| """Render Darcy coefficient/GT/prediction/signed-error panels and diagnostics.""" |
| from __future__ import annotations |
|
|
| import csv |
| import json |
| from pathlib import Path |
|
|
| import matplotlib |
| matplotlib.use('Agg') |
| import matplotlib.pyplot as plt |
| import numpy as np |
|
|
| from common import atomic_json, configuration, read_config |
|
|
|
|
| def main(): |
| args=configuration('Visualize the full HNO Darcy experiment').parse_args() |
| c=read_config(args.config) |
| output=Path(c['output']['results_dir']) |
| data=np.load(output/'predictions.npz') |
| metrics=json.loads((output/'metrics.json').read_text()) |
| with (output/'training_history.csv').open() as f: |
| history=list(csv.DictReader(f)) |
| epochs=np.asarray([int(r['epoch']) for r in history]) |
| if len(history)!=c['training']['epochs'] or list(epochs)!=list(range(1,c['training']['epochs']+1)): |
| raise ValueError('Visualization requires the complete formal training history') |
| if len(data['prediction'])!=c['data']['n_eval']: |
| raise ValueError('Incomplete inference data') |
| dpi=c['visualization']['dpi'] |
| fig,axes=plt.subplots(1,3,figsize=(14,3.8),constrained_layout=True) |
| for name,label in [('train_relative_l2','Training relative L2'),('eval_relative_l2','Evaluation relative L2')]: |
| axes[0].semilogy(epochs,[float(r[name]) for r in history],label=label,lw=1) |
| axes[0].axvline(metrics['best_epoch'],color='gray',ls=':',label='Best checkpoint') |
| axes[0].axhline(metrics['paper_relative_l2'],color='black',ls='--',label='Paper: 0.00446') |
| axes[0].set(xlabel='Epoch',ylabel='Mean relative L2');axes[0].legend(fontsize=8) |
| axes[1].semilogy(epochs,[float(r['train_loss']) for r in history],label='Total loss / sample') |
| axes[1].semilogy(epochs,[float(r['train_derivative']) for r in history],label='Derivative relative loss') |
| axes[1].set(xlabel='Epoch',ylabel='Loss');axes[1].legend(fontsize=8) |
| axes[2].plot(epochs,[float(r['learning_rate']) for r in history]);axes[2].set(xlabel='Epoch',ylabel='Learning rate') |
| fig.savefig(output/'training_curves.png',dpi=dpi);fig.savefig(output/'training_curves.pdf');plt.close(fig) |
| chosen=c['visualization']['sample_indices'] |
| fig,axes=plt.subplots(len(chosen),4,figsize=(13,3*len(chosen)),squeeze=False,constrained_layout=True) |
| titles=['Coefficient a','Ground truth u','HNO prediction','Signed error (prediction - GT)'] |
| for row,index in enumerate(chosen): |
| a,u,p=(data[k][index] for k in ['coefficient','target','prediction']) |
| residual=p-u |
| vmin=min(u.min(),p.min());vmax=max(u.max(),p.max()) |
| error_bound=max(float(np.abs(residual).max()),1e-12) |
| specifications=[(a,'viridis',None,None),(u,'viridis',vmin,vmax),(p,'viridis',vmin,vmax), |
| (residual,'RdBu_r',-error_bound,error_bound)] |
| for col,(field,cmap,lo,hi) in enumerate(specifications): |
| ax=axes[row,col] |
| im=ax.imshow(field,origin='lower',extent=(0,1,0,1),cmap=cmap,vmin=lo,vmax=hi) |
| ax.set(xlabel='x',ylabel='y',aspect='equal') |
| if row==0:ax.set_title(titles[col],fontsize=10) |
| fig.colorbar(im,ax=ax,fraction=.046,pad=.04) |
| axes[row,0].set_ylabel(f'Sample {index}\ny') |
| axes[row,2].text(.02,.98,f'Rel. L2={data["relative_l2"][index]:.5f}',transform=axes[row,2].transAxes, |
| va='top',fontsize=8,bbox=dict(facecolor='white',alpha=.8,edgecolor='none')) |
| fig.savefig(output/'darcy_predictions.png',dpi=dpi);fig.savefig(output/'darcy_predictions.pdf');plt.close(fig) |
| fig,ax=plt.subplots(figsize=(6,4),constrained_layout=True) |
| ax.hist(data['relative_l2'],bins=24,color='#35668b',edgecolor='white') |
| ax.axvline(metrics['mean_relative_l2'],label=f'Mean: {metrics["mean_relative_l2"]:.6f}',color='black') |
| ax.set(xlabel='Per-sample relative L2',ylabel='Evaluation samples',title='Darcy: 200 best-checkpoint predictions') |
| ax.legend();fig.savefig(output/'error_distribution.png',dpi=dpi);plt.close(fig) |
| |
| residual=data['prediction']-data['target'] |
| error_energy=np.abs(np.fft.fft2(residual,axes=(-2,-1)))**2 |
| target_energy=np.abs(np.fft.fft2(data['target'],axes=(-2,-1)))**2 |
| size=residual.shape[-1];freq=np.fft.fftfreq(size)*size |
| yy,xx=np.meshgrid(freq,freq,indexing='ij');bins=np.floor(np.sqrt(xx**2+yy**2)).astype(int) |
| spectrum=[] |
| for b in range(bins.max()+1): |
| select=bins==b;den=target_energy[:,select].sum() |
| spectrum.append(float(np.sqrt(error_energy[:,select].sum()/max(den,1e-30)))) |
| fig,ax=plt.subplots(figsize=(6,4),constrained_layout=True) |
| ax.semilogy(np.arange(len(spectrum)),spectrum);ax.set(xlabel='Radial frequency |k|',ylabel='Relative error spectrum') |
| fig.savefig(output/'error_spectrum.png',dpi=dpi);plt.close(fig) |
| atomic_json(output/'visualization_manifest.json',{'status':'completed','n_eval':len(data['prediction']), |
| 'epochs':len(history),'sample_indices':chosen,'dpi':dpi,'colormaps':['viridis','RdBu_r'], |
| 'figures':['training_curves.png','training_curves.pdf','darcy_predictions.png','darcy_predictions.pdf', |
| 'error_distribution.png','error_spectrum.png'],'radial_error_spectrum':spectrum}) |
| print(f'VISUALIZATION_COMPLETED results={output} epochs={len(history)} samples={len(data["prediction"])}',flush=True) |
|
|
|
|
| if __name__=='__main__': |
| main() |
|
|
