JeanRadig
commented
3 months ago - Update train.sh. You can modify train.sh to accommodate for the right arguments.
- Modified train.sh in function of the arguments that can actually be passed, as defined for each function. Previous commands could not run due to unexpected arguments such as e.g. --num_genes for the diffusion and classifier. Also modified the arguments name, when they did not correspond to those as defined in the functions.
cd VAE
echo "Training Autoencoder, this might take a long time"
CUDA_VISIBLE_DEVICES=0 python VAE_train.py --data_dir 'path/to/anndata.h5ad' --num_genes 18996 --save_dir 'path/to/saved_VAE_model' --max_steps 200000
echo "Training Autoencoder done"
cd ..
echo "Training diffusion backbone"
CUDA_VISIBLE_DEVICES=0 python cell_train.py --data_dir 'path/to/anndata.h5ad' --vae_path 'path/to/saved_VAE_model/VAE_checkpoint.pt' \
--save_dir 'path/to/saved_diffusion_model' --model_name 'my_diffusion' --save_interval 20000
echo "Training diffusion backbone done"
echo "Training classifier"
CUDA_VISIBLE_DEVICES=0 python classifier_train.py --data_dir 'path/to/anndata.h5ad' --model_path "path/to/saved_classifier_model" \
--iterations 40000 --vae_path 'path/to/saved_VAE_model/VAE_checkpoint.pt'
echo "Training classifier, done"- Guide the user. You can insert some more explanation at the beginning of your notebooks to guide the user. Example on top of script_diffusion_umap.ipynb:
Before running the following code, ensure that you have trained the VAE model, the diffusion model, and the classifier model.
Given an AnnData file you wish to analyze, named anndata.h5ad, please follow these steps in your command line to create the necessary input data:
-
Train VAE:
- Prior to training, create the folder
path/to/saved_VAE_model. - In the terminal, navigate to the
VAEdirectory:cd VAE - Run the following command to train the Autoencoder:
echo "Training Autoencoder, this might take a long time" CUDA_VISIBLE_DEVICES=0 python VAE_train.py --data_dir 'path/to/anndata.h5ad' --num_genes 18996 --save_dir 'path/to/saved_VAE_model' --max_steps 200000 echo "Training Autoencoder done"
- Prior to training, create the folder
-
Train the Diffusion Model:
- Prior to training, create the folder
path/to/saved_diffusion_model. - In the terminal, navigate back to the root directory:
cd .. - Run the following command to train the diffusion backbone:
echo "Training diffusion backbone" CUDA_VISIBLE_DEVICES=0 python cell_train.py --data_dir 'path/to/anndata.h5ad' --vae_path 'path/to/saved_VAE_model/VAE_checkpoint.pt' \ --save_dir 'path/to/saved_diffusion_model' --model_name 'my_diffusion' --save_interval 20000 echo "Training diffusion backbone done"
- Prior to training, create the folder
-
Train the Classifier:
- Prior to training, create the folder
path/to/saved_classifier_model. - Run the following command to train the classifier:
echo "Training classifier" CUDA_VISIBLE_DEVICES=0 python classifier_train.py --data_dir 'path/to/anndata.h5ad' --model_path "path/to/saved_classifier_model" \ --iterations 40000 --vae_path 'path/to/saved_VAE_model/VAE_checkpoint.pt' echo "Training classifier, done"
- Prior to training, create the folder
Once the models are trained, you need to create the .npz files which will serve as input for the following steps. To do this:
-
Unconditional Sampling (Generate Data from Diffusion Model):
- Prior to sampling, create the folder
path/to/saved_unconditional_sampling. - Run the following command to perform unconditional sampling:
# Unconditional sampling python cell_sample.py --model_path "path/to/saved_diffusion_model/checkpoint.pt" --sample_dir "path/to/saved_unconditional_sampling"
- Prior to sampling, create the folder
-
Conditional Sampling (Generate Data from Classifier Model):
- Prior to sampling, create the folder
path/to/saved_conditional_sampling. - You also need to modify the
main()function inclassifier_sample.pyto create the samples based on your specified condition. NEEDS MORE DOCUMENTATION: PROVIDE CLEAR EXAMPLE - Run the following command to perform conditional sampling:
# Conditional sampling python classifier_sample.py --model_path "path/to/saved_diffusion_model/checkpoint.pt" --classifier_path "path/to/saved_classifier_model/checkpoint.pt" --sample_dir "path/to/saved_conditional_sampling"
- Prior to sampling, create the folder
Ensure that you replace 'path/to/anndata.h5ad', 'path/to/saved_VAE_model', 'path/to/saved_diffusion_model', and 'path/to/saved_classifier_model' with the actual paths in your system. Additionally, make sure to adjust any other parameters according to your specific setup and requirements.
Sbs12
EperLuo
Hi!
I am trying to reproduce Fig. 3, in particular the umap of muris with the different cell types.
I have trained the models on the muris dataset. I try to get the conditional samples. I do not understand what parameters I should pass in the function main of classifier_sample.py to get the 12 samples, as needed in Jupyter notebook "script_diffusion_map", lines:
index2 = [0,1,2,3,4,5,6,7,8,9,10,11] for i in index2: npzfile=np.load(f'/data1/lep/Workspace/guided-diffusion/output/muriscondi/muris{i}_scimilarity.npz',allow_pickle=True)
Could you please let me know how you have done it?