• A tf train framework for auto train the nice mode.
• This project now is developing, now as code showing

• ON single 8-V100, 936 per sec at imagenet batch 1024

• sample API & module: keep sample & keep easy to replace.
• multi GPU support:
  • replcated for P2P with GPU such as NVLINK
  • BlacePlament for low latency and limit bandwith such as PCI-E*16

1. distribute support: horovd(MPI) and broadcast_proxy.

   • horovd (MPI & Roce)

     • each a singe process

   • broadcast_proxy： seastear, DPDK

2. Auto hyterparam select: detect the cpu, gpu, network limits, as much use evey resource
   • detect system configure
   • train and eval speed measure
3. Model zoo & Data zoo: shared state of art some models.
   • imagenet

###　Benchmark

• DGX workstation: V100*4, NVLink 100GB, 1100 sample/sec for imagenet resnet50 batchsize:128