get_freer_gpu

deepinv.utils.get_freer_gpu(verbose=True, use_torch_api=True, hide_warnings=False)

Returns the GPU device with the most free memory.

Use in conjunction with torch.cuda.is_available().

If use_torch_api=True then attempts to select GPU using only torch commands, otherwise uses system driver to detect GPUs (via nvidia-smi command). The first method may be slower but is more reliable as the former depends on environment settings. If system method is chosen and fails, the call falls back to using torch commands and a warning is printed. If no CUDA devices are detected, then None is returned.

Parameters:

• verbose (bool) – print selected GPU index and memory

• use_torch_api (bool) – use torch commands if True, or Nvidia driver otherwise

• hide_warnings (bool) – supress all warnings for all methods

Return torch.device device:

selected cuda device

Warning

GPU indices in nvidia-smi may not match those in PyTorch if in your environment CUDA_DEVICE_ORDER is not set to PCI_BUS_ID: https://discuss.pytorch.org/t/gpu-devices-nvidia-smi-and-cuda-get-device-name-output-appear-inconsistent/13150 If the variable is not set or has different value, the call to will print a warning (if not supressed with hide_warnings=True) but will not change the device.

Examples using get_freer_gpu:

Blind deblurring with kernel estimation network

Blind deblurring with kernel estimation network

Blind denoising with noise level estimation

Blind denoising with noise level estimation

Single-pixel imaging with Spyrit

Single-pixel imaging with Spyrit

Inverse scattering problem

Inverse scattering problem

Poisson denoising using Poisson2Sparse

Poisson denoising using Poisson2Sparse

Reducing the memory and computational complexity of unfolded network training

Reducing the memory and computational complexity of unfolded network training