MorphDiff

The codes in MorphDiff are modified based on the stable diffusion v1 main framework.

Original framework: CompVis/stable-diffusion

---

Training Process

The training process consists of two main stages:

1. Training the Variational AutoEncoder (VAE)

• Example script: MorphDiff/vae.sh
• Configuration file: MorphDiff/configs/autoencoder/autoencoder_kl_32x32x4_5c.yaml

Detailed information about input file paths and parameters can be found in the configuration file.

2. Training the Latent Diffusion Model

• Example script: MorphDiff/dm.sh
• Configuration file: MorphDiff/configs/ldm/morph_5c.yaml

Detailed information about input file paths and parameters can be found in the configuration file.

---

Inference Stage

Examples for both gene to image and image to image modes are provided as shell scripts: MorphDiff/scripts/gene2img.sh and MorphDiff/scripts/img2img.sh.

1. Gene to Image Mode

Example command from gene2img.sh:

python gene2img_dpm.py \
  --gene_path data/gene_expression.npy \
  --outdir result/g2i/ \
  --output_name /result/g2i.npy \
  --H 128 \
  --W 128 \
  --model_path /model/drug_ldm.ckpt \
  --config_path config/morph_5c.yaml \
  --scale 1.0 \
  --batch_size 2 \
  --ddim_steps 100

2. Image to Image Mode

Example command from img2img.sh:

python MorphDiff/scripts/gene_img2img_5c.py \
  --init_img_path data/drug_base_demo.npy \
  --gene_path data/gene_expression_demo.npy \
  --outdir ../results/drug_inference_demo/ \
  --output_name ../results/drug_inference_demo.npy \
  --model_path MorphDiff/checkpoints/drug_ldm.ckpt \
  --config_path MorphDiff/configs/ldm/morph_5c.yaml \
  --scale 1.0 \
  --strength 0.55 \
  --seed 10

---