Questions about creating pertdata with own data files

snap-stanford / GEARS

GEARS is a geometric deep learning model that predicts outcomes of novel multi-gene perturbations

MIT License

200 stars 39 forks source link

Questions about creating pertdata with own data files #79

Open HelloWorldLTY opened 3 months ago

HelloWorldLTY commented 3 months ago

Hi, thanks for your work. I wonder if it is possible for us to access the sample_adata.h5ad file in the dataset tutorial. I intend to test if creating dataset functions have other errors or not. Thanks.

HelloWorldLTY commented 2 months ago

Moreover, I found that I met nan when training GEARs with own datasets

What could be the possible reasons? Thanks.

Start Training...
Epoch 1 Step 1 Train Loss: 0.2541
Epoch 1 Step 51 Train Loss: 0.0840
Epoch 1 Step 101 Train Loss: 0.1225
Epoch 1 Step 151 Train Loss: 0.1427
Epoch 1 Step 201 Train Loss: 0.0407
Epoch 1 Step 251 Train Loss: nan

yhr91 commented 2 months ago

Can you provide more details on what dataset you are using for training. Did you use the data tutorial as is or did you make modifications

Very likely this is a problem of data normalization or maybe there are zeroes or nans in your data

HelloWorldLTY commented 2 months ago

Hi, I used datasets from this paper (https://www.biorxiv.org/content/10.1101/2024.07.03.601903v1.abstract) which contains two large-scale perturbation datasets. I follow the recommended pre-processing step and verify that there are no nan values.

Thanks.