Skip to content

FAQ.md

Latest commit

 

History

History
44 lines (31 loc) · 1.98 KB

FAQ.md

File metadata and controls

44 lines (31 loc) · 1.98 KB
Raw

FAQ

We will keep adding more sections on this page, as we receive more questions from users

1. how to load a pre-train model for transfer learning?

Usually, you can do this by setting model.model_extra.pre-train with the checkpoint path. A special case is for loading a pre-trained FCN as the generator of a GAN model, while the discriminator will still be initialized regularly with Kaiming method. In this case, you need to set model.net.generator.init_weight with the path to the pretrained FCN model. See an example HERE.

2. how to use multi-GPU training or half precision training to better utilize GPU(s)?

This is very simple, just setting the parameters in trainer.

trainer:
  verbose: True
  params:
    gpus: 5
    precision: 16
    max_epochs: 10000
    detect_anomaly: True

3. How to monitor training progress?

Tensorboard is used for tracking the training progress by default. Everything is saved under lightning_logs/version_X. Then after installing tensorboard, you can simply run tensorboard --logdir=./lightning_logs/version_X. For advacned users, if you wish to change what is being logged with tensorboard, you can just modify the corresponding lightning module file, like here.

How to select different GPUs?

By default, when you run run_im2im --config myconfig.yaml, all GPUs available on your machine are usable by the program. Then, if you select to use 1 GPU, then the first GPU will be used. If you want to run on a specific GPU(s), you can do CUDA_VISIBLE_DEVICES=3 run_im2im --config myconfig.yaml or CUDA_VISIBLE_DEVICES=1,3,5 run_im2im --config myconfig.yaml

How to automatically select which GPU to use?

set device = "auto" in trainer. For example:

trainer:
  verbose: True
  params:
    accelerator: "gpu"
    devices: "auto"
    precision: 16
    max_epochs: 2000
    detect_anomaly: True