"""Independently reload epoch-500 weights; evaluate all 200 samples at raw scale.""" import hashlib import importlib.metadata from copy import deepcopy import numpy as np import torch from common import (ROOT, MoNoLight, atomic_json, config_args, environment, evaluate, load_data, normalized, path, read_config, require_compute_node) def main(): args = config_args(__doc__).parse_args() cfg = read_config(args.config) device = require_compute_node() checkpoint = path(cfg['training']['checkpoint']) state = torch.load(checkpoint, map_location='cpu', weights_only=False) assert state['epoch'] == 500 and state['global_step'] == 125000 trained_config = deepcopy(state['config']) trained_config['data']['directory'] = cfg['data']['directory'] assert trained_config == cfg model = MoNoLight(**cfg['model']).to(device) model.load_state_dict(state['model'], strict=True) raw = load_data(cfg) data, stats = normalized(raw, state['normalization'], device) errors, predictions = evaluate(model, data, stats, return_predictions=True) assert predictions.shape == (200, 7225, 1) # Independent double precision NumPy recomputation of the requested metric. pred = predictions.numpy().reshape(200,85,85) target = raw['test_sol'].numpy().reshape(200,85,85) coeff = raw['test_coeff'].numpy().reshape(200,85,85) per_sample = np.linalg.norm((pred.astype(np.float64)-target).reshape(200,-1),axis=1) / np.linalg.norm(target.astype(np.float64).reshape(200,-1),axis=1) np.testing.assert_allclose(errors.numpy(), per_sample, rtol=2e-5, atol=1e-8) result_dir = path(cfg['results']) np.savez(result_dir/'predictions.npz', prediction=pred, target=target, coeff=coeff, coordinates=raw['coordinates_raw'].numpy().reshape(85,85,2), relative_l2=per_sample, sample_id=np.arange(200)) np.savetxt(result_dir/'per_sample_relative_l2.csv', np.column_stack((np.arange(200),per_sample)), delimiter=',', header='sample_id,relative_l2_raw', comments='', fmt=['%d','%.12g']) mean = float(per_sample.mean()) reference = cfg['paper_relative_l2'] metrics = dict(mean_relative_l2_raw=mean, test_samples=200, actual_training_epochs=state['epoch'], optimizer_updates=state['global_step'], paper_table3=reference, signed_gap=mean-reference, absolute_gap=abs(mean-reference), relative_gap_percent=100*(mean/reference-1), selection=state['selection'], checkpoint=str(checkpoint), checkpoint_sha256=hashlib.sha256(checkpoint.read_bytes()).hexdigest(), min=float(per_sample.min()), max=float(per_sample.max()), median=float(np.median(per_sample)), training_environment=state['environment'], inference_environment=environment(), software_versions={name:importlib.metadata.version(name) for name in ['torch','triton','onescience','numpy','scipy','PyYAML','matplotlib']}, configured_cluster='SCNet compute node for the original training', gpu_total_memory_bytes=torch.cuda.get_device_properties(device).total_memory, limitations=[ 'Single DCU/PyTorch 2.4.1 HIP backend versus paper RTX4090/PyTorch2.8 CUDA12.8.', 'Configured system RAM allocation is 30G, versus up to 600GB reported by the paper; all 1000/200 samples and model dimensions retained.', 'Installed Triton is 3.0.0 with DTK customizations, versus paper Triton3.4.0; author fused kernels are not used.', 'Independent implementation, packed QKV/SDPA attention, and RNG consumption differ from explicit author attention; not bitwise reproduction.', 'Portable PyTorch log-domain Sinkhorn instead of author fused Triton kernels; same finite-step equations and differentiation.', 'Identical Darcy geometry is computed once and broadcast over batch; arithmetic ordering differs.', 'Paper does not specify seed uncertainty or a tolerance for the Table3 result.', 'Final epoch checkpoint named best_model.pth; no test-set model selection.', 'Units unspecified by supplied data; metric uses stored pressure scale.']) atomic_json(result_dir/'metrics.json', metrics) print('INFERENCE', metrics, flush=True) if __name__ == '__main__': main()