MoNo-light / scripts /common.py
yushuang88's picture
Upload folder using huggingface_hub
db32cff verified
Raw History Blame Contribute Delete
6.51 kB
"""Data and evaluation contracts shared by the independently written runners."""
import argparse
import hashlib
import json
import os
from pathlib import Path
import platform
import random
import sys
import numpy as np
from scipy.io import loadmat, whosmat
import torch
import yaml
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT))
from model.MoNo import MoNoLight
def config_args(description):
parser = argparse.ArgumentParser(description=description)
parser.add_argument('--config', default=str(ROOT / 'config/config.yaml'))
return parser
def read_config(path):
with open(path) as f:
cfg = yaml.safe_load(f)
assert (cfg['data']['train_samples'], cfg['data']['test_samples']) == (1000, 200)
assert cfg['data']['stride'] == 5 and cfg['data']['grid_size'] == 85
t = cfg['training']
assert (t['epochs'], t['batch_size'], t['warmup_epochs'], t['learning_rate']) == (500, 4, 100, .001)
return cfg
def path(value):
p = Path(value)
return p if p.is_absolute() else ROOT / p
def atomic_json(destination, obj):
destination = Path(destination)
destination.parent.mkdir(parents=True, exist_ok=True)
tmp = destination.with_suffix(destination.suffix + '.tmp')
tmp.write_text(json.dumps(obj, indent=2, ensure_ascii=False, allow_nan=False))
os.replace(tmp, destination)
def atomic_save(destination, obj):
destination = Path(destination)
destination.parent.mkdir(parents=True, exist_ok=True)
tmp = destination.with_suffix(destination.suffix + '.tmp')
torch.save(obj, tmp)
os.replace(tmp, destination)
def seed_all(seed):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
def require_compute_node():
assert torch.cuda.is_available(), 'Assigned GPU/DCU unavailable; no CPU fallback'
torch.set_num_threads(int(os.environ.get('OMP_NUM_THREADS', '8')))
return torch.device('cuda')
def environment():
return dict(python=sys.version, torch=torch.__version__, hip=torch.version.hip,
cuda=torch.version.cuda, device=torch.cuda.get_device_name(),
device_count=torch.cuda.device_count(), platform=platform.platform(),
hostname=platform.node(), slurm_job_id=os.environ.get('SLURM_JOB_ID'))
def source_signature(cfg):
d = cfg['data']
return [dict(file=str(Path(d['directory']) / d[k]),
size=(Path(d['directory']) / d[k]).stat().st_size,
mtime_ns=(Path(d['directory']) / d[k]).stat().st_mtime_ns)
for k in ['train_file', 'test_file']]
def load_data(cfg):
d = cfg['data']
cache = path(d['cache'])
signature = source_signature(cfg)
if cache.exists():
with np.load(cache) as f:
assert json.loads(str(f['signature'])) == signature, 'Data source changed'
arrays = {k: f[k].copy() for k in ['train_coeff', 'train_sol', 'test_coeff', 'test_sol']}
else:
arrays = {}
schema = {}
for split, count in [('train', 1000), ('test', 200)]:
src = Path(d['directory']) / d[split + '_file']
schema[split] = whosmat(src)
print(f'DATA reading {src}, first {count} samples; stride=5', flush=True)
mat = loadmat(src, variable_names=['coeff', 'sol'])
for field in ['coeff', 'sol']:
assert mat[field].shape == (1024, 421, 421), (src, field, mat[field].shape)
a = np.ascontiguousarray(mat[field][:count, ::5, ::5], dtype=np.float32)
assert a.shape == (count, 85, 85) and np.isfinite(a).all()
arrays[split + '_' + field] = a
del mat
cache.parent.mkdir(parents=True, exist_ok=True)
tmp = cache.with_suffix('.tmp.npz')
np.savez(tmp, **arrays, signature=json.dumps(signature))
os.replace(tmp, cache)
atomic_json(path(cfg['results']) / 'data_manifest.json', dict(
sources=signature, schema=schema, train_samples=1000, test_samples=200,
split='file1[0:1000], file2[0:200]', downsample='[:,::5,::5]',
grid=[85,85], coeff='permeability', sol='pressure', units='stored numerical scale; unspecified'))
lin = np.linspace(0, 1, 85, dtype=np.float32)
grid = np.stack(np.meshgrid(lin, lin, indexing='xy'), -1).reshape(-1, 2)
coordinates = torch.from_numpy(grid).unsqueeze(0)
out = {'coordinates_raw': coordinates}
for split in ['train', 'test']:
coeff = torch.from_numpy(arrays[split + '_coeff']).reshape(-1, 7225, 1)
sol = torch.from_numpy(arrays[split + '_sol']).reshape(-1, 7225, 1)
out[split + '_coeff'] = coeff
out[split + '_sol'] = sol
out[split + '_features'] = torch.cat([coordinates.expand(coeff.shape[0], -1, -1), coeff], -1)
return out
def compute_stats(data):
x, y = data['train_features'], data['train_sol']
return dict(x_mean=x.mean((0,1), keepdim=True),
x_std=x.std((0,1), keepdim=True) + 1e-8,
y_mean=y.mean((0,1), keepdim=True),
y_std=y.std((0,1), keepdim=True) + 1e-8)
def normalized(data, stats, device):
s = {k: v.to(device) for k, v in stats.items()}
out = {'coordinates': ((data['coordinates_raw'].to(device) - s['x_mean'][..., :2]) /
s['x_std'][..., :2])}
for split in ['train', 'test']:
out[split + '_x'] = (data[split + '_features'].to(device) - s['x_mean']) / s['x_std']
out[split + '_y'] = data[split + '_sol'].to(device)
return out, s
def relative_l2(prediction, target):
return (prediction - target).flatten(1).norm(dim=1) / target.flatten(1).norm(dim=1)
@torch.no_grad()
def evaluate(model, data, stats, batch_size=4, return_predictions=False):
model.eval()
errors, predictions = [], []
for i in range(0, data['test_x'].shape[0], batch_size):
y = model(data['coordinates'], data['test_x'][i:i+batch_size])
y = y * stats['y_std'] + stats['y_mean']
errors.append(relative_l2(y, data['test_y'][i:i+batch_size]).cpu())
if return_predictions:
predictions.append(y.cpu())
return torch.cat(errors), (torch.cat(predictions) if return_predictions else None)
def code_hashes():
return {str(f.relative_to(ROOT)): hashlib.sha256(f.read_bytes()).hexdigest()
for f in [ROOT/'model/MoNo.py', ROOT/'config/config.yaml',
ROOT/'scripts/train.py', ROOT/'scripts/common.py']}