optimizer speed for resnet101 about pipedream HOT 3 OPEN

I think I find a problem and make improvement in my fork.
Could you verify it? it gives much faster result for resnet101, but I'm not sure it is really correct.

https://github.com/jeageun/pipedream/blob/master/graph/graph.py

Thanks

from pipedream.

Can you summarize what the intended purpose of the change is?

from pipedream.

I think your implementation label the depth for all layers and your implementation uses the Depth First Search approach. This approach is good if there is no merge point, but if there is merge point, all labeling using a shorter path is not valid. Since you don't remove this invalid path from a work queue, you need to go through the invalid path until it ends. After this invalid path ends, the other valid path always updates depth value.

In my modification code,

1. I remove these redundant paths before it ends. However, if we use DFS, we can't realize it is valid or not, I use BFS.
2. I skip building depth if there is the previous result for depth.

from pipedream.