Jing
实现了:
-
数据增强
- 随机水平翻转
- 随机竖直翻转
- 随机亮度值(brightness)
- 随机色调(hue)
- 随机饱和度(saturation)
- 随机对比度(contrast)
-
Pytorch加载训练集的pipeline
-
基于ImageNet预训练的模型微调的PNASNet及训练
-
单张图像和视频的推断
训练的参数设置:
- batch size:16
- epochs:300
- Loss: CrossEntropyLoss
- optim: Adam
- lr: feature_param:0.0001, linear_param: 0.001
- 硬件:ubuntu 16.04 64G, Tesla V100(32G)
模型训练:
python3 model.py
推断模型:
python3 inference.py
python3 inference_video.py
测试结果: