syntheticSuccess/c19/tabsyn/tabsyn-c19-20260426_203054/_tabsyn_train.py

import os, sys, subprocess

work_dir = "/work/output-SpecializedModels/c19/tabsyn/tabsyn-c19-20260426_203054"
dataname = "tabsyn_c19"
tabsyn_root = "/workspace/tabsyn"

assert os.path.exists(tabsyn_root), f"TabSyn source not mounted: {tabsyn_root}"

old = os.environ.get("PYTHONPATH", "")
os.environ["PYTHONPATH"] = tabsyn_root + (os.pathsep + old if old else "")
sys.path.insert(0, tabsyn_root)

os.chdir(tabsyn_root)

# Symlink data dir into TabSyn data/
data_link = os.path.join(tabsyn_root, "data", dataname)
data_src = os.path.join(work_dir, "data", dataname)
os.makedirs(os.path.join(tabsyn_root, "data"), exist_ok=True)
if os.path.exists(data_link):
    os.remove(data_link)
os.symlink(data_src, data_link)

env = os.environ.copy()
env.setdefault("TABSYN_RESUME", "1")
env.setdefault("TABSYN_VAE_BATCH_SIZE", "1024")
_te = 1000
if _te is not None:
    env["TABSYN_VAE_EPOCHS"] = str(_te)
    env["TABSYN_DIFFUSION_MAX_EPOCHS"] = str(max(_te + 1, 2))

# Data preprocessing is done on the host side (_prepare_data_dir)
# which creates .npy files, train/test CSVs, and info.json

# Step 1: Train VAE (produces latent embeddings)
print(f"[TabSyn] Step 1/2: Training VAE in {tabsyn_root}, dataname={dataname}")
ret = subprocess.run(
    [sys.executable, "main.py",
     "--dataname", dataname,
     "--mode", "train",
     "--method", "vae",
     "--gpu", "0"],
    cwd=tabsyn_root,
    env=env
)
if ret.returncode != 0:
    print("[TabSyn] VAE training failed")
    sys.exit(ret.returncode)

# Step 2: Train diffusion model on latent space
print(f"[TabSyn] Step 2/2: Training diffusion model")
ret = subprocess.run(
    [sys.executable, "main.py",
     "--dataname", dataname,
     "--mode", "train",
     "--method", "tabsyn",
     "--gpu", "0"],
    cwd=tabsyn_root,
    env=env
)
if ret.returncode != 0:
    print("[TabSyn] Diffusion training failed")
    sys.exit(ret.returncode)
print("[TabSyn] Training complete (VAE + Diffusion)")