Files

51 lines
2.5 KiB
Python

"""Check model resources, checkpoint loading and a short simulation rollout."""
from pathlib import Path
from dataclasses import asdict
import hashlib,json,xml.etree.ElementTree as ET
import mujoco,torch
import src.tasks
from mjlab.envs import ManagerBasedRlEnv
from mjlab.rl import RslRlVecEnvWrapper
from mjlab.tasks.registry import load_env_cfg,load_rl_cfg,load_runner_cls
from src.policy import load_policy
ROOT=Path(__file__).resolve().parents[1]
def main():
result={}
urdf=ROOT/'models/superspider/superspider.urdf'
tree=ET.parse(urdf)
for mesh in tree.findall('.//mesh'):
assert (urdf.parent/mesh.attrib['filename']).is_file()
model=mujoco.MjModel.from_xml_path(str(urdf))
assert model.njnt==18
result['urdf']={'links':len(tree.findall('link')),'joints':model.njnt}
model=mujoco.MjModel.from_xml_path(str(ROOT/'models/superspider_mjlab/superspider.xml'))
result['mjcf']={'nq':model.nq,'nv':model.nv}
for entry in json.loads((ROOT/'checkpoints/manifest.json').read_text()):
assert hashlib.sha256((ROOT/entry['file']).read_bytes()).hexdigest()==entry['sha256']
result['checkpoint_hashes']=5
device='cuda:0' if torch.cuda.is_available() else 'cpu'
for task,files in [('Custom-Superspider-Stairs-Distill',['student_5999.pt','student_1600.pt']),('Custom-Superspider-Flat',['flat_5999.pt']),('Custom-Superspider-Stairs-Teacher',['teacher_3800.pt']),('Custom-Superspider-Stairs',['blind_10000.pt'])]:
cfg=load_env_cfg(task,play=True);cfg.scene.num_envs=1
agent=load_rl_cfg(task)
env=ManagerBasedRlEnv(cfg,device=device)
wrapped=RslRlVecEnvWrapper(env,clip_actions=agent.clip_actions)
runner=load_runner_cls(task)(wrapped,asdict(agent),device=device)
for name in files:
path=ROOT/'checkpoints'/name
policy=load_policy(runner,path,device)
saved=torch.load(path,map_location=device,weights_only=False)
key='student' if 'student_state_dict' in saved else 'actor'
loaded=getattr(runner.alg,key).state_dict()
assert all(torch.equal(loaded[k],v) for k,v in saved[key+'_state_dict'].items())
obs,_=wrapped.reset()
with torch.no_grad():
for _ in range(8):
action=policy(obs);assert action.shape==(1,18) and torch.isfinite(action).all()
obs,*_=wrapped.step(action)
result[name]={'loaded_exactly':True,'actions':18,'simulation_steps':8}
env.close()
print('VERIFICATION_JSON '+json.dumps(result))
if __name__=='__main__':main()