A novel method named stDiff investigates the potential of employing diffusion models for single-cell omics generation.

if gene num < 512, batchsize = 2048, hiddensize = 512;
if gene num > 512 and num < 1024, batchsize = 512, hiddensize = 1024;
if gene num > 1024 and num < 2048, batchsize = 512, hiddensize = 2048.

The specific parameter settings follow those in the Spatial Benchamark, and the baseline code is adapted from Spatial Benchmark, whose parameter settings also follow the default repositories of their respective models.
The code for uniport is referenced in the example on its official website Impute genes for MERFISH.
The code for SpatialScope is referenced in SpatialScope for imputation. \

https://zenodo.org/records/10613236?token=eyJhbGciOiJIUzUxMiJ9.eyJpZCI6IjI5NDk0ZWE2LWMxZWEtNGU5ZS05ODQ1LTAyZTU5ZThhOTQ4NiIsImRhdGEiOnt9LCJyYW5kb20iOiJhODEzZjZhMDkxOTIwMjk5Y2I1ZjVkYzNmNWUwNjI0NyJ9.KX2t2EqEFatcx7PfIf_gkjmMkBjhknThlgJC3tuZOotwcDLwah8nqMvfy3uqmmiTariWNS3PLkZHjdIyzmqZPQ

Five sets of cross-validated checkpoints(random_state = 0) for all datasets have been uploaded to https://drive.google.com/drive/folders/1Qe1npltWkCuSZFBBG26cx7PUiM8K1Ghe?usp=sharing.

conda create -n stDiff python=3.6
conda activate stDiff
pip install -r requirements.txt

or

pip install stDiff-sc

The datasets 2-16 in the experiment were all from the paper Benchmarking spatial and single-cell transcriptomics integration methods for transcript distribution prediction and cell type. The raw data were initially processed in 'process/data_process.py' to convert the txt file to h5ad, and only the genes shared by both ST and scRNA-seq were retained. \

test-stDiff for stDiff
test-baseline for baselines
test-spscope for SpatialScope

from model.stDiff_model import DiT_stDiff
from model.stDiff_scheduler import NoiseScheduler
from model.stDiff_train import normal_train_stDiff
from model.sample import sample_stDiff
from process import *

python test-stDiff.py --sc_data 'sc_dataset(h5ad)' --sp_data 'sp_dataset(h5ad)' --document 'stDiff_result_name' --batch_size 512 --hidden_size 1024
python test-baseline.py --sc_data 'sc_dataset(h5ad)' --sp_data 'sp_dataset(h5ad)' --document 'base_result_name'   

Use bash run.sh run both methods

Error is reported on RTX 3070, 3090 GPUs, the reason is that the default torch=1.9.0 cuda version does not match the GPU, after installing with pip, try pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html

RuntimeError: CUDA error: no kernel image is available for execution on the device
CUDA kernel errors might be asynchronously reported at some other API call,so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.