Label_prop.py memory requirement

[drchangliu]

It is said that "256GB CPU memory required to run this script" for label_prop.py. We ran this on a machine with 256GB main memory and still ran out of memory. The OS used a little bit of memory but not much. There was no other process running on the machine. So did you ran this successfully on a machine with 256GB physical memory? Just confirming the memory requirement. Thanks.

Chang

[rusty1s]

Sorry for the inconveniences. The label_prop.py example indeed may consume slightly more memory than 256GB. I added a --low-memory option to the script that converts the one-hot labels to torch.half. That should be fine to run the script with 256GB of memory (consumes about 160GB of memory for me). The validation accuracy is nearly identical (58.43%).

[drchangliu]

The new version works. Thanks a lot!

[zbn123]

The following error occurs when I converts the one-hot labels to torch.half

[zbn123]

How do you solve this problem

[weihua916]

Hmm, it works fine on my end:

python label_prop.py --low-memory

Namespace(alpha=0.9, low_memory=True, num_layers=3)
Reading adjacency matrix... Done! [250.81s]
Propagating labels... Done! [380.19s]
Train: 0.9909, Valid: 0.5843