SpiceMix GPU

[Jieran-S]

SpaceHack2023/method/SpiceMix/SpiceMix.py

Lines 163 to 174 in 57c4710

	device = config.pop("device", "cpu")
	dtype = config.pop("dtype", "float32")
	if dtype == "float32":
	dtype = torch.float32
	elif dtype == "float64":
	dtype = torch.float64
	torch_context = {
	"device": device,
	"dtype": dtype,
	}

SpaceHack2023/method/SpiceMix/config/config_1.json

Lines 1 to 8 in 57c4710

	{
	"K": 15,
	"lambda_Sigma_x_inv": 1e-4,
	"device": "cuda:0",
	"dtype": "float64",
	"num_preiterations": 5,
	"num_iterations": 200
	}

SpiceMix uses GPU (cuda) as device and it raises an error when running our pipeline atm as we don't have GPU access. I don't know much about GPU but I added the line below to make it work. It seems it runs slowly and returns me this warning (last code chunk). Does it make sense? Do you have any recommendations?

torch_context["device"] = device if torch.cuda.is_available() else 'cpu'

warning:

/home/jovyan/scratch/SpaceHack2/userfolders/jsun/workflow/SpaceHack2023/workflows/.snakemake/conda/397c14e72e8de8734c04bce3a0f813c6_/lib/python3.8/site-packages/popari/util.py:499: UserWarning: Creating a tensor from a list of numpy.ndarrays is extremely slow. Please consider converting the list to a single numpy.ndarray with numpy.array() before converting to a tensor. (Triggered internally at ../torch/csrc/utils/tensor_new.cpp:230.)
  i = torch.LongTensor(indices)

@alam-shahul (maybe you know)

[niklasmueboe]

Given that the warning is created inside popari I don't think we have much influence on that, unless we are passing "incorrrect" input, which I don't think is the case. So I guess we can ignore the warning for now.

[alam-shahul]

The warning is not that important and it might also trigger when using the GPU. It will be fixed in future versions of Popari.

More importantly, SpiceMix/Popari is designed to be used with a GPU; although it is possible to run with a CPU, this is computationally infeasible for any sizable spatial transcriptomics dataset.