Store classfication.
StoreTag, or you can find it here.
-
0: Split train and test.
-
1: Building an ssd(more about ssd) detector in paddlepaddle, train the neural network with train dataset. With special thanks to this blog, we can fix an agonizing bug(Float error when start training).
-
2: Building a ResNet to classify the tag area obtained in step1, this part of job is done by my teammate Xu.
-
To run the ssd net, you need gpus, we have tested on Debian9, Cuda8, cudnn7, python2.7.15(
pip2 install paddlepaddle-gpu==0.14.0.post87
) and kali-rolling, Cuda9.1, cudnn7, python2.7.15(pip2 install paddlepaddle-gpu
) -
To start training, you need a pre-trained model, or there will be float errors. Modify the pre-trained model as what your net is, here you need to delete all the files with 'mbox'(
rm *mbox*
).