princeton-vl / px2graph

Training code for "Pixels to Graphs by Associative Embedding"

BSD 3-Clause "New" or "Revised" License

133 stars 22 forks source link

Questions about training details #4

Open KaihuaTang opened 6 years ago

KaihuaTang commented 6 years ago

I've noticed that in your model.hg "

cnv = layers.res_block

cnv = lambda x,y: layers.cnv(x,y,3) " So does it mean that in your px2graph network, the hourglass structure uses a single conv to replace res_block and it improves the performance ?