================================================================================ UNIVERSAL SVAE DIAGNOSTIC BATTERY ================================================================================ v18_johanna_curriculum/checkpoints/epoch(…): 100%  203M/203M [00:05<00:00, 110MB/s] Loading: /root/.cache/huggingface/hub/models--AbstractPhil--geolip-SVAE/snapshots/83f2d093ce03997df16e8e4c721d37968d43285c/v18_johanna_curriculum/checkpoints/epoch_0300.pt Epoch: 300, MSE: 0.237272 Config: {'V': 256, 'D': 16, 'patch_size': 16, 'hidden': 768, 'depth': 4, 'n_cross_layers': 2, 'target_cv': 0.125, 'dataset': 'curriculum_noise', 'img_size': 64, 'lr': 0.0003} Params: 16,942,419 Resolution: 64×64, samples_per_test: 64 ================================================================================ IMAGE DATASET BATTERY (64×64, n=64) ================================================================================ dataset MSE std min max | S0 SD ratio erank ---------------------------------------------------------------------------------------------------- CIFAR-10→64 0.001435 0.003511 0.000367 0.028044 | 4.297 1.840 2.34 15.76 MNIST→64 0.001426 0.000384 0.000948 0.002757 | 4.297 1.837 2.34 15.76 TinyImageNet→64 0.004736 0.006803 0.000353 0.048924 | 4.297 1.844 2.33 15.76 ImageNet-128→64 0.002932 0.008305 0.000511 0.066514 | 4.296 1.845 2.33 15.76 Resolving data files: 100%  40/40 [00:00<00:00, 14651.31it/s] Resolving data files: 100%  40/40 [00:00<00:00, 15094.21it/s] ImageNet-256→64 0.003374 0.009390 0.000527 0.074961 | 4.296 1.845 2.33 15.76 ================================================================================ NOISE TYPE BATTERY (64×64, n=64) ================================================================================ type MSE std | S0 SD ratio erank | byte_acc ±1_acc ---------------------------------------------------------------------------------------------------- gaussian 0.288616 0.007752 | 4.298 1.839 2.34 15.76 | 2.32% 6.98% uniform 0.096594 0.002739 | 4.297 1.836 2.34 15.76 | 3.95% 11.84% uniform_scaled 0.385065 0.007132 | 4.297 1.840 2.34 15.76 | 1.98% 5.95% poisson 0.056313 0.059723 | 4.297 1.837 2.34 15.76 | 6.76% 20.14% pink 0.001325 0.001499 | 4.331 2.216 1.95 15.80 | 73.38% 90.99% brown 0.001068 0.001300 | 4.333 2.215 1.96 15.80 | 72.69% 92.38% salt_pepper 0.726683 0.028651 | 4.296 1.870 2.30 15.76 | 2.00% 6.00% sparse 0.189028 0.008378 | 4.297 1.837 2.34 15.76 | 4.08% 12.20% block 0.011099 0.013453 | 4.298 1.842 2.33 15.76 | 25.02% 59.09% gradient 0.073688 0.001773 | 4.296 1.842 2.33 15.76 | 4.63% 13.78% checkerboard 0.026502 0.000645 | 4.298 1.839 2.34 15.76 | 7.74% 22.84% mixed 0.105270 0.036889 | 4.297 1.837 2.34 15.76 | 4.02% 11.96% structural 0.175447 0.006575 | 4.307 1.949 2.21 15.77 | 21.28% 27.41% cauchy 0.842779 0.020405 | 4.297 1.846 2.33 15.76 | 1.34% 4.00% exponential 0.267518 0.007822 | 4.298 1.838 2.34 15.76 | 2.58% 7.71% laplace 0.542029 0.013302 | 4.298 1.841 2.33 15.76 | 1.72% 5.21% ================================================================================ TEXT BYTE RECONSTRUCTION (64×64) ================================================================================ In: 'Hello, world! This is a test of the geometric encoder.' Out: 'EagbZ,)]hjaP$"LGM`7T\0B/\cg[5WY TdT0S^ec`jh`O2YOJib_\'' MSE: 0.001095 Byte: 3.7% In: 'The quick brown fox jumps over the lazy dog. 0123456789' Out: 'IcS2\ib\S1Pfkl_!Mia3XigfZ4YkbZ:FKU4Uala2RdU+&-#"...10%' MSE: 0.001085 Byte: 1.8% In: 'import torch; model = PatchSVAE(); output = model(x)' Out: 'Ynkgi[5[gh]U6,]QGbX)+%BXf`[NI<;,+Zmnjj[*)+ZgeJB?W'' MSE: 0.001068 Byte: 0.0% In: 'E = mc² — Albert Einstein, theoretical physicist, 1905' Out: '2)-*Wl��O��t/?>#:GG>0^RHbf[1B@?D9+R\.9/=OB>=#$%)& 6B8%-.'' MSE: 0.001033 Byte: 1.9% ================================================================================ PIECEMEAL 256→64 TILED RECONSTRUCTION ================================================================================ gaussian : 16 tiles, MSE=0.288182 uniform : 16 tiles, MSE=0.096244 pink : 16 tiles, MSE=0.000639 salt_pepper : 16 tiles, MSE=0.725523 cauchy : 16 tiles, MSE=0.844530 ================================================================================ SIGNAL ENERGY SURVIVAL ================================================================================ source survival SNR_dB orig_E recon_E ---------------------------------------------------------------------- CIFAR-10→64 100.4% 28.1dB 0.9352 0.9392 MNIST→64 100.2% 27.3dB 0.7701 0.7720 TinyImageNet→64 99.4% 22.7dB 0.8920 0.8866 ImageNet-128→64 100.4% 26.6dB 1.3260 1.3309 Resolving data files: 100%  40/40 [00:00<00:00, 13208.33it/s] Resolving data files: 100%  40/40 [00:00<00:00, 14013.71it/s] ImageNet-256→64 100.3% 26.0dB 1.3343 1.3389 noise/gaussian 69.4% 5.4dB 0.9998 0.6940 noise/pink 99.9% 31.4dB 2.2975 2.2953 noise/salt_pepper 89.4% 7.4dB 4.0102 3.5838 noise/cauchy 69.3% 5.5dB 2.9580 2.0497 ================================================================================ ALPHA PROFILE ================================================================================ Layer 0: mean=0.05629 max=0.06007 min=0.05116 std=0.002453 α[ 0]: 0.06007 █████████████████████████████████████████████████ α[ 1]: 0.05917 █████████████████████████████████████████████████ α[ 2]: 0.05941 █████████████████████████████████████████████████ α[ 3]: 0.05935 █████████████████████████████████████████████████ α[ 4]: 0.05680 ███████████████████████████████████████████████ α[ 5]: 0.05789 ████████████████████████████████████████████████ α[ 6]: 0.05694 ███████████████████████████████████████████████ α[ 7]: 0.05661 ███████████████████████████████████████████████ α[ 8]: 0.05465 █████████████████████████████████████████████ α[ 9]: 0.05605 ██████████████████████████████████████████████ α[10]: 0.05461 █████████████████████████████████████████████ α[11]: 0.05472 █████████████████████████████████████████████ α[12]: 0.05488 █████████████████████████████████████████████ α[13]: 0.05434 █████████████████████████████████████████████ α[14]: 0.05403 ████████████████████████████████████████████ α[15]: 0.05116 ██████████████████████████████████████████ Layer 1: mean=0.05758 max=0.06146 min=0.05559 std=0.001789 α[ 0]: 0.05910 ████████████████████████████████████████████████ α[ 1]: 0.05979 ████████████████████████████████████████████████ α[ 2]: 0.05954 ████████████████████████████████████████████████ α[ 3]: 0.05954 ████████████████████████████████████████████████ α[ 4]: 0.06146 █████████████████████████████████████████████████ α[ 5]: 0.05700 ██████████████████████████████████████████████ α[ 6]: 0.05784 ███████████████████████████████████████████████ α[ 7]: 0.05729 ██████████████████████████████████████████████ α[ 8]: 0.05714 ██████████████████████████████████████████████ α[ 9]: 0.05561 █████████████████████████████████████████████ α[10]: 0.05702 ██████████████████████████████████████████████ α[11]: 0.05618 █████████████████████████████████████████████ α[12]: 0.05568 █████████████████████████████████████████████ α[13]: 0.05612 █████████████████████████████████████████████ α[14]: 0.05635 █████████████████████████████████████████████ α[15]: 0.05559 █████████████████████████████████████████████ ================================================================================ COMPRESSION METRICS ================================================================================ Input: 64×64×3 = 12,288 values Latent: 16×16 = 256 omega tokens Ratio: 48.0:1 8-bit: input=12.0KB latent=0.2KB ratio=48.0:1 16-bit: input=24.0KB latent=0.5KB ratio=48.0:1 32-bit: input=48.0KB latent=1.0KB ratio=48.0:1 ================================================================================ RECONSTRUCTION GRID ================================================================================ Resolving data files: 100%  40/40 [00:00<00:00, 13156.54it/s] Resolving data files: 100%  40/40 [00:00<00:00, 14544.62it/s] Saved: universal_diagnostic_grid.png Results: diagnostic_v18_johanna_curriculum_checkpoints_epoch_0300.json ================================================================================ DIAGNOSTIC COMPLETE ================================================================================