Using pytorch then using tfrecords reading data about second.pytorch

[traveller59]

Actually there is a complete tensorflow implementation of this project which shares all code except ./pytorch and contains a faster sparse convnet implementation. I have no plan to release tf code because of two reasons:

1. tf sparseconv is done in a company.
2. due to limitations of tensorflow input (it uses c++ multi-thread rather than multiprocessing in pytorch) which isn't compatible with python pre-processing function, the training speed of tensorflow code is very slow (~1.5x-2x slower).

[lucasjinreal]

@traveller59 So that it actually didn't used in this project? I think pytorch official SparseConvNet should be reliable and fast though

[traveller59]

@jinfagang some opinions are only for tensorflow and aren't used in pytorch code. the official SparseConvNet actually has some performance problems (rule generation, hand-written gemm).