diff --git a/demo/README.md b/demo/README.md
index 4720b53b6de61d4aa3c9b77e67ab3f75f79c0981..334f5a92e3253d92e9729a1db999b9f1f7ccb83a 100644
--- a/demo/README.md
+++ b/demo/README.md
@@ -25,3 +25,10 @@ By adding VisualDL as callbacks to `model.fit`,
 we can use the Python SDK in MxNet,
 but it seems that only the outside program can only retrieve parameters in epoch callbacks,
 that limits the number of steps for visualization.
+
+## PyTorch CNN
+Locates in `./pytorch`.
+
+This shows how to use VisualDL in PyTorch for a CNN on `cifar10` dataset. We visualize the loss in Scalar,
+two convolutional layers in Image, the change trend of conv1 weights in Histogram and the final model graph
+in Graph.