Update README.md

README.md

@@ -25,6 +25,10 @@ For the branch **master**, the training of the transductive learning on Cora tas
 
 A small note about initial sparse matrix operations of https://github.com/tkipf/pygcn: they have been removed. Therefore, the current model take ~7GB on GRAM.
 
+# Sparse version GAT
+
+We develop a sparse version GAT using pytorch. There are numerically instability because of softmax function. Therefore, you need to initialize carefully. To use sparse version GAT, add flag `--sparse`. The performance of sparse version is similar with tensorflow. On a Titan Xp takes 0.08~0.14 sec.
+
 # Requirements
 
 pyGAT relies on Python 3.5 and PyTorch 0.4.1 (due to torch.sparse_coo_tensor).