MoNo-light / scripts /result.py
yushuang88's picture
Upload folder using huggingface_hub
db32cff verified
Raw History Blame Contribute Delete
5.56 kB
"""Render the fields, errors, learning curves, and audited reproduction report."""
import json
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import numpy as np
from common import config_args, path, read_config
def main():
args = config_args(__doc__).parse_args()
cfg = read_config(args.config)
results = path(cfg['results'])
metrics = json.loads((results/'metrics.json').read_text())
with np.load(results/'predictions.npz') as data:
assert data['prediction'].shape == data['target'].shape == (200,85,85)
fig, axs = plt.subplots(4,4,figsize=(14,12),constrained_layout=True)
for i in range(4):
gt, pred = data['target'][i], data['prediction'][i]
fields = [data['coeff'][i], gt, pred, np.abs(pred-gt)]
for j, (field, title) in enumerate(zip(fields,['Permeability','True pressure','Predicted pressure','Absolute error'])):
kw = dict(vmin=min(gt.min(),pred.min()),vmax=max(gt.max(),pred.max())) if j in [1,2] else {}
im=axs[i,j].imshow(field,origin='lower',extent=(0,1,0,1),cmap='magma' if j==3 else 'viridis',**kw)
axs[i,j].set_title(f'Sample {i}: {title}')
axs[i,j].set_xlabel('x'); axs[i,j].set_ylabel('y')
fig.colorbar(im,ax=axs[i,j],shrink=.8)
fig.savefig(results/'pressure_fields.png',dpi=180,bbox_inches='tight')
plt.close(fig)
fig,ax=plt.subplots(figsize=(7,4),constrained_layout=True)
ax.hist(data['relative_l2'],bins=25,color='#286ba5')
ax.axvline(metrics['mean_relative_l2_raw'],label='Reproduction mean',color='#d44b28')
ax.axvline(.0070,label='Paper Table 3',color='black',linestyle='--')
ax.set(xlabel='Raw-scale relative L2 per test sample',ylabel='Count')
ax.legend(); fig.savefig(results/'test_error_distribution.png',dpi=180,bbox_inches='tight'); plt.close(fig)
h=np.genfromtxt(results/'history.csv',delimiter=',',names=True)
assert len(h)==500 and np.array_equal(h['epoch'],np.arange(1,501))
fig,axs=plt.subplots(1,2,figsize=(11,4),constrained_layout=True)
axs[0].semilogy(h['epoch'],h['train_relative_l2'],label='Training')
axs[0].semilogy(h['epoch'],h['test_relative_l2'],label='Test monitoring')
axs[0].axhline(.007,linestyle='--',color='black',label='Paper Table 3')
axs[0].set(xlabel='Epoch',ylabel='Raw-scale relative L2'); axs[0].legend()
axs[1].plot(h['epoch'],h['learning_rate']); axs[1].set(xlabel='Epoch',ylabel='Learning rate')
fig.savefig(results/'learning_curves.png',dpi=180,bbox_inches='tight'); plt.close(fig)
limitations='\n'.join('- '+x for x in metrics['limitations'])
report=f'''# MoNo-light Darcy 85×85 reproduction
Paper: https://arxiv.org/pdf/2608.09764v1, Table 3 and Appendix F.
## Measured result
- Raw-scale per-sample relative L2 mean over all 200 test samples: **{metrics['mean_relative_l2_raw']:.10f}**.
- Paper Table 3: 0.0070; signed difference: **{metrics['signed_gap']:+.10f}**; absolute difference: {metrics['absolute_gap']:.10f}; relative difference: {metrics['relative_gap_percent']:+.4f}%.
- Actual completed training: **500 epochs / 125000 optimizer updates**, batch size 4, max LR .001, 100 warmup epochs.
- Training elapsed time: {float(h['elapsed_seconds'][-1])/3600:.4f} hours.
- Weight choice: **final epoch 500**, saved as `weight/best_model.pth`. Test metrics never select or tune weights. No validation split or early stopping.
- Checkpoint SHA256: `{metrics['checkpoint_sha256']}`.
- Parameters: {metrics['training_environment']['trainable_parameters']:,} trainable; {metrics['training_environment']['all_parameters']:,} including the frozen anchor copy (about 1.73M total).
- Actual package versions: `{metrics['software_versions']}`. Slurm resources: `{metrics['configured_cluster']}`.
## Data and metric
Reuse supplied `piececonst_r421_N1024_smooth1.mat[0:1000]` and `piececonst_r421_N1024_smooth2.mat[0:200]`, fields coeff/sol. Slice `[:,::5,::5]` to 85×85. Global per-channel unbiased train statistics only; reuse for test. Normalize input coordinates/coeff and pressure; inverse-transform pressure before loss and evaluation. Metric is arithmetic mean of each sample's norm(pred-target)/norm(target), recomputed independently with float64 NumPy after checkpoint reload.
## Alignment and limitations
{limitations}
Implementation sources and paper-versus-supplement distinction are preserved in `.trainer_work/darcy85/source_capture.md`. Official code was consulted read-only for paper-unstated details, including a resume audit; no repository downloaded, cloned, copied, or imported. This session reused the saved paper and training specifications.
## Artifacts
- `predictions.npz`: all 200 raw-scale predictions, targets, permeability, coordinates and errors.
- `per_sample_relative_l2.csv`, `metrics.json`, `history.csv`, `data_manifest.json`, `training_environment.json`.
- `pressure_fields.png`: fixed test samples 0--3, common pressure color scale within each sample.
- `test_error_distribution.png`, `learning_curves.png`.
- Runtime stdout/stderr and Slurm evidence: `../logs/` and `../.onescience/`.
'''
report_path=results/'reproduction_report.md'
if report_path.exists():
previous=report_path.read_text()
if '## Execution provenance' in previous:
report+='\n## Execution provenance'+previous.split('## Execution provenance',1)[1]
report_path.write_text(report)
print(report,flush=True)
if __name__ == '__main__':
main()