Ignite - Githubissues

pennpolygons / cv-boilerplate

Open-source boilerplate for computer vision research

MIT License

16 stars 2 forks source link

Ignite #18

Closed goodmattg closed 4 years ago

goodmattg commented 4 years ago

Sets up train.py to use Ignite with simple logging callback
Runs on demo MNIST
Makefile is cleaned up (run make lint to black format the code

mchiquier commented 4 years ago

Looks great, except in datasets/init.py your get_dataloaders function's argumentsare currently the config and num_workers - do you not think that should also be specified as part of the config file?

goodmattg commented 4 years ago

Sure I’ll change it to max_num_workers. Typically, you set the number of workers to the number of cores on your cpu/gpu.

goodmattg commented 4 years ago

I spoke too early; how to optimally set num_workers is still up for debate, but the workers on CPU, which loads data and then sends to GPU with .to(device) or cuda()

https://discuss.pytorch.org/t/guidelines-for-assigning-num-workers-to-dataloader/813/6