Evaluate on dataset v1.1

Browse files

Files changed (2) hide show

README.md +19 -17
eval/metrics.json +152 -151

README.md CHANGED Viewed

@@ -33,13 +33,13 @@ model-index:
       split: test
     metrics:
     - type: psnr
-      value: 21.28
       name: PSNR (dB)
     - type: ssim
-      value: 0.5694
       name: SSIM
     - type: lpips
-      value: 0.1879
       name: LPIPS (AlexNet)
 ---
 ![PSRGAN grid](psrgan_grid.jpg)
@@ -51,9 +51,9 @@ model-index:
 # Key Features
 1. Every output comes with an uncertainty map that combines aleatoric uncertainty (missing information in the input) and epistemic uncertainty (uncertainty of the model itself, from SWAG weight samples).
-2. The uncertainty tracks the error. Pixel-wise uncertainty and absolute error correlate with Pearson r = 0.59 (Spearman ρ = 0.70), and the 10% most uncertain pixels have 14.3× the error of the 10% most confident ones.
-3. On 1,056 test patches from unseen regions, PSRGAN reaches an LPIPS of 0.188, against 0.348 for Real-ESRGAN x4plus and 0.599 for bicubic upsampling. It also beats Real-ESRGAN in PSNR (21.28 vs. 20.22 dB) and SSIM (0.569 vs. 0.475).
-4. Two sets of weights. The final adversarial weights have the lowest LPIPS, the weights after the likelihood stage the highest PSNR (21.94 dB) and SSIM (0.598) of all four methods tested.
 5. Trained on randomized degradations (Gaussian, motion or defocus blur, then bicubic, bilinear or area downsampling), so it learns actual super-resolution instead of inverting one fixed downscaling operation.
 # How the uncertainty map is made
@@ -113,16 +113,16 @@ The model expects Sentinel-2 L2A RGB (B04, B03, B02) after a per-image 1 to 99%
 # Evaluation
-All numbers are on the test split of [PSRGAN Sentinel-2](https://huggingface.co/datasets/marklpd/PSRGAN-Sentinel-2) (1,056 patches), whose tiles lie in 0.5° blocks that share no pixel with training or validation. PSNR, SSIM and LPIPS are averaged per image (± is the 95% confidence interval), VGG is the perceptual loss used in training and OT the patchwise Sinkhorn distance between pixel color distributions.
 | method | PSNR (dB) ↑ | SSIM ↑ | LPIPS ↓ | VGG ↓ | OT ↓ |
 |---|---:|---:|---:|---:|---:|
-| Bicubic | 21.56 ± 0.17 | 0.5428 ± 0.0077 | 0.5987 ± 0.0050 | 8.135 | 0.00321 |
-| Real-ESRGAN | 20.22 ± 0.15 | 0.4753 ± 0.0068 | 0.3482 ± 0.0043 | 8.640 | 0.00399 |
-| PSRGAN (PSNR) | **21.94 ± 0.17** | **0.5981 ± 0.0063** | 0.2082 ± 0.0027 | **4.877** | 0.00267 |
-| **PSRGAN** | 21.28 ± 0.17 | 0.5694 ± 0.0063 | **0.1879 ± 0.0026** | 5.859 | **0.00135** |
-PSRGAN beats Real-ESRGAN in PSNR on 97% of test patches and in LPIPS on 100%. Compared with the likelihood stage, the adversarial stage gives up 0.66 dB PSNR and changes LPIPS from 0.208 to 0.188 (lower is better). Bicubic upsampling keeps a higher PSNR than the final model on 70% of patches, because PSNR rewards smooth averages over sharp but slightly misplaced detail. LPIPS, which follows perceived similarity, is lower for PSRGAN than for bicubic on 100% of patches.
 On the 128 test patches from Herzogenaurach, Wolfsburg and München:
@@ -142,10 +142,10 @@ On the 128 test patches from Herzogenaurach, Wolfsburg and München:
 | | ECE (σ vs. RMSE) | error within 1σ | error within 2σ |
 |---|---:|---:|---:|
-| Aleatoric + SWAG | 0.0240 | 86.6% | 98.4% |
-| Aleatoric + Bayesian heads | 0.0289 | 88.4% | 98.7% |
-A perfectly calibrated Gaussian would cover 68.3% and 95.4%. Both variants cover more, and the predicted σ is larger than the observed RMSE in every calibration bin, so the uncertainty is conservative. The uncertainty-error correlation is stronger in flat areas (Pearson r = 0.59) than on edges (r = 0.37).
 # Training
@@ -156,12 +156,14 @@ PSRGAN starts from the Real-ESRGAN x4plus generator (RRDBNet, 23 blocks) and its
 Augmentation (flips, 90° rotations) is applied to the training split only, and the test split is only used for the final evaluation. Training ran on one NVIDIA A100-SXM4-40GB in Google Colab.
 # Limitations
 - The model is trained on synthetically degraded Sentinel-2 images. It sharpens Sentinel-2 data at its own resolution and does not recover detail from a finer sensor. Details beyond what the input contains are estimates, not measurements.
 - Its training data are 98.6% Alpine tiles (Switzerland, Austria, northern Italy, south-eastern France, southern Germany, Slovenia) from late September and early October 2025. Expect worse results on other climates, other seasons, deserts and especially dense city centres.
-- The uncertainty maps rank pixels by expected error, but the absolute σ is too large: 87% of pixel errors fall within 1σ instead of 68%. Treat σ as a conservative bound, or recalibrate it on your own data.
-- The final model trades pixel accuracy for sharpness. Its PSNR is 0.29 dB below bicubic upsampling, so use `psrgan_x4_psnr.safetensors` when pixel accuracy matters more than texture.
 - The training pixels are not square (about 10 m east-west, 15.6 m north-south), because tiles were requested on a square WGS84 grid.
 # Out-of-Scope Use

       split: test
     metrics:
     - type: psnr
+      value: 21.22
       name: PSNR (dB)
     - type: ssim
+      value: 0.5639
       name: SSIM
     - type: lpips
+      value: 0.1889
       name: LPIPS (AlexNet)
 ---
 ![PSRGAN grid](psrgan_grid.jpg)
 # Key Features
 1. Every output comes with an uncertainty map that combines aleatoric uncertainty (missing information in the input) and epistemic uncertainty (uncertainty of the model itself, from SWAG weight samples).
+2. The uncertainty tracks the error. Pixel-wise uncertainty and absolute error correlate with Pearson r = 0.57 (Spearman ρ = 0.66), and the 10% most uncertain pixels have 12.8× the error of the 10% most confident ones.
+3. On 1,328 test patches from unseen regions, PSRGAN reaches an LPIPS of 0.189, against 0.353 for Real-ESRGAN x4plus and 0.604 for bicubic upsampling. It also beats Real-ESRGAN in PSNR (21.22 vs. 20.20 dB) and SSIM (0.564 vs. 0.469).
+4. Two sets of weights. The final adversarial weights have the lowest LPIPS, the weights after the likelihood stage the highest PSNR (21.87 dB) and SSIM (0.593) of all four methods tested.
 5. Trained on randomized degradations (Gaussian, motion or defocus blur, then bicubic, bilinear or area downsampling), so it learns actual super-resolution instead of inverting one fixed downscaling operation.
 # How the uncertainty map is made
 # Evaluation
+All numbers are on the test split of [PSRGAN Sentinel-2](https://huggingface.co/datasets/marklpd/PSRGAN-Sentinel-2) version 1.1 (1,328 patches), whose tiles lie in 0.5° blocks that share no pixel with training or validation. PSNR, SSIM and LPIPS are averaged per image (± is the 95% confidence interval), VGG is the perceptual loss used in training and OT the patchwise Sinkhorn distance between pixel color distributions.
 | method | PSNR (dB) ↑ | SSIM ↑ | LPIPS ↓ | VGG ↓ | OT ↓ |
 |---|---:|---:|---:|---:|---:|
+| Bicubic | 21.50 ± 0.14 | 0.5364 ± 0.0064 | 0.6042 ± 0.0040 | 8.274 | 0.00323 |
+| Real-ESRGAN | 20.20 ± 0.13 | 0.4692 ± 0.0057 | 0.3529 ± 0.0035 | 8.760 | 0.00399 |
+| PSRGAN (PSNR) | **21.87 ± 0.15** | **0.5927 ± 0.0054** | 0.2108 ± 0.0023 | **4.943** | 0.00275 |
+| **PSRGAN** | 21.22 ± 0.14 | 0.5639 ± 0.0053 | **0.1889 ± 0.0021** | 5.925 | **0.00138** |
+PSRGAN beats Real-ESRGAN in PSNR on 98% of test patches and in LPIPS on 100%. Compared with the likelihood stage, the adversarial stage gives up 0.65 dB PSNR and changes LPIPS from 0.211 to 0.189 (lower is better). Bicubic upsampling keeps a higher PSNR than the final model on 70% of patches, because PSNR rewards smooth averages over sharp but slightly misplaced detail. LPIPS, which follows perceived similarity, is lower for PSRGAN than for bicubic on 100% of patches.
 On the 128 test patches from Herzogenaurach, Wolfsburg and München:
 | | ECE (σ vs. RMSE) | error within 1σ | error within 2σ |
 |---|---:|---:|---:|
+| Aleatoric + SWAG | 0.0255 | 84.7% | 98.1% |
+| Aleatoric + Bayesian heads | 0.0313 | 86.6% | 98.4% |
+A perfectly calibrated Gaussian would cover 68.3% and 95.4%. Both variants cover more, and the predicted σ is larger than the observed RMSE in every calibration bin, so the uncertainty is conservative. The uncertainty-error correlation is stronger in flat areas (Pearson r = 0.57) than on edges (r = 0.39).
 # Training
 Augmentation (flips, 90° rotations) is applied to the training split only, and the test split is only used for the final evaluation. Training ran on one NVIDIA A100-SXM4-40GB in Google Colab.
+The model was trained on version 1.0 of the dataset (8,171 training and 1,008 validation patches). 1,728 of those training patches come from 108 tiles that turned out to be pixel copies of 54 other training tiles, so these places counted three times. Version 1.1 removes the copies and adds new tiles. Its new training and validation patches were not used, and the evaluation above uses its test split.
 # Limitations
 - The model is trained on synthetically degraded Sentinel-2 images. It sharpens Sentinel-2 data at its own resolution and does not recover detail from a finer sensor. Details beyond what the input contains are estimates, not measurements.
 - Its training data are 98.6% Alpine tiles (Switzerland, Austria, northern Italy, south-eastern France, southern Germany, Slovenia) from late September and early October 2025. Expect worse results on other climates, other seasons, deserts and especially dense city centres.
+- The uncertainty maps rank pixels by expected error, but the absolute σ is too large: 85% of pixel errors fall within 1σ instead of 68%. Treat σ as a conservative bound, or recalibrate it on your own data.
+- The final model trades pixel accuracy for sharpness. Its PSNR is 0.28 dB below bicubic upsampling, so use `psrgan_x4_psnr.safetensors` when pixel accuracy matters more than texture.
 - The training pixels are not square (about 10 m east-west, 15.6 m north-south), because tiles were requested on a square WGS84 grid.
 # Out-of-Scope Use

eval/metrics.json CHANGED Viewed

@@ -1,63 +1,63 @@
 {
   "test": {
-    "n_patches": 1056,
     "results": {
       "Bicubic": {
-        "PSNR": 21.5607197952993,
-        "PSNR \u00b1": 0.1711068225145145,
-        "SSIM": 0.5427891769813318,
-        "SSIM \u00b1": 0.007653797755572996,
-        "LPIPS": 0.5987126496479367,
-        "LPIPS \u00b1": 0.004994828619425444,
-        "VGG": 8.134999072912967,
-        "OT": 0.003209914177811394
       },
       "Real-ESRGAN": {
-        "PSNR": 20.215020929322097,
-        "PSNR \u00b1": 0.15033991375476158,
-        "SSIM": 0.47526671260482434,
-        "SSIM \u00b1": 0.006763881811721902,
-        "LPIPS": 0.34822498551202996,
-        "LPIPS \u00b1": 0.004308759976577521,
-        "VGG": 8.63969765287457,
-        "OT": 0.00399009935614286
       },
       "PSRGAN (PSNR)": {
-        "PSNR": 21.936194533651527,
-        "PSNR \u00b1": 0.17383839192396056,
-        "SSIM": 0.5981000966646455,
-        "SSIM \u00b1": 0.006328675695824681,
-        "LPIPS": 0.20817970013895043,
-        "LPIPS \u00b1": 0.0027403351497943132,
-        "VGG": 4.8766192811908144,
-        "OT": 0.002672349598442176
       },
       "PSRGAN": {
-        "PSNR": 21.275709489981335,
-        "PSNR \u00b1": 0.16942986974481397,
-        "SSIM": 0.5693713306929126,
-        "SSIM \u00b1": 0.006270623736088413,
-        "LPIPS": 0.18790546653679374,
-        "LPIPS \u00b1": 0.002612060048786746,
-        "VGG": 5.858746337168144,
-        "OT": 0.0013482042025619496
       }
     },
     "psrgan_win_rate": {
       "Bicubic": {
-        "PSNR": 0.29829545454545453,
-        "SSIM": 0.6903409090909091,
         "LPIPS": 1.0
       },
       "Real-ESRGAN": {
-        "PSNR": 0.9706439393939394,
         "SSIM": 1.0,
         "LPIPS": 1.0
       },
       "PSRGAN (PSNR)": {
-        "PSNR": 0.000946969696969697,
-        "SSIM": 0.002840909090909091,
-        "LPIPS": 0.9261363636363636
       }
     }
   },
@@ -65,156 +65,156 @@
     "n_patches": 128,
     "results": {
       "Bicubic": {
-        "PSNR": 21.145912691950798,
-        "PSNR \u00b1": 0.2489663205630983,
         "SSIM": 0.5521608623676002,
         "SSIM \u00b1": 0.020581239817082063,
-        "LPIPS": 0.6825457969680429,
-        "LPIPS \u00b1": 0.014193458656650901,
         "VGG": 9.868039011955261,
-        "OT": 0.003211126459063962
       },
       "Real-ESRGAN": {
-        "PSNR": 20.24102108180523,
-        "PSNR \u00b1": 0.22995863706651176,
-        "SSIM": 0.5146662218030542,
-        "SSIM \u00b1": 0.01896918156259728,
-        "LPIPS": 0.39876320981420577,
-        "LPIPS \u00b1": 0.014998464281860064,
-        "VGG": 9.620419383049011,
-        "OT": 0.003078749228734523
       },
       "PSRGAN (PSNR)": {
-        "PSNR": 21.299197003245354,
-        "PSNR \u00b1": 0.27709311095960687,
-        "SSIM": 0.5960347084328532,
-        "SSIM \u00b1": 0.019889227648886978,
-        "LPIPS": 0.21178198768757284,
-        "LPIPS \u00b1": 0.00864762965393086,
-        "VGG": 5.9829049706459045,
-        "OT": 0.003099578752880916
       },
       "PSRGAN": {
-        "PSNR": 20.610211864113808,
-        "PSNR \u00b1": 0.2630898367096838,
-        "SSIM": 0.5620126791764051,
-        "SSIM \u00b1": 0.018756582941474018,
-        "LPIPS": 0.19378441805019975,
-        "LPIPS \u00b1": 0.008681403045909196,
-        "VGG": 7.124138653278351,
-        "OT": 0.001463380001951009
       }
     }
   },
   "calibration": {
     "swag": {
-      "ECE": 0.02397674263321663,
-      "coverage_1sigma": 0.8664602438608805,
-      "coverage_2sigma": 0.9841360251108805,
       "reliability": {
         "pred_std": [
-          0.020123129710555077,
-          0.042232051491737366,
-          0.06536165624856949,
-          0.08905356377363205,
-          0.11278203874826431,
-          0.1365380734205246,
-          0.16020965576171875,
-          0.18388722836971283,
-          0.2074737250804901,
-          0.2310594618320465,
-          0.2546865940093994,
-          0.2783450484275818,
-          0.3019865155220032,
-          0.32563266158103943,
-          0.3493105471134186
         ],
         "rmse": [
-          0.014095807448029518,
-          0.02998480387032032,
-          0.04914991185069084,
-          0.0686262995004654,
-          0.08680041134357452,
-          0.1045115739107132,
-          0.12249297648668289,
-          0.14000065624713898,
-          0.1562192142009735,
-          0.17163224518299103,
-          0.18531100451946259,
-          0.19668403267860413,
-          0.20926719903945923,
-          0.21913963556289673,
-          0.2296137809753418
         ]
       }
     },
     "bayesian_heads": {
-      "ECE": 0.02886166655259949,
-      "coverage_1sigma": 0.8842587471008301,
-      "coverage_2sigma": 0.986927350362142,
       "reliability": {
         "pred_std": [
-          0.02026936411857605,
-          0.04286534711718559,
-          0.0662725418806076,
-          0.08985785394906998,
-          0.11369309574365616,
-          0.1376895308494568,
-          0.16160856187343597,
-          0.18553175032138824,
-          0.20935258269309998,
-          0.233138307929039,
-          0.25700125098228455,
-          0.28088292479515076,
-          0.30468234419822693,
-          0.32858404517173767,
-          0.35245150327682495
         ],
         "rmse": [
-          0.012933187186717987,
-          0.027884116396307945,
-          0.045200906693935394,
-          0.0637541264295578,
-          0.08282162249088287,
-          0.10149621963500977,
-          0.1198882907629013,
-          0.13830198347568512,
-          0.15485383570194244,
-          0.17070721089839935,
-          0.1843496859073639,
-          0.19707782566547394,
-          0.2094695121049881,
-          0.21937879920005798,
-          0.23037448525428772
         ]
       }
     }
   },
   "uncertainty_error_correlation": {
-    "pearson": 0.5869906544685364,
-    "spearman": 0.7000287473727114,
     "by_region": {
       "edges": {
-        "pearson": 0.3714364469051361,
-        "spearman": 0.37498411900628265
       },
       "flat": {
-        "pearson": 0.5925312638282776,
-        "spearman": 0.6675135785054067
       }
     }
   },
   "error_by_uncertainty_decile": [
-    0.009789826348423958,
-    0.016854647547006607,
-    0.024157894775271416,
-    0.03273877874016762,
-    0.04215969517827034,
-    0.052353471517562866,
-    0.06426148116588593,
-    0.07902297377586365,
-    0.09887537360191345,
-    0.13998572528362274
   ],
   "training": {
     "train_patches": 8171,
@@ -243,5 +243,6 @@
     },
     "gpu": "NVIDIA A100-SXM4-40GB",
     "seed": 42
-  }
 }

 {
   "test": {
+    "n_patches": 1328,
     "results": {
       "Bicubic": {
+        "PSNR": 21.49776914895299,
+        "PSNR \u00b1": 0.14387326826672933,
+        "SSIM": 0.5364394646631666,
+        "SSIM \u00b1": 0.0064225451768439145,
+        "LPIPS": 0.6041575976120062,
+        "LPIPS \u00b1": 0.003990337435301,
+        "VGG": 8.273993061249515,
+        "OT": 0.0032254697611914522
       },
       "Real-ESRGAN": {
+        "PSNR": 20.196650752102038,
+        "PSNR \u00b1": 0.1276592253712529,
+        "SSIM": 0.46924692207206803,
+        "SSIM \u00b1": 0.005693634651190977,
+        "LPIPS": 0.3528626701087657,
+        "LPIPS \u00b1": 0.003529675200247061,
+        "VGG": 8.75981970292976,
+        "OT": 0.003986208264839398
       },
       "PSRGAN (PSNR)": {
+        "PSNR": 21.871579875428992,
+        "PSNR \u00b1": 0.14741332287890474,
+        "SSIM": 0.5926794328738049,
+        "SSIM \u00b1": 0.005390192711499601,
+        "LPIPS": 0.21083843581804848,
+        "LPIPS \u00b1": 0.0022550128517949573,
+        "VGG": 4.942558713706143,
+        "OT": 0.002748513919671616
       },
       "PSRGAN": {
+        "PSNR": 21.21997445893575,
+        "PSNR \u00b1": 0.14396838880837962,
+        "SSIM": 0.5639128608920847,
+        "SSIM \u00b1": 0.005315456142309424,
+        "LPIPS": 0.18892271170415073,
+        "LPIPS \u00b1": 0.0021353417087285735,
+        "VGG": 5.92484254147633,
+        "OT": 0.0013823630232139524
       }
     },
     "psrgan_win_rate": {
       "Bicubic": {
+        "PSNR": 0.29593373493975905,
+        "SSIM": 0.7063253012048193,
         "LPIPS": 1.0
       },
       "Real-ESRGAN": {
+        "PSNR": 0.9751506024096386,
         "SSIM": 1.0,
         "LPIPS": 1.0
       },
       "PSRGAN (PSNR)": {
+        "PSNR": 0.0007530120481927711,
+        "SSIM": 0.0015060240963855422,
+        "LPIPS": 0.9427710843373494
       }
     }
   },
     "n_patches": 128,
     "results": {
       "Bicubic": {
+        "PSNR": 21.14591260254383,
+        "PSNR \u00b1": 0.2489663064062437,
         "SSIM": 0.5521608623676002,
         "SSIM \u00b1": 0.020581239817082063,
+        "LPIPS": 0.6825458155944943,
+        "LPIPS \u00b1": 0.014193455320916534,
         "VGG": 9.868039011955261,
+        "OT": 0.0032111122854985297
       },
       "Real-ESRGAN": {
+        "PSNR": 20.241013795137405,
+        "PSNR \u00b1": 0.22995885194599489,
+        "SSIM": 0.5146660783793777,
+        "SSIM \u00b1": 0.018968534919615254,
+        "LPIPS": 0.3987554812338203,
+        "LPIPS \u00b1": 0.014997665450986498,
+        "VGG": 9.62031614780426,
+        "OT": 0.003078677167650312
       },
       "PSRGAN (PSNR)": {
+        "PSNR": 21.299035504460335,
+        "PSNR \u00b1": 0.2770925555614416,
+        "SSIM": 0.5960352646652609,
+        "SSIM \u00b1": 0.019889318496348913,
+        "LPIPS": 0.21179563156329095,
+        "LPIPS \u00b1": 0.008647334268858733,
+        "VGG": 5.982898652553558,
+        "OT": 0.003099947818554938
       },
       "PSRGAN": {
+        "PSNR": 20.610222026705742,
+        "PSNR \u00b1": 0.26308919913873674,
+        "SSIM": 0.5620116831269115,
+        "SSIM \u00b1": 0.01875657800639376,
+        "LPIPS": 0.19378626625984907,
+        "LPIPS \u00b1": 0.008680669962247933,
+        "VGG": 7.124137580394745,
+        "OT": 0.0014629952056566253
       }
     }
   },
   "calibration": {
     "swag": {
+      "ECE": 0.02546852860957056,
+      "coverage_1sigma": 0.8469663461049398,
+      "coverage_2sigma": 0.9806195894877116,
       "reliability": {
         "pred_std": [
+          0.02145407721400261,
+          0.04436695948243141,
+          0.06854496151208878,
+          0.09341390430927277,
+          0.11819290369749069,
+          0.143020361661911,
+          0.16783550381660461,
+          0.19264240562915802,
+          0.21738982200622559,
+          0.24216853082180023,
+          0.2670060694217682,
+          0.2918231785297394,
+          0.3166002929210663,
+          0.34143877029418945,
+          0.3662784695625305
         ],
         "rmse": [
+          0.01695420779287815,
+          0.03484932705760002,
+          0.054482631385326385,
+          0.07356802374124527,
+          0.09181062132120132,
+          0.11039590835571289,
+          0.1300036609172821,
+          0.14939245581626892,
+          0.167398601770401,
+          0.18336381018161774,
+          0.19742253422737122,
+          0.21018892526626587,
+          0.22215959429740906,
+          0.2338602989912033,
+          0.2432245910167694
         ]
       }
     },
     "bayesian_heads": {
+      "ECE": 0.03130538757300307,
+      "coverage_1sigma": 0.8664686679840088,
+      "coverage_2sigma": 0.9843335946400961,
       "reliability": {
         "pred_std": [
+          0.021405110135674477,
+          0.04486752673983574,
+          0.0694613978266716,
+          0.09442617744207382,
+          0.11953892558813095,
+          0.14435626566410065,
+          0.16953186690807343,
+          0.1946171224117279,
+          0.21965409815311432,
+          0.24466313421726227,
+          0.26977258920669556,
+          0.29481807351112366,
+          0.31992027163505554,
+          0.34500011801719666,
+          0.37008485198020935
         ],
         "rmse": [
+          0.015972202643752098,
+          0.03202086687088013,
+          0.05116646736860275,
+          0.06973199546337128,
+          0.08717330545186996,
+          0.10517378151416779,
+          0.12632086873054504,
+          0.14633670449256897,
+          0.16502711176872253,
+          0.18130864202976227,
+          0.1961313635110855,
+          0.20941568911075592,
+          0.22148063778877258,
+          0.23357439041137695,
+          0.24296487867832184
         ]
       }
     }
   },
   "uncertainty_error_correlation": {
+    "pearson": 0.5736827254295349,
+    "spearman": 0.6573789566355632,
     "by_region": {
       "edges": {
+        "pearson": 0.3913707435131073,
+        "spearman": 0.38998334134778145
       },
       "flat": {
+        "pearson": 0.5683034062385559,
+        "spearman": 0.6238047961941732
       }
     }
   },
   "error_by_uncertainty_decile": [
+    0.011378942988812923,
+    0.020646333694458008,
+    0.028598524630069733,
+    0.036584850400686264,
+    0.04477774724364281,
+    0.05396171286702156,
+    0.06510858237743378,
+    0.07936927676200867,
+    0.0998813733458519,
+    0.1451248973608017
   ],
   "training": {
     "train_patches": 8171,
     },
     "gpu": "NVIDIA A100-SXM4-40GB",
     "seed": 42
+  },
+  "dataset_version": "1.1"
 }