0x-Khalil commited on
Commit
7ce4d7d
·
verified ·
1 Parent(s): 7f0f555

Upload OSAE study results

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. README.md +299 -1
  2. configs/default.yaml +96 -0
  3. configs/fast.yaml +18 -0
  4. configs/kaggle.yaml +7 -0
  5. configs/kaggle_fast.yaml +14 -0
  6. configs/smoke.yaml +54 -0
  7. data_splits/splits.json +0 -0
  8. data_splits/splits_summary.md +16 -0
  9. data_splits/stats.json +13 -0
  10. downstream/C128_l0.01_clean/finetune/f0.1/s0/metrics.json +149 -0
  11. downstream/C128_l0.01_clean/finetune/f0.1/s0/test_metrics.json +170 -0
  12. downstream/C128_l0.01_clean/finetune/f0.1/s0/test_predictions.npy +3 -0
  13. downstream/C128_l0.01_clean/finetune/f0.1/s1/metrics.json +149 -0
  14. downstream/C128_l0.01_clean/finetune/f0.1/s1/test_metrics.json +170 -0
  15. downstream/C128_l0.01_clean/finetune/f0.1/s1/test_predictions.npy +3 -0
  16. downstream/C128_l0.01_clean/finetune/f0.1/s2/metrics.json +149 -0
  17. downstream/C128_l0.01_clean/finetune/f0.1/s2/test_metrics.json +170 -0
  18. downstream/C128_l0.01_clean/finetune/f0.1/s2/test_predictions.npy +3 -0
  19. downstream/C128_l0.01_clean/finetune/f0.25/s0/metrics.json +149 -0
  20. downstream/C128_l0.01_clean/finetune/f0.25/s0/test_metrics.json +170 -0
  21. downstream/C128_l0.01_clean/finetune/f0.25/s0/test_predictions.npy +3 -0
  22. downstream/C128_l0.01_clean/finetune/f0.25/s1/metrics.json +149 -0
  23. downstream/C128_l0.01_clean/finetune/f0.25/s1/test_metrics.json +170 -0
  24. downstream/C128_l0.01_clean/finetune/f0.25/s1/test_predictions.npy +3 -0
  25. downstream/C128_l0.01_clean/finetune/f0.25/s2/metrics.json +149 -0
  26. downstream/C128_l0.01_clean/finetune/f0.25/s2/test_metrics.json +170 -0
  27. downstream/C128_l0.01_clean/finetune/f0.25/s2/test_predictions.npy +3 -0
  28. downstream/C128_l0.01_clean/finetune/f0.5/s0/metrics.json +149 -0
  29. downstream/C128_l0.01_clean/finetune/f0.5/s0/test_metrics.json +170 -0
  30. downstream/C128_l0.01_clean/finetune/f0.5/s0/test_predictions.npy +3 -0
  31. downstream/C128_l0.01_clean/finetune/f0.5/s1/metrics.json +149 -0
  32. downstream/C128_l0.01_clean/finetune/f0.5/s1/test_metrics.json +170 -0
  33. downstream/C128_l0.01_clean/finetune/f0.5/s1/test_predictions.npy +3 -0
  34. downstream/C128_l0.01_clean/finetune/f0.5/s2/metrics.json +149 -0
  35. downstream/C128_l0.01_clean/finetune/f0.5/s2/test_metrics.json +170 -0
  36. downstream/C128_l0.01_clean/finetune/f0.5/s2/test_predictions.npy +3 -0
  37. downstream/C128_l0.01_clean/finetune/f1/s0/metrics.json +149 -0
  38. downstream/C128_l0.01_clean/finetune/f1/s0/test_metrics.json +170 -0
  39. downstream/C128_l0.01_clean/finetune/f1/s0/test_predictions.npy +3 -0
  40. downstream/C128_l0.01_clean/finetune/f1/s1/metrics.json +149 -0
  41. downstream/C128_l0.01_clean/finetune/f1/s1/test_metrics.json +170 -0
  42. downstream/C128_l0.01_clean/finetune/f1/s1/test_predictions.npy +3 -0
  43. downstream/C128_l0.01_clean/finetune/f1/s2/metrics.json +149 -0
  44. downstream/C128_l0.01_clean/finetune/f1/s2/test_metrics.json +170 -0
  45. downstream/C128_l0.01_clean/finetune/f1/s2/test_predictions.npy +3 -0
  46. downstream/C128_l0.01_clean/frozen/f0.1/s0/metrics.json +149 -0
  47. downstream/C128_l0.01_clean/frozen/f0.1/s0/test_metrics.json +170 -0
  48. downstream/C128_l0.01_clean/frozen/f0.1/s0/test_predictions.npy +3 -0
  49. downstream/C128_l0.01_clean/frozen/f0.1/s1/metrics.json +149 -0
  50. downstream/C128_l0.01_clean/frozen/f0.1/s1/test_metrics.json +170 -0
README.md CHANGED
@@ -1,3 +1,301 @@
1
  ---
2
- license: apache-2.0
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ tags:
3
+ - autoencoder
4
+ - sparse-coding
5
+ - representation-learning
6
+ - self-supervised
7
+ - stl10
8
+ - resnet18
9
  ---
10
+
11
+ # Overcomplete Sparse Autoencoders for Visual Representation Learning (STL-10)
12
+
13
+ Controlled study of whether an **overcomplete, L1-sparse convolutional autoencoder** learns features that transfer
14
+ better to STL-10 classification than a **conventional undercomplete autoencoder**, against a **supervised
15
+ ResNet-18 trained from scratch**, at 10 / 25 / 50 / 100 % of the labelled data.
16
+
17
+ Code: https://github.com/0x-khalil/OSAE
18
+
19
+ * Shared feature extractor: ResNet-18 adapted to 96×96 (3×3/2 stem, no max-pool); identical downstream classifier for all models.
20
+ * Standard AE latent: 128×6×6 = 4,608 coefficients (0.17× input). Overcomplete AE latent: 2048×6×6 = 73,728 (2.67× input).
21
+ * Pre-training on the 100,000 unlabeled images; λ* = 0.01 (selected on validation).
22
+ * Downstream: full fine-tuning and frozen-encoder protocols, 3 seeds, identical stratified subsets,
23
+ validation-only model selection, test split used once for final evaluation.
24
+
25
+ ## Files
26
+
27
+ | Path | Content |
28
+ |---|---|
29
+ | `results/RESULTS.md` | headline table, ablations, paired statistics, costs |
30
+ | `results/figures/` | label-efficiency curves, confusion matrices, per-class accuracy |
31
+ | `results/representation/` | sparsity, reconstruction, separability, PCA / t-SNE / UMAP, CKA |
32
+ | `encoders/<variant>/encoder.pt` | pre-trained ResNet-18 trunk (`torch.load(...)["encoder"]`) |
33
+ | `downstream/` | per-run metrics (and checkpoints if uploaded) |
34
+ | `configs/`, `data_splits/` | exact configuration and split indices |
35
+
36
+ Load a pre-trained trunk with the code from the repository:
37
+
38
+ ```python
39
+ import torch
40
+ from osae.models import Classifier
41
+ clf = Classifier()
42
+ clf.encoder.load_state_dict(torch.load("encoders/oae_sparse/encoder.pt")["encoder"])
43
+ ```
44
+
45
+ ---
46
+
47
+ # Results
48
+
49
+ All numbers: mean ± std over seeds, in %. Test = official STL-10 test split (8,000 images); every choice below was made on the validation split.
50
+
51
+ ## finetune — test_top1
52
+
53
+ | Model | variant | 10% | 25% | 50% | 100% |
54
+ |:-------------------|:--------------------|:-------------|:-------------|:-------------|:-------------|
55
+ | ResNet-18 baseline | scratch | 44.11 ± 0.62 | 54.44 ± 1.86 | 65.16 ± 0.97 | 72.77 ± 0.40 |
56
+ | Standard AE | C128_l0_clean | 49.19 ± 0.28 | 58.44 ± 0.29 | 65.80 ± 0.12 | 68.91 ± 0.64 |
57
+ | Sparse OAE | C2048_l0.01_mask0.5 | 49.55 ± 0.84 | 59.21 ± 0.54 | 67.40 ± 0.14 | 69.82 ± 4.37 |
58
+
59
+ ## finetune — test_macro_f1
60
+
61
+ | Model | variant | 10% | 25% | 50% | 100% |
62
+ |:-------------------|:--------------------|:-------------|:-------------|:-------------|:-------------|
63
+ | ResNet-18 baseline | scratch | 44.00 ± 0.65 | 54.40 ± 1.93 | 65.21 ± 0.97 | 72.87 ± 0.39 |
64
+ | Standard AE | C128_l0_clean | 49.02 ± 0.28 | 58.46 ± 0.28 | 65.84 ± 0.16 | 68.81 ± 0.77 |
65
+ | Sparse OAE | C2048_l0.01_mask0.5 | 49.36 ± 0.82 | 59.16 ± 0.46 | 67.45 ± 0.15 | 69.76 ± 4.46 |
66
+
67
+ ## finetune — val_top1
68
+
69
+ | Model | variant | 10% | 25% | 50% | 100% |
70
+ |:-------------------|:--------------------|:-------------|:-------------|:-------------|:-------------|
71
+ | ResNet-18 baseline | scratch | 47.47 ± 2.31 | 57.20 ± 1.22 | 64.40 ± 1.59 | 70.33 ± 1.10 |
72
+ | Standard AE | C128_l0_clean | 51.13 ± 0.42 | 57.07 ± 0.95 | 62.27 ± 1.50 | 64.33 ± 1.30 |
73
+ | Sparse OAE | C2048_l0.01_mask0.5 | 50.40 ± 1.83 | 59.33 ± 0.31 | 64.07 ± 0.70 | 66.80 ± 0.53 |
74
+
75
+ ## finetune — all variants (test_top1)
76
+
77
+ | variant | latent | lambda | noise | 10% | 25% | 50% | 100% | val (all f) |
78
+ |:---------------------|:---------|:---------|:---------|:-------------|:-------------|:-------------|:-------------|--------------:|
79
+ | ae_sparse | 128 | 0.01 | none | 47.90 ± 0.87 | 57.55 ± 0.64 | 64.12 ± 0.63 | 67.90 ± 0.79 | 59.02 |
80
+ | ae_sparse_denoise | 128 | 0.01 | gaussian | 48.93 ± 0.15 | 57.14 ± 0.50 | 63.78 ± 0.94 | 68.45 ± 0.51 | 59 |
81
+ | ae | 128 | 0.0 | none | 49.19 ± 0.28 | 58.44 ± 0.29 | 65.80 ± 0.12 | 68.91 ± 0.64 | 58.7 |
82
+ | ae_denoise | 128 | 0.0 | gaussian | 48.88 ± 0.48 | 58.58 ± 0.24 | 65.35 ± 0.73 | 69.50 ± 0.35 | 59.17 |
83
+ | oae_l0.01/oae_sparse | 2048 | 0.01 | none | 49.22 ± 0.51 | 57.58 ± 0.64 | 64.71 ± 0.39 | 67.90 ± 0.99 | 57.92 |
84
+ | oae_sparse_denoise | 2048 | 0.01 | gaussian | 48.85 ± 0.66 | 57.60 ± 0.53 | 65.05 ± 0.52 | 69.51 ± 0.14 | 58.5 |
85
+ | oae_sparse_mask | 2048 | 0.01 | mask | 49.55 ± 0.84 | 59.21 ± 0.54 | 67.40 ± 0.14 | 69.82 ± 4.37 | 60.15 |
86
+ | oae | 2048 | 0.0 | none | 49.77 ± 0.46 | 55.69 ± 1.58 | 65.09 ± 0.33 | 69.55 ± 0.48 | 58.33 |
87
+ | oae_denoise | 2048 | 0.0 | gaussian | 48.93 ± 0.40 | 57.85 ± 0.46 | 65.55 ± 0.16 | 69.77 ± 0.21 | 59.35 |
88
+ | scratch | – | – | – | 44.11 ± 0.62 | 54.44 ± 1.86 | 65.16 ± 0.97 | 72.77 ± 0.40 | 59.85 |
89
+
90
+ ## finetune — downstream training cost (mean over runs)
91
+
92
+ | canonical | train_time_s | peak_train_mem_mb | params_total | params_trainable |
93
+ |:---------------------|---------------:|--------------------:|---------------:|-------------------:|
94
+ | C128_l0.01_clean | 147.5 | 919.2 | 1.1174e+07 | 1.1174e+07 |
95
+ | C128_l0.01_gauss0.5 | 145.1 | 919.2 | 1.1174e+07 | 1.1174e+07 |
96
+ | C128_l0_clean | 153 | 919.2 | 1.1174e+07 | 1.1174e+07 |
97
+ | C128_l0_gauss0.5 | 149.9 | 919.2 | 1.1174e+07 | 1.1174e+07 |
98
+ | C2048_l0.01_clean | 148.4 | 919.2 | 1.1174e+07 | 1.1174e+07 |
99
+ | C2048_l0.01_gauss0.5 | 142 | 919.2 | 1.1174e+07 | 1.1174e+07 |
100
+ | C2048_l0.01_mask0.5 | 143.9 | 919.2 | 1.1174e+07 | 1.1174e+07 |
101
+ | C2048_l0_clean | 148.7 | 919.2 | 1.1174e+07 | 1.1174e+07 |
102
+ | C2048_l0_gauss0.5 | 147.4 | 919.2 | 1.1174e+07 | 1.1174e+07 |
103
+ | scratch | 151 | 919.2 | 1.1174e+07 | 1.1174e+07 |
104
+
105
+ ## finetune — paired comparisons (test top-1, B − A in points)
106
+
107
+ | A | B | fraction | n_seeds | diff_B_minus_A | diff_std | p_paired_t | p_paired_t_holm | cohens_dz | p_mcnemar_max_over_seeds |
108
+ |:------------------|:---------------------|-----------:|----------:|-----------------:|-----------:|-------------:|------------------:|------------:|---------------------------:|
109
+ | scratch | C128_l0_clean | 0.1 | 3 | 5.083 | 0.719 | 0.006602 | 0.02641 | 7.07 | 2.226e-14 |
110
+ | scratch | C128_l0_clean | 0.25 | 3 | 4.004 | 2.073 | 0.07893 | 0.1579 | 1.931 | 0.002828 |
111
+ | scratch | C128_l0_clean | 0.5 | 3 | 0.6417 | 1.079 | 0.4112 | 0.4112 | 0.5947 | 0.3509 |
112
+ | scratch | C128_l0_clean | 1 | 3 | -3.858 | 0.5837 | 0.007542 | 0.02641 | -6.61 | 7.597e-11 |
113
+ | scratch | C2048_l0.01_mask0.5 | 0.1 | 3 | 5.438 | 0.9135 | 0.009278 | 0.03711 | 5.952 | 6.667e-17 |
114
+ | scratch | C2048_l0.01_mask0.5 | 0.25 | 3 | 4.775 | 1.338 | 0.0252 | 0.07559 | 3.568 | 2.118e-09 |
115
+ | scratch | C2048_l0.01_mask0.5 | 0.5 | 3 | 2.237 | 1.038 | 0.06488 | 0.1298 | 2.155 | 0.01869 |
116
+ | scratch | C2048_l0.01_mask0.5 | 1 | 3 | -2.95 | 4.179 | 0.346 | 0.346 | -0.7059 | 0.5993 |
117
+ | C128_l0_clean | C2048_l0.01_mask0.5 | 0.1 | 3 | 0.3542 | 1.122 | 0.6394 | 1 | 0.3157 | 0.3909 |
118
+ | C128_l0_clean | C2048_l0.01_mask0.5 | 0.25 | 3 | 0.7708 | 0.8079 | 0.2402 | 0.7207 | 0.9541 | 0.7619 |
119
+ | C128_l0_clean | C2048_l0.01_mask0.5 | 0.5 | 3 | 1.596 | 0.1938 | 0.004879 | 0.01952 | 8.235 | 0.004674 |
120
+ | C128_l0_clean | C2048_l0.01_mask0.5 | 1 | 3 | 0.9083 | 3.722 | 0.7136 | 1 | 0.244 | 4.28e-09 |
121
+ | C128_l0_clean | C2048_l0_clean | 0.1 | 3 | 0.5792 | 0.6411 | 0.2581 | 0.3788 | 0.9034 | 0.8093 |
122
+ | C128_l0_clean | C2048_l0_clean | 0.25 | 3 | -2.75 | 1.875 | 0.1263 | 0.3788 | -1.467 | 0.05449 |
123
+ | C128_l0_clean | C2048_l0_clean | 0.5 | 3 | -0.7125 | 0.219 | 0.03008 | 0.1203 | -3.253 | 0.2517 |
124
+ | C128_l0_clean | C2048_l0_clean | 1 | 3 | 0.6375 | 0.553 | 0.1839 | 0.3788 | 1.153 | 1 |
125
+ | C2048_l0_clean | C2048_l0.01_clean | 0.1 | 3 | -0.55 | 0.2382 | 0.05719 | 0.2288 | -2.309 | 0.3679 |
126
+ | C2048_l0_clean | C2048_l0.01_clean | 0.25 | 3 | 1.887 | 1.003 | 0.08257 | 0.2477 | 1.883 | 0.01547 |
127
+ | C2048_l0_clean | C2048_l0.01_clean | 0.5 | 3 | -0.3792 | 0.7142 | 0.4549 | 0.4549 | -0.5309 | 0.3409 |
128
+ | C2048_l0_clean | C2048_l0.01_clean | 1 | 3 | -1.646 | 1.06 | 0.1149 | 0.2477 | -1.553 | 0.1841 |
129
+ | C128_l0_clean | C128_l0.01_clean | 0.1 | 3 | -1.288 | 0.7253 | 0.0915 | 0.2745 | -1.775 | 0.1862 |
130
+ | C128_l0_clean | C128_l0.01_clean | 0.25 | 3 | -0.8917 | 0.7381 | 0.1715 | 0.343 | -1.208 | 0.7835 |
131
+ | C128_l0_clean | C128_l0.01_clean | 0.5 | 3 | -1.683 | 0.5693 | 0.03608 | 0.1443 | -2.957 | 0.01671 |
132
+ | C128_l0_clean | C128_l0.01_clean | 1 | 3 | -1.008 | 1.264 | 0.3012 | 0.343 | -0.7977 | 0.4219 |
133
+ | C2048_l0.01_clean | C2048_l0.01_gauss0.5 | 0.1 | 3 | -0.3708 | 0.4537 | 0.2925 | 0.8775 | -0.8174 | 0.824 |
134
+ | C2048_l0.01_clean | C2048_l0.01_gauss0.5 | 0.25 | 3 | 0.01667 | 1.096 | 0.9814 | 0.9814 | 0.01521 | 0.2753 |
135
+ | C2048_l0.01_clean | C2048_l0.01_gauss0.5 | 0.5 | 3 | 0.3417 | 0.6825 | 0.4773 | 0.9546 | 0.5006 | 0.6171 |
136
+ | C2048_l0.01_clean | C2048_l0.01_gauss0.5 | 1 | 3 | 1.613 | 0.8928 | 0.08878 | 0.3551 | 1.806 | 0.02627 |
137
+ | C128_l0_clean | C128_l0_gauss0.5 | 0.1 | 3 | -0.3167 | 0.3224 | 0.231 | 0.9241 | -0.9821 | 0.9317 |
138
+ | C128_l0_clean | C128_l0_gauss0.5 | 0.25 | 3 | 0.1417 | 0.4161 | 0.6152 | 0.9852 | 0.3404 | 0.4568 |
139
+ | C128_l0_clean | C128_l0_gauss0.5 | 0.5 | 3 | -0.4583 | 0.6193 | 0.3284 | 0.9852 | -0.7401 | 0.6854 |
140
+ | C128_l0_clean | C128_l0_gauss0.5 | 1 | 3 | 0.5917 | 0.9863 | 0.4079 | 0.9852 | 0.5999 | 0.5875 |
141
+
142
+ ## frozen — test_top1
143
+
144
+ | Model | variant | 10% | 25% | 50% | 100% |
145
+ |:----------------------------------|:--------------------|:-------------|:-------------|:-------------|:-------------|
146
+ | Random frozen ResNet-18 (control) | random | 35.58 ± 0.28 | 39.48 ± 0.34 | 41.61 ± 0.72 | 42.82 ± 1.13 |
147
+ | Standard AE | C128_l0_clean | 43.90 ± 0.29 | 48.54 ± 1.39 | 50.90 ± 0.56 | 52.55 ± 0.21 |
148
+ | Sparse OAE | C2048_l0.01_mask0.5 | 45.72 ± 0.18 | 49.67 ± 1.44 | 53.93 ± 0.09 | 55.66 ± 0.30 |
149
+
150
+ ## frozen — test_macro_f1
151
+
152
+ | Model | variant | 10% | 25% | 50% | 100% |
153
+ |:----------------------------------|:--------------------|:-------------|:-------------|:-------------|:-------------|
154
+ | Random frozen ResNet-18 (control) | random | 35.26 ± 0.24 | 39.22 ± 0.38 | 41.17 ± 0.66 | 42.36 ± 1.14 |
155
+ | Standard AE | C128_l0_clean | 43.83 ± 0.30 | 48.31 ± 1.34 | 50.68 ± 0.62 | 52.26 ± 0.26 |
156
+ | Sparse OAE | C2048_l0.01_mask0.5 | 45.50 ± 0.19 | 49.52 ± 1.48 | 53.86 ± 0.08 | 55.54 ± 0.33 |
157
+
158
+ ## frozen — val_top1
159
+
160
+ | Model | variant | 10% | 25% | 50% | 100% |
161
+ |:----------------------------------|:--------------------|:-------------|:-------------|:-------------|:-------------|
162
+ | Random frozen ResNet-18 (control) | random | 37.20 ± 1.56 | 41.93 ± 1.01 | 42.80 ± 3.33 | 45.27 ± 2.34 |
163
+ | Standard AE | C128_l0_clean | 45.60 ± 0.20 | 50.27 ± 0.81 | 50.73 ± 1.14 | 52.33 ± 1.21 |
164
+ | Sparse OAE | C2048_l0.01_mask0.5 | 46.33 ± 0.50 | 51.20 ± 0.20 | 54.53 ± 0.83 | 56.60 ± 0.53 |
165
+
166
+ ## frozen — all variants (test_top1)
167
+
168
+ | variant | latent | lambda | noise | 10% | 25% | 50% | 100% | val (all f) |
169
+ |:---------------------|:---------|:---------|:---------|:-------------|:-------------|:-------------|:-------------|--------------:|
170
+ | ae_sparse | 128 | 0.01 | none | 44.01 ± 0.35 | 48.34 ± 0.25 | 50.41 ± 0.13 | 51.15 ± 0.36 | 49 |
171
+ | ae_sparse_denoise | 128 | 0.01 | gaussian | 42.26 ± 0.89 | 47.79 ± 0.43 | 49.28 ± 0.30 | 50.71 ± 0.29 | 47.5 |
172
+ | ae | 128 | 0.0 | none | 43.90 ± 0.29 | 48.54 ± 1.39 | 50.90 ± 0.56 | 52.55 ± 0.21 | 49.73 |
173
+ | ae_denoise | 128 | 0.0 | gaussian | 42.89 ± 0.56 | 48.73 ± 0.47 | 51.08 ± 0.47 | 52.66 ± 0.31 | 48.43 |
174
+ | oae_l0.001 | 2048 | 0.001 | none | 44.61 | – | – | 53.02 | 48.3 |
175
+ | oae_l0.01/oae_sparse | 2048 | 0.01 | none | 42.94 ± 0.62 | 48.20 ± 0.14 | 50.65 ± 0.57 | 52.07 ± 0.42 | 49.72 |
176
+ | oae_sparse_denoise | 2048 | 0.01 | gaussian | 43.00 ± 0.61 | 48.59 ± 1.46 | 51.33 ± 0.41 | 52.85 ± 0.26 | 49.53 |
177
+ | oae_sparse_mask | 2048 | 0.01 | mask | 45.72 ± 0.18 | 49.67 ± 1.44 | 53.93 ± 0.09 | 55.66 ± 0.30 | 52.17 |
178
+ | oae_l0.1 | 2048 | 0.1 | none | 42.27 | – | – | 50.56 | 46.5 |
179
+ | oae | 2048 | 0.0 | none | 44.22 ± 0.91 | 49.39 ± 0.86 | 52.15 ± 0.20 | 52.88 ± 0.85 | 51.03 |
180
+ | oae_denoise | 2048 | 0.0 | gaussian | 44.13 ± 0.45 | 50.03 ± 0.24 | 52.55 ± 0.27 | 54.14 ± 0.10 | 50.67 |
181
+ | random | – | – | – | 35.58 ± 0.28 | 39.48 ± 0.34 | 41.61 ± 0.72 | 42.82 ± 1.13 | 41.8 |
182
+
183
+ ## frozen — downstream training cost (mean over runs)
184
+
185
+ | canonical | train_time_s | peak_train_mem_mb | params_total | params_trainable |
186
+ |:---------------------|---------------:|--------------------:|---------------:|-------------------:|
187
+ | C128_l0.01_clean | 59.8 | 832.6 | 1.1174e+07 | 5130 |
188
+ | C128_l0.01_gauss0.5 | 57.3 | 832.6 | 1.1174e+07 | 5130 |
189
+ | C128_l0_clean | 56.6 | 832.6 | 1.1174e+07 | 5130 |
190
+ | C128_l0_gauss0.5 | 59.4 | 832.6 | 1.1174e+07 | 5130 |
191
+ | C2048_l0.001_clean | 64.2 | 832.6 | 1.1174e+07 | 5130 |
192
+ | C2048_l0.01_clean | 56.5 | 832.6 | 1.1174e+07 | 5130 |
193
+ | C2048_l0.01_gauss0.5 | 56.6 | 832.6 | 1.1174e+07 | 5130 |
194
+ | C2048_l0.01_mask0.5 | 57.9 | 832.6 | 1.1174e+07 | 5130 |
195
+ | C2048_l0.1_clean | 66.5 | 832.6 | 1.1174e+07 | 5130 |
196
+ | C2048_l0_clean | 61 | 832.6 | 1.1174e+07 | 5130 |
197
+ | C2048_l0_gauss0.5 | 56.5 | 832.6 | 1.1174e+07 | 5130 |
198
+ | random | 232.8 | 2418.4 | 1.1174e+07 | 5130 |
199
+
200
+ ## frozen — paired comparisons (test top-1, B − A in points)
201
+
202
+ | A | B | fraction | n_seeds | diff_B_minus_A | diff_std | p_paired_t | p_paired_t_holm | cohens_dz | p_mcnemar_max_over_seeds |
203
+ |:------------------|:---------------------|-----------:|----------:|-----------------:|-----------:|-------------:|------------------:|------------:|---------------------------:|
204
+ | C128_l0_clean | C2048_l0.01_mask0.5 | 0.1 | 3 | 1.825 | 0.1146 | 0.001311 | 0.005244 | 15.93 | 0.0018 |
205
+ | C128_l0_clean | C2048_l0.01_mask0.5 | 0.25 | 3 | 1.133 | 2.788 | 0.5543 | 0.5543 | 0.4066 | 0.0002473 |
206
+ | C128_l0_clean | C2048_l0.01_mask0.5 | 0.5 | 3 | 3.029 | 0.5236 | 0.009812 | 0.02457 | 5.786 | 6.927e-06 |
207
+ | C128_l0_clean | C2048_l0.01_mask0.5 | 1 | 3 | 3.117 | 0.4915 | 0.00819 | 0.02457 | 6.341 | 9.169e-07 |
208
+ | C128_l0_clean | C2048_l0_clean | 0.1 | 3 | 0.3208 | 0.657 | 0.4867 | 1 | 0.4884 | 0.8107 |
209
+ | C128_l0_clean | C2048_l0_clean | 0.25 | 3 | 0.85 | 2.234 | 0.5776 | 1 | 0.3805 | 0.0009237 |
210
+ | C128_l0_clean | C2048_l0_clean | 0.5 | 3 | 1.242 | 0.5646 | 0.06253 | 0.2501 | 2.199 | 0.2492 |
211
+ | C128_l0_clean | C2048_l0_clean | 1 | 3 | 0.3333 | 0.7821 | 0.5372 | 1 | 0.4262 | 0.2995 |
212
+ | C2048_l0_clean | C2048_l0.01_clean | 0.1 | 3 | -1.279 | 0.8055 | 0.1107 | 0.332 | -1.588 | 0.4022 |
213
+ | C2048_l0_clean | C2048_l0.01_clean | 0.25 | 3 | -1.192 | 0.9306 | 0.1568 | 0.332 | -1.281 | 0.8221 |
214
+ | C2048_l0_clean | C2048_l0.01_clean | 0.5 | 3 | -1.496 | 0.6377 | 0.05557 | 0.2223 | -2.346 | 0.09288 |
215
+ | C2048_l0_clean | C2048_l0.01_clean | 1 | 3 | -0.8042 | 1.172 | 0.3567 | 0.3567 | -0.6861 | 0.3488 |
216
+ | C128_l0_clean | C128_l0.01_clean | 0.1 | 3 | 0.1125 | 0.5717 | 0.7657 | 1 | 0.1968 | 0.5233 |
217
+ | C128_l0_clean | C128_l0.01_clean | 0.25 | 3 | -0.2 | 1.61 | 0.8496 | 1 | -0.1242 | 0.2949 |
218
+ | C128_l0_clean | C128_l0.01_clean | 0.5 | 3 | -0.4958 | 0.6831 | 0.3356 | 1 | -0.7259 | 1 |
219
+ | C128_l0_clean | C128_l0.01_clean | 1 | 3 | -1.396 | 0.2306 | 0.008975 | 0.0359 | -6.053 | 0.01661 |
220
+ | C2048_l0.01_clean | C2048_l0.01_gauss0.5 | 0.1 | 3 | 0.0625 | 0.598 | 0.873 | 1 | 0.1045 | 0.5002 |
221
+ | C2048_l0.01_clean | C2048_l0.01_gauss0.5 | 0.25 | 3 | 0.3958 | 1.325 | 0.6564 | 1 | 0.2988 | 0.05433 |
222
+ | C2048_l0.01_clean | C2048_l0.01_gauss0.5 | 0.5 | 3 | 0.675 | 0.3356 | 0.07345 | 0.2938 | 2.011 | 0.5933 |
223
+ | C2048_l0.01_clean | C2048_l0.01_gauss0.5 | 1 | 3 | 0.7792 | 0.4661 | 0.1014 | 0.3043 | 1.672 | 0.5116 |
224
+ | C128_l0_clean | C128_l0_gauss0.5 | 0.1 | 3 | -1.008 | 0.593 | 0.09853 | 0.3941 | -1.7 | 0.5056 |
225
+ | C128_l0_clean | C128_l0_gauss0.5 | 0.25 | 3 | 0.1958 | 0.9284 | 0.7499 | 1 | 0.2109 | 0.203 |
226
+ | C128_l0_clean | C128_l0_gauss0.5 | 0.5 | 3 | 0.175 | 0.4265 | 0.5509 | 1 | 0.4103 | 0.9387 |
227
+ | C128_l0_clean | C128_l0_gauss0.5 | 1 | 3 | 0.1125 | 0.142 | 0.3036 | 0.9107 | 0.7924 | 0.9418 |
228
+
229
+ ## Frozen vs fine-tuned (test_top1, mean over seeds)
230
+
231
+ | | finetune | frozen | finetune_minus_frozen |
232
+ |:-------------------------------|-----------:|---------:|------------------------:|
233
+ | ('C128_l0.01_clean', 0.1) | 47.9 | 44.01 | 3.9 |
234
+ | ('C128_l0.01_clean', 0.25) | 57.55 | 48.34 | 9.21 |
235
+ | ('C128_l0.01_clean', 0.5) | 64.12 | 50.41 | 13.71 |
236
+ | ('C128_l0.01_clean', 1.0) | 67.9 | 51.15 | 16.75 |
237
+ | ('C128_l0.01_gauss0.5', 0.1) | 48.93 | 42.26 | 6.67 |
238
+ | ('C128_l0.01_gauss0.5', 0.25) | 57.14 | 47.79 | 9.35 |
239
+ | ('C128_l0.01_gauss0.5', 0.5) | 63.78 | 49.28 | 14.5 |
240
+ | ('C128_l0.01_gauss0.5', 1.0) | 68.45 | 50.71 | 17.74 |
241
+ | ('C128_l0_clean', 0.1) | 49.19 | 43.9 | 5.3 |
242
+ | ('C128_l0_clean', 0.25) | 58.44 | 48.54 | 9.9 |
243
+ | ('C128_l0_clean', 0.5) | 65.8 | 50.9 | 14.9 |
244
+ | ('C128_l0_clean', 1.0) | 68.91 | 52.55 | 16.36 |
245
+ | ('C128_l0_gauss0.5', 0.1) | 48.88 | 42.89 | 5.99 |
246
+ | ('C128_l0_gauss0.5', 0.25) | 58.58 | 48.73 | 9.85 |
247
+ | ('C128_l0_gauss0.5', 0.5) | 65.35 | 51.08 | 14.27 |
248
+ | ('C128_l0_gauss0.5', 1.0) | 69.5 | 52.66 | 16.84 |
249
+ | ('C2048_l0.01_clean', 0.1) | 49.22 | 42.94 | 6.28 |
250
+ | ('C2048_l0.01_clean', 0.25) | 57.58 | 48.2 | 9.38 |
251
+ | ('C2048_l0.01_clean', 0.5) | 64.71 | 50.65 | 14.06 |
252
+ | ('C2048_l0.01_clean', 1.0) | 67.9 | 52.08 | 15.82 |
253
+ | ('C2048_l0.01_gauss0.5', 0.1) | 48.85 | 43 | 5.85 |
254
+ | ('C2048_l0.01_gauss0.5', 0.25) | 57.6 | 48.59 | 9 |
255
+ | ('C2048_l0.01_gauss0.5', 0.5) | 65.05 | 51.32 | 13.73 |
256
+ | ('C2048_l0.01_gauss0.5', 1.0) | 69.51 | 52.85 | 16.66 |
257
+ | ('C2048_l0.01_mask0.5', 0.1) | 49.55 | 45.72 | 3.83 |
258
+ | ('C2048_l0.01_mask0.5', 0.25) | 59.21 | 49.67 | 9.54 |
259
+ | ('C2048_l0.01_mask0.5', 0.5) | 67.4 | 53.93 | 13.47 |
260
+ | ('C2048_l0.01_mask0.5', 1.0) | 69.82 | 55.66 | 14.15 |
261
+ | ('C2048_l0_clean', 0.1) | 49.77 | 44.22 | 5.55 |
262
+ | ('C2048_l0_clean', 0.25) | 55.69 | 49.39 | 6.3 |
263
+ | ('C2048_l0_clean', 0.5) | 65.09 | 52.15 | 12.95 |
264
+ | ('C2048_l0_clean', 1.0) | 69.55 | 52.88 | 16.67 |
265
+ | ('C2048_l0_gauss0.5', 0.1) | 48.93 | 44.13 | 4.8 |
266
+ | ('C2048_l0_gauss0.5', 0.25) | 57.85 | 50.03 | 7.82 |
267
+ | ('C2048_l0_gauss0.5', 0.5) | 65.55 | 52.55 | 13.01 |
268
+ | ('C2048_l0_gauss0.5', 1.0) | 69.77 | 54.14 | 15.63 |
269
+
270
+ ## Autoencoder pre-training
271
+
272
+ | variant | latent_dim | overcomplete_x | params_encoder | params_decoder | train_time_h | img_per_s | peak_mem_mb | holdout_psnr | frac_active |
273
+ |:-------------------|-------------:|-----------------:|-----------------:|-----------------:|---------------:|------------:|--------------:|---------------:|--------------:|
274
+ | oae_l0.001 | 73728 | 2.667 | 12219456 | 2617155 | 0.3322 | 818.7 | 2550 | 25.1 | 0.22 |
275
+ | oae_l0.01 | 73728 | 2.667 | 12219456 | 2617155 | 0.3115 | 873.3 | 2550 | 25.05 | 0.1399 |
276
+ | oae_l0.1 | 73728 | 2.667 | 12219456 | 2617155 | 0.3331 | 816.6 | 2550 | 24.79 | 0.05403 |
277
+ | ae | 4608 | 0.167 | 11234496 | 1634115 | 0.2988 | 910.4 | 2507 | 25.1 | 0.7798 |
278
+ | ae_sparse | 4608 | 0.167 | 11234496 | 1634115 | 0.317 | 858.2 | 2507 | 24.78 | 0.396 |
279
+ | ae_denoise | 4608 | 0.167 | 11234496 | 1634115 | 0.294 | 925.1 | 2674 | 24.74 | 0.7822 |
280
+ | ae_sparse_denoise | 4608 | 0.167 | 11234496 | 1634115 | 0.3021 | 900.3 | 2674 | 24.44 | 0.3752 |
281
+ | oae | 73728 | 2.667 | 12219456 | 2617155 | 0.313 | 869.1 | 2550 | 25.17 | 0.2507 |
282
+ | oae_sparse | 73728 | 2.667 | 12219456 | 2617155 | 0.3115 | 873.3 | 2550 | 25.05 | 0.1399 |
283
+ | oae_denoise | 73728 | 2.667 | 12219456 | 2617155 | 0.2901 | 937.7 | 2564 | 24.79 | 0.2423 |
284
+ | oae_sparse_denoise | 73728 | 2.667 | 12219456 | 2617155 | 0.2947 | 923.1 | 2564 | 24.73 | 0.1336 |
285
+ | oae_sparse_mask | 73728 | 2.667 | 12219456 | 2617155 | 0.2928 | 928.9 | 2564 | 24.01 | 0.08135 |
286
+
287
+ ## Inference cost (identical downstream architecture for all models)
288
+
289
+ Parameters: 11,173,962
290
+
291
+ | | latency_ms_median | latency_ms_p90 | throughput_img_s | peak_gpu_mem_mb |
292
+ |:------|--------------------:|-----------------:|-------------------:|------------------:|
293
+ | bs1 | 3.06 | 3.21 | 326 | 76.4 |
294
+ | bs256 | 45 | 45.2 | 5.69e+03 | 442 |
295
+
296
+ ## Selections (validation only)
297
+
298
+ ```
299
+ {'finetune/ResNet-18 baseline': {'chosen': 'scratch', 'mean_val_top1': {'scratch': np.float64(0.5984999999999999)}}, 'finetune/Standard AE': {'chosen': 'C128_l0_clean', 'mean_val_top1': {'C128_l0_clean': np.float64(0.587)}}, 'finetune/Sparse OAE': {'chosen': 'C2048_l0.01_mask0.5', 'mean_val_top1': {'C2048_l0.01_clean': np.float64(0.5791666666666667), 'C2048_l0.01_gauss0.5': np.float64(0.585), 'C2048_l0.01_mask0.5': np.float64(0.6015)}}, 'frozen/Random frozen ResNet-18 (control)': {'chosen': 'random', 'mean_val_top1': {'random': np.float64(0.418)}}, 'frozen/Standard AE': {'chosen': 'C128_l0_clean', 'mean_val_top1': {'C128_l0_clean': np.float64(0.49733333333333346)}}, 'frozen/Sparse OAE': {'chosen': 'C2048_l0.01_mask0.5', 'mean_val_top1': {'C2048_l0.01_clean': np.float64(0.4971666666666666), 'C2048_l0.01_gauss0.5': np.float64(0.4953333333333334), 'C2048_l0.01_mask0.5': np.float64(0.5216666666666668)}}}
300
+ ```
301
+
configs/default.yaml ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Full experimental protocol. Run with: python scripts/run_all.py --config configs/default.yaml
2
+ # Every setting that must be identical across experiments lives here, once.
3
+
4
+ data_dir: data/stl10
5
+ runs_dir: runs
6
+ results_dir: results
7
+ amp: auto # bf16 on Ampere+ (RTX 30xx-50xx, A100), fp16 + loss scaling on older GPUs (T4); "none" = fp32
8
+
9
+ # Jobs run concurrently on the one GPU. 1 = sequential (fine for an 8 GB laptop GPU).
10
+ # On a 24-32 GB GPU (RTX 4090 / 5090): downstream_jobs 4, pretrain_jobs 2. CLI: --jobs / --pretrain_jobs.
11
+ parallel:
12
+ downstream_jobs: 1
13
+ pretrain_jobs: 1
14
+
15
+ splits:
16
+ fractions: [0.1, 0.25, 0.5, 1.0] # of the 4,500-image labelled pool (after removing validation)
17
+ val_per_class: 50 # 500-image stratified validation split (model selection / early stopping)
18
+ split_seed: 0 # fixed -> identical subsets for every method and seed
19
+ unlabeled_holdout: 2000 # unlabeled images held out to monitor reconstruction
20
+
21
+ seeds: [0, 1, 2] # downstream seeds (init of head / scratch model, data order, augmentation)
22
+ pretrain_seeds: [0] # [0] = one AE per variant shared by all downstream seeds;
23
+ # [0, 1, 2] = downstream seed s uses AE pre-trained with seed s (3x cost)
24
+
25
+ # ---------------------------------------------------------------------------- AE pre-training
26
+ pretrain:
27
+ epochs: 20 # over 98,000 unlabeled images
28
+ batch_size: 128
29
+ lr: 0.001 # AdamW, linear warm-up then cosine decay
30
+ weight_decay: 0.0001
31
+ warmup_epochs: 1
32
+ noise_std: 0.5 # Gaussian denoising: std in normalised units
33
+ mask_ratio: 0.5 # masking denoising: fraction of 8x8 patches zeroed
34
+ mask_patch: 8
35
+ max_steps: null
36
+
37
+ # Stage 1: sparsity sweep on the overcomplete AE (clean inputs). lambda* = the lambda > 0 with the best
38
+ # mean *validation* accuracy of a frozen-encoder probe on the selection grid below.
39
+ lambda_sweep:
40
+ latent_channels: 2048
41
+ values: [0.001, 0.01, 0.1] # lambda = 0 (non-sparse OAE) is always trained as variant `oae`
42
+ selection:
43
+ protocol: frozen
44
+ fractions: [0.1, 1.0]
45
+ seeds: [0]
46
+ full_downstream_grid: false # true -> also run every protocol/fraction/seed for every sweep lambda
47
+
48
+ # Stage 2: factorial design {undercomplete, overcomplete} x {lambda=0, lambda*} x {clean, Gaussian}
49
+ # + masking-denoising OAE. `lambda_star` is resolved after stage 1.
50
+ variants:
51
+ ae: {latent_channels: 128, l1_lambda: 0.0, noise: none} # Experiment B
52
+ ae_sparse: {latent_channels: 128, l1_lambda: lambda_star, noise: none}
53
+ ae_denoise: {latent_channels: 128, l1_lambda: 0.0, noise: gaussian}
54
+ ae_sparse_denoise: {latent_channels: 128, l1_lambda: lambda_star, noise: gaussian}
55
+ oae: {latent_channels: 2048, l1_lambda: 0.0, noise: none}
56
+ oae_sparse: {latent_channels: 2048, l1_lambda: lambda_star, noise: none} # Experiment C
57
+ oae_denoise: {latent_channels: 2048, l1_lambda: 0.0, noise: gaussian}
58
+ oae_sparse_denoise: {latent_channels: 2048, l1_lambda: lambda_star, noise: gaussian} # Experiment C
59
+ oae_sparse_mask: {latent_channels: 2048, l1_lambda: lambda_star, noise: mask} # Experiment C
60
+
61
+ # Rows of the headline table. With several candidates, the one with the best mean validation accuracy wins.
62
+ main_table:
63
+ ResNet-18 baseline: [scratch]
64
+ Random frozen ResNet-18 (control): [random]
65
+ Standard AE: [ae]
66
+ Sparse OAE: [oae_sparse, oae_sparse_denoise, oae_sparse_mask]
67
+
68
+ # Paired statistical comparisons (A, B) -> reports B - A. Row names of main_table or variant names.
69
+ comparisons:
70
+ - [ResNet-18 baseline, Standard AE]
71
+ - [ResNet-18 baseline, Sparse OAE]
72
+ - [Standard AE, Sparse OAE]
73
+ - [ae, oae] # overcompleteness alone
74
+ - [oae, oae_sparse] # sparsity given overcompleteness
75
+ - [ae, ae_sparse] # sparsity given undercompleteness
76
+ - [oae_sparse, oae_sparse_denoise] # denoising given sparse OAE
77
+ - [ae, ae_denoise] # denoising given undercomplete AE
78
+
79
+ # ---------------------------------------------------------------------------- downstream
80
+ downstream:
81
+ protocols: [finetune, frozen]
82
+ common: # identical for every method, fraction and seed
83
+ steps: 2000 # fixed optimisation budget (batch 128 -> ~57 epochs at 100%, ~570 at 10%)
84
+ batch_size: 128
85
+ momentum: 0.9 # SGD + Nesterov, linear warm-up then cosine decay
86
+ warmup_steps: 100
87
+ eval_every: 100 # validation check; the best-validation checkpoint is kept (early stopping)
88
+ patience: 0
89
+ label_smoothing: 0.0
90
+ finetune: {lr: 0.05, weight_decay: 0.0005}
91
+ frozen: {lr: 0.1, weight_decay: 0.0}
92
+
93
+ analysis:
94
+ embed_samples: 2000 # points in t-SNE / UMAP plots
95
+ include_supervised_reference: true # add the 100%-label scratch model (seed 0) to the analysis
96
+ pairs: ["ae,oae_sparse", "ae,oae", "oae,oae_sparse", "scratch_100,oae_sparse"]
configs/fast.yaml ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Same protocol as default.yaml (all variants, both protocols, 4 fractions, 3 seeds) with a halved
2
+ # training budget: ~7-7.5 h on an RTX 3070 Laptop instead of ~16-17 h. Results go to separate folders
3
+ # so they never mix with a full-budget run.
4
+ base: default.yaml
5
+
6
+ runs_dir: runs_fast
7
+ results_dir: results_fast
8
+
9
+ parallel:
10
+ downstream_jobs: 2 # 2 concurrent downstream runs fit easily in 8 GB
11
+ pretrain_jobs: 1 # one AE pre-training (~4.5 GB) at a time on 8 GB
12
+
13
+ pretrain:
14
+ epochs: 10 # default: 20
15
+
16
+ downstream:
17
+ common:
18
+ steps: 1000 # default: 2000 (still ~28 epochs at 100% labels, ~280 at 10%)
configs/kaggle.yaml ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ # Kaggle free tier (2x Tesla T4, 12 h sessions). Same scientific protocol as default.yaml; only the
2
+ # infrastructure differs. Run with scripts/kaggle_session.py (see README, "Running on Kaggle").
3
+ base: default.yaml
4
+
5
+ data_dir: /tmp/stl10 # outside /kaggle/working (20 GB output limit); re-downloaded each session
6
+ delete_raw_data: true # drop the 2.6 GB archive after conversion
7
+ amp: auto # T4 has no bf16 -> fp16 with loss scaling (bf16 on Ampere+ GPUs)
configs/kaggle_fast.yaml ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Kaggle (2x T4) with the halved training budget: fits in ONE 12-hour session (~6-7.5 h estimated).
2
+ # Same design as the full study (all variants, both protocols, 4 fractions, 3 seeds); only
3
+ # pre-training epochs (20 -> 10) and downstream steps (2000 -> 1000) are halved.
4
+ base: kaggle.yaml
5
+
6
+ runs_dir: runs_fast
7
+ results_dir: results_fast
8
+
9
+ pretrain:
10
+ epochs: 10
11
+
12
+ downstream:
13
+ common:
14
+ steps: 1000
configs/smoke.yaml ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Tiny end-to-end check of the whole pipeline (minutes, not hours). Numbers are meaningless.
2
+ # python scripts/run_all.py --config configs/smoke.yaml
3
+ data_dir: data/stl10
4
+ runs_dir: runs_smoke
5
+ results_dir: results_smoke
6
+ amp: bf16
7
+
8
+ # Jobs run concurrently on the one GPU. 1 = sequential (fine for an 8 GB laptop GPU).
9
+ # On a 24-32 GB GPU (RTX 4090 / 5090): downstream_jobs 4, pretrain_jobs 2. CLI: --jobs / --pretrain_jobs.
10
+ parallel:
11
+ downstream_jobs: 1
12
+ pretrain_jobs: 1
13
+
14
+ splits: {fractions: [0.1, 0.25, 0.5, 1.0], val_per_class: 50, split_seed: 0, unlabeled_holdout: 2000}
15
+ seeds: [0, 1]
16
+ pretrain_seeds: [0]
17
+
18
+ pretrain: {epochs: 1, batch_size: 64, lr: 0.001, weight_decay: 0.0001, warmup_epochs: 1,
19
+ noise_std: 0.5, mask_ratio: 0.5, mask_patch: 8, max_steps: 30, holdout_max: 256}
20
+
21
+ lambda_sweep:
22
+ latent_channels: 2048
23
+ values: [0.01, 0.1]
24
+ selection: {protocol: frozen, fractions: [0.1], seeds: [0]}
25
+ full_downstream_grid: false
26
+
27
+ variants:
28
+ ae: {latent_channels: 128, l1_lambda: 0.0, noise: none}
29
+ oae: {latent_channels: 2048, l1_lambda: 0.0, noise: none}
30
+ oae_sparse: {latent_channels: 2048, l1_lambda: lambda_star, noise: none}
31
+ oae_sparse_mask: {latent_channels: 2048, l1_lambda: lambda_star, noise: mask}
32
+
33
+ main_table:
34
+ ResNet-18 baseline: [scratch]
35
+ Random frozen ResNet-18 (control): [random]
36
+ Standard AE: [ae]
37
+ Sparse OAE: [oae_sparse, oae_sparse_mask]
38
+
39
+ comparisons:
40
+ - [ResNet-18 baseline, Sparse OAE]
41
+ - [Standard AE, Sparse OAE]
42
+ - [ae, oae]
43
+
44
+ downstream:
45
+ protocols: [finetune, frozen]
46
+ common: {steps: 20, batch_size: 64, momentum: 0.9, warmup_steps: 5, eval_every: 10, patience: 0, label_smoothing: 0.0}
47
+ finetune: {lr: 0.05, weight_decay: 0.0005}
48
+ frozen: {lr: 0.1, weight_decay: 0.0}
49
+
50
+ analysis:
51
+ embed_samples: 300
52
+ max_pool: 500
53
+ include_supervised_reference: true
54
+ pairs: ["ae,oae_sparse"]
data_splits/splits.json ADDED
The diff for this file is too large to render. See raw diff
 
data_splits/splits_summary.md ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # STL-10 splits
2
+
3
+ split_seed = 0; validation = 50 images/class drawn from the official 5,000-image training split; the official 8,000-image test split is untouched.
4
+
5
+ | subset | images | airplane | bird | car | cat | deer | dog | horse | monkey | ship | truck |
6
+ |---|---|---|---|---|---|---|---|---|---|---|---|
7
+ | validation | 500 | 50 | 50 | 50 | 50 | 50 | 50 | 50 | 50 | 50 | 50 |
8
+ | train 10% | 450 | 45 | 45 | 45 | 45 | 45 | 45 | 45 | 45 | 45 | 45 |
9
+ | train 25% | 1120 | 112 | 112 | 112 | 112 | 112 | 112 | 112 | 112 | 112 | 112 |
10
+ | train 50% | 2250 | 225 | 225 | 225 | 225 | 225 | 225 | 225 | 225 | 225 | 225 |
11
+ | train 100% | 4500 | 450 | 450 | 450 | 450 | 450 | 450 | 450 | 450 | 450 | 450 |
12
+
13
+ Unlabeled: 98000 images for AE training, 2000 held out for reconstruction monitoring.
14
+
15
+ Labelled subsets are nested (10% ⊂ 25% ⊂ 50% ⊂ 100%) and identical for every method and seed.
16
+ Exact indices (into the official train / unlabeled arrays) are in `splits.json`.
data_splits/stats.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "mean": [
3
+ 0.44058638793147475,
4
+ 0.42731113699108136,
5
+ 0.38579501157407375
6
+ ],
7
+ "std": [
8
+ 0.2686545583839396,
9
+ 0.2612521862515845,
10
+ 0.26849920555418677
11
+ ],
12
+ "computed_on": "unlabeled (100,000 images)"
13
+ }
downstream/C128_l0.01_clean/finetune/f0.1/s0/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/finetune/f0.1/s0",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "finetune",
8
+ "fraction": 0.1,
9
+ "seed": 0,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.05,
13
+ "weight_decay": 0.0005,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 450,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 45,
26
+ 45,
27
+ 45,
28
+ 45,
29
+ 45,
30
+ 45,
31
+ 45,
32
+ 45,
33
+ 45,
34
+ 45
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.492,
38
+ "val_macro_f1": 0.4918337981398479,
39
+ "val_loss": 3.370607177734375,
40
+ "step": 700
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 1.1665714979171753,
46
+ "val_loss": 51.0107421875,
47
+ "val_top1": 0.214,
48
+ "val_macro_f1": 0.16285571706420549,
49
+ "lr": 0.05,
50
+ "train_time_s": 18.63870734699958
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 0.2513880431652069,
55
+ "val_loss": 2.9067998046875,
56
+ "val_top1": 0.428,
57
+ "val_macro_f1": 0.4104706526137661,
58
+ "lr": 0.04852201922385564,
59
+ "train_time_s": 34.00254133699946
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 0.08250879496335983,
64
+ "val_loss": 3.075227294921875,
65
+ "val_top1": 0.486,
66
+ "val_macro_f1": 0.4797197036292201,
67
+ "lr": 0.04420708808983809,
68
+ "train_time_s": 49.588061432998984
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 0.013659990392625332,
73
+ "val_loss": 3.1420498046875,
74
+ "val_top1": 0.476,
75
+ "val_macro_f1": 0.47547102210778835,
76
+ "lr": 0.03757549866575588,
77
+ "train_time_s": 65.25845655399917
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 0.013041661120951176,
82
+ "val_loss": 3.532505859375,
83
+ "val_top1": 0.474,
84
+ "val_macro_f1": 0.47650066206638153,
85
+ "lr": 0.029427118507989587,
86
+ "train_time_s": 80.96726245599984
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 0.007182898931205273,
91
+ "val_loss": 3.29467431640625,
92
+ "val_top1": 0.474,
93
+ "val_macro_f1": 0.4742693909972022,
94
+ "lr": 0.020744762520849194,
95
+ "train_time_s": 96.76758273499945
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.00309723149985075,
100
+ "val_loss": 3.370607177734375,
101
+ "val_top1": 0.492,
102
+ "val_macro_f1": 0.4918337981398479,
103
+ "lr": 0.012575650974311119,
104
+ "train_time_s": 112.53901157199925
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.004616633057594299,
109
+ "val_loss": 3.354072265625,
110
+ "val_top1": 0.474,
111
+ "val_macro_f1": 0.4732649107774015,
112
+ "lr": 0.005905099284133953,
113
+ "train_time_s": 128.49585584799934
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.003953780047595501,
118
+ "val_loss": 3.34121728515625,
119
+ "val_top1": 0.48,
120
+ "val_macro_f1": 0.480697628964535,
121
+ "lr": 0.001537674431011102,
122
+ "train_time_s": 144.57307902999946
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.0033200725447386503,
127
+ "val_loss": 3.363711669921875,
128
+ "val_top1": 0.48,
129
+ "val_macro_f1": 0.47998055458343564,
130
+ "lr": 1.5230855524017706e-07,
131
+ "train_time_s": 160.50339227699897
132
+ }
133
+ ],
134
+ "train_time_s": 160.50344123299874,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 919.21826171875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 11173962,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/finetune/f0.1/s0/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.4885,
3
+ "macro_f1": 0.4872389199270398,
4
+ "balanced_accuracy": 0.4885,
5
+ "per_class_accuracy": [
6
+ 0.55,
7
+ 0.4,
8
+ 0.65625,
9
+ 0.4225,
10
+ 0.39625,
11
+ 0.28125,
12
+ 0.545,
13
+ 0.45,
14
+ 0.665,
15
+ 0.51875
16
+ ],
17
+ "per_class_f1": [
18
+ 0.6019151846785227,
19
+ 0.4139715394566624,
20
+ 0.6529850746268656,
21
+ 0.37828763290430883,
22
+ 0.41546526867627787,
23
+ 0.3129346314325452,
24
+ 0.5272067714631198,
25
+ 0.4463732176069436,
26
+ 0.6211325160537069,
27
+ 0.5021173623714459
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 440,
32
+ 34,
33
+ 47,
34
+ 13,
35
+ 9,
36
+ 9,
37
+ 12,
38
+ 9,
39
+ 131,
40
+ 96
41
+ ],
42
+ [
43
+ 51,
44
+ 320,
45
+ 26,
46
+ 125,
47
+ 62,
48
+ 63,
49
+ 43,
50
+ 66,
51
+ 29,
52
+ 15
53
+ ],
54
+ [
55
+ 28,
56
+ 2,
57
+ 525,
58
+ 8,
59
+ 1,
60
+ 11,
61
+ 14,
62
+ 3,
63
+ 52,
64
+ 156
65
+ ],
66
+ [
67
+ 6,
68
+ 68,
69
+ 17,
70
+ 338,
71
+ 85,
72
+ 79,
73
+ 51,
74
+ 126,
75
+ 12,
76
+ 18
77
+ ],
78
+ [
79
+ 10,
80
+ 75,
81
+ 15,
82
+ 153,
83
+ 317,
84
+ 60,
85
+ 98,
86
+ 57,
87
+ 7,
88
+ 8
89
+ ],
90
+ [
91
+ 16,
92
+ 109,
93
+ 6,
94
+ 117,
95
+ 68,
96
+ 225,
97
+ 121,
98
+ 122,
99
+ 6,
100
+ 10
101
+ ],
102
+ [
103
+ 9,
104
+ 36,
105
+ 11,
106
+ 27,
107
+ 98,
108
+ 103,
109
+ 436,
110
+ 56,
111
+ 2,
112
+ 22
113
+ ],
114
+ [
115
+ 2,
116
+ 79,
117
+ 2,
118
+ 161,
119
+ 61,
120
+ 64,
121
+ 56,
122
+ 360,
123
+ 4,
124
+ 11
125
+ ],
126
+ [
127
+ 66,
128
+ 10,
129
+ 29,
130
+ 27,
131
+ 12,
132
+ 12,
133
+ 5,
134
+ 5,
135
+ 532,
136
+ 102
137
+ ],
138
+ [
139
+ 34,
140
+ 13,
141
+ 130,
142
+ 18,
143
+ 13,
144
+ 12,
145
+ 18,
146
+ 9,
147
+ 138,
148
+ 415
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }
downstream/C128_l0.01_clean/finetune/f0.1/s0/test_predictions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:417774b1e35f6ce811b45020658dc50208c6555e68e6547b2ba42f2cdaf81e85
3
+ size 8128
downstream/C128_l0.01_clean/finetune/f0.1/s1/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/finetune/f0.1/s1",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "finetune",
8
+ "fraction": 0.1,
9
+ "seed": 1,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.05,
13
+ "weight_decay": 0.0005,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 450,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 45,
26
+ 45,
27
+ 45,
28
+ 45,
29
+ 45,
30
+ 45,
31
+ 45,
32
+ 45,
33
+ 45,
34
+ 45
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.494,
38
+ "val_macro_f1": 0.4925019110594578,
39
+ "val_loss": 3.339408203125,
40
+ "step": 500
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 1.0008782148361206,
46
+ "val_loss": 3.219765625,
47
+ "val_top1": 0.314,
48
+ "val_macro_f1": 0.27123795811247864,
49
+ "lr": 0.05,
50
+ "train_time_s": 9.458997052000086
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 0.3748807907104492,
55
+ "val_loss": 2.83027099609375,
56
+ "val_top1": 0.42,
57
+ "val_macro_f1": 0.4127956873669282,
58
+ "lr": 0.04852201922385564,
59
+ "train_time_s": 24.474821982999856
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 0.08123943954706192,
64
+ "val_loss": 4.577466796875,
65
+ "val_top1": 0.42,
66
+ "val_macro_f1": 0.414446287034932,
67
+ "lr": 0.04420708808983809,
68
+ "train_time_s": 39.64488251399962
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 0.020724279806017876,
73
+ "val_loss": 3.254974609375,
74
+ "val_top1": 0.468,
75
+ "val_macro_f1": 0.46220013311116476,
76
+ "lr": 0.03757549866575588,
77
+ "train_time_s": 54.81532114699985
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 0.005773973651230335,
82
+ "val_loss": 3.339408203125,
83
+ "val_top1": 0.494,
84
+ "val_macro_f1": 0.4925019110594578,
85
+ "lr": 0.029427118507989587,
86
+ "train_time_s": 69.20235361400046
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 0.0073205092921853065,
91
+ "val_loss": 3.25009375,
92
+ "val_top1": 0.472,
93
+ "val_macro_f1": 0.47493385085967554,
94
+ "lr": 0.020744762520849194,
95
+ "train_time_s": 76.77247189599984
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.003716814797371626,
100
+ "val_loss": 3.264702880859375,
101
+ "val_top1": 0.482,
102
+ "val_macro_f1": 0.48445411821421186,
103
+ "lr": 0.012575650974311119,
104
+ "train_time_s": 92.00777965499947
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.002123554702848196,
109
+ "val_loss": 3.29340673828125,
110
+ "val_top1": 0.482,
111
+ "val_macro_f1": 0.4836453666089384,
112
+ "lr": 0.005905099284133953,
113
+ "train_time_s": 107.35191676900013
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.0037158969789743423,
118
+ "val_loss": 3.308322265625,
119
+ "val_top1": 0.472,
120
+ "val_macro_f1": 0.47600220747562594,
121
+ "lr": 0.001537674431011102,
122
+ "train_time_s": 123.05702766200011
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.0023842889349907637,
127
+ "val_loss": 3.309263671875,
128
+ "val_top1": 0.474,
129
+ "val_macro_f1": 0.47641109038895546,
130
+ "lr": 1.5230855524017706e-07,
131
+ "train_time_s": 138.77393605400084
132
+ }
133
+ ],
134
+ "train_time_s": 138.77395815300133,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 919.21826171875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 11173962,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/finetune/f0.1/s1/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.47125,
3
+ "macro_f1": 0.46664404510136326,
4
+ "balanced_accuracy": 0.47125000000000006,
5
+ "per_class_accuracy": [
6
+ 0.56375,
7
+ 0.3825,
8
+ 0.7625,
9
+ 0.3925,
10
+ 0.4075,
11
+ 0.2125,
12
+ 0.50875,
13
+ 0.42625,
14
+ 0.6175,
15
+ 0.43875
16
+ ],
17
+ "per_class_f1": [
18
+ 0.5957727873183619,
19
+ 0.38107098381070986,
20
+ 0.6479022835900159,
21
+ 0.35182072829131655,
22
+ 0.40471756672873993,
23
+ 0.2607361963190184,
24
+ 0.4990803188228081,
25
+ 0.422029702970297,
26
+ 0.627700127064803,
27
+ 0.47560975609756095
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 451,
32
+ 31,
33
+ 71,
34
+ 12,
35
+ 17,
36
+ 4,
37
+ 6,
38
+ 15,
39
+ 109,
40
+ 84
41
+ ],
42
+ [
43
+ 58,
44
+ 306,
45
+ 36,
46
+ 120,
47
+ 83,
48
+ 50,
49
+ 41,
50
+ 69,
51
+ 23,
52
+ 14
53
+ ],
54
+ [
55
+ 27,
56
+ 5,
57
+ 610,
58
+ 8,
59
+ 5,
60
+ 11,
61
+ 12,
62
+ 4,
63
+ 28,
64
+ 90
65
+ ],
66
+ [
67
+ 13,
68
+ 75,
69
+ 25,
70
+ 314,
71
+ 98,
72
+ 64,
73
+ 60,
74
+ 128,
75
+ 8,
76
+ 15
77
+ ],
78
+ [
79
+ 15,
80
+ 77,
81
+ 16,
82
+ 142,
83
+ 326,
84
+ 53,
85
+ 100,
86
+ 62,
87
+ 4,
88
+ 5
89
+ ],
90
+ [
91
+ 27,
92
+ 139,
93
+ 13,
94
+ 132,
95
+ 76,
96
+ 170,
97
+ 121,
98
+ 110,
99
+ 4,
100
+ 8
101
+ ],
102
+ [
103
+ 14,
104
+ 68,
105
+ 20,
106
+ 36,
107
+ 105,
108
+ 83,
109
+ 407,
110
+ 53,
111
+ 1,
112
+ 13
113
+ ],
114
+ [
115
+ 5,
116
+ 84,
117
+ 7,
118
+ 174,
119
+ 67,
120
+ 50,
121
+ 52,
122
+ 341,
123
+ 9,
124
+ 11
125
+ ],
126
+ [
127
+ 72,
128
+ 7,
129
+ 59,
130
+ 34,
131
+ 19,
132
+ 9,
133
+ 6,
134
+ 15,
135
+ 494,
136
+ 85
137
+ ],
138
+ [
139
+ 32,
140
+ 14,
141
+ 226,
142
+ 13,
143
+ 15,
144
+ 10,
145
+ 26,
146
+ 19,
147
+ 94,
148
+ 351
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }
downstream/C128_l0.01_clean/finetune/f0.1/s1/test_predictions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:86ce06debafe1e8d18a2c14bd5642f1b3b0476ea7e9be5e99526021e7c636cd6
3
+ size 8128
downstream/C128_l0.01_clean/finetune/f0.1/s2/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/finetune/f0.1/s2",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "finetune",
8
+ "fraction": 0.1,
9
+ "seed": 2,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.05,
13
+ "weight_decay": 0.0005,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 450,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 45,
26
+ 45,
27
+ 45,
28
+ 45,
29
+ 45,
30
+ 45,
31
+ 45,
32
+ 45,
33
+ 45,
34
+ 45
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.502,
38
+ "val_macro_f1": 0.5025969863809201,
39
+ "val_loss": 3.207138671875,
40
+ "step": 500
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 0.7579313516616821,
46
+ "val_loss": 3.37785693359375,
47
+ "val_top1": 0.388,
48
+ "val_macro_f1": 0.36111624998462777,
49
+ "lr": 0.05,
50
+ "train_time_s": 16.186432792000232
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 0.3175930380821228,
55
+ "val_loss": 2.66659375,
56
+ "val_top1": 0.46,
57
+ "val_macro_f1": 0.4578665785490942,
58
+ "lr": 0.04852201922385564,
59
+ "train_time_s": 31.47894563800037
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 0.07602322101593018,
64
+ "val_loss": 3.0446904296875,
65
+ "val_top1": 0.482,
66
+ "val_macro_f1": 0.4834471697652252,
67
+ "lr": 0.04420708808983809,
68
+ "train_time_s": 46.758263552000244
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 0.01683623716235161,
73
+ "val_loss": 3.111078125,
74
+ "val_top1": 0.484,
75
+ "val_macro_f1": 0.48789790249900805,
76
+ "lr": 0.03757549866575588,
77
+ "train_time_s": 62.06071158600025
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 0.007751318626105785,
82
+ "val_loss": 3.207138671875,
83
+ "val_top1": 0.502,
84
+ "val_macro_f1": 0.5025969863809201,
85
+ "lr": 0.029427118507989587,
86
+ "train_time_s": 69.53512524800044
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 0.0019714594818651676,
91
+ "val_loss": 3.3405,
92
+ "val_top1": 0.486,
93
+ "val_macro_f1": 0.4863927881096918,
94
+ "lr": 0.020744762520849194,
95
+ "train_time_s": 83.47451420899961
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.00385871808975935,
100
+ "val_loss": 3.23151806640625,
101
+ "val_top1": 0.49,
102
+ "val_macro_f1": 0.49309516734223086,
103
+ "lr": 0.012575650974311119,
104
+ "train_time_s": 98.8063177920003
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.002976941177621484,
109
+ "val_loss": 3.264876953125,
110
+ "val_top1": 0.502,
111
+ "val_macro_f1": 0.5028797205781366,
112
+ "lr": 0.005905099284133953,
113
+ "train_time_s": 114.13058873400041
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.0015659760683774948,
118
+ "val_loss": 3.25403173828125,
119
+ "val_top1": 0.5,
120
+ "val_macro_f1": 0.5015497467989943,
121
+ "lr": 0.001537674431011102,
122
+ "train_time_s": 129.4627581450004
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.003472910961136222,
127
+ "val_loss": 3.27770068359375,
128
+ "val_top1": 0.496,
129
+ "val_macro_f1": 0.4962385696368175,
130
+ "lr": 1.5230855524017706e-07,
131
+ "train_time_s": 138.14243283400083
132
+ }
133
+ ],
134
+ "train_time_s": 138.14245812100035,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 919.21826171875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 11173962,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/finetune/f0.1/s2/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.477375,
3
+ "macro_f1": 0.4774554768330976,
4
+ "balanced_accuracy": 0.47737500000000005,
5
+ "per_class_accuracy": [
6
+ 0.48125,
7
+ 0.3775,
8
+ 0.64875,
9
+ 0.495,
10
+ 0.4375,
11
+ 0.30375,
12
+ 0.4225,
13
+ 0.4075,
14
+ 0.7325,
15
+ 0.4675
16
+ ],
17
+ "per_class_f1": [
18
+ 0.5720653789004457,
19
+ 0.4097693351424695,
20
+ 0.6352509179926561,
21
+ 0.3829787234042554,
22
+ 0.41791044776119407,
23
+ 0.32443257676902537,
24
+ 0.48493543758967,
25
+ 0.4343770819453697,
26
+ 0.6318059299191374,
27
+ 0.4810289389067524
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 385,
32
+ 38,
33
+ 39,
34
+ 29,
35
+ 10,
36
+ 20,
37
+ 10,
38
+ 7,
39
+ 172,
40
+ 90
41
+ ],
42
+ [
43
+ 42,
44
+ 302,
45
+ 28,
46
+ 165,
47
+ 83,
48
+ 58,
49
+ 25,
50
+ 59,
51
+ 28,
52
+ 10
53
+ ],
54
+ [
55
+ 25,
56
+ 5,
57
+ 519,
58
+ 15,
59
+ 3,
60
+ 11,
61
+ 10,
62
+ 4,
63
+ 59,
64
+ 149
65
+ ],
66
+ [
67
+ 3,
68
+ 57,
69
+ 15,
70
+ 396,
71
+ 98,
72
+ 76,
73
+ 33,
74
+ 93,
75
+ 15,
76
+ 14
77
+ ],
78
+ [
79
+ 8,
80
+ 60,
81
+ 10,
82
+ 197,
83
+ 350,
84
+ 60,
85
+ 53,
86
+ 44,
87
+ 9,
88
+ 9
89
+ ],
90
+ [
91
+ 11,
92
+ 86,
93
+ 12,
94
+ 155,
95
+ 94,
96
+ 243,
97
+ 74,
98
+ 105,
99
+ 13,
100
+ 7
101
+ ],
102
+ [
103
+ 9,
104
+ 52,
105
+ 22,
106
+ 42,
107
+ 140,
108
+ 129,
109
+ 338,
110
+ 47,
111
+ 4,
112
+ 17
113
+ ],
114
+ [
115
+ 2,
116
+ 60,
117
+ 6,
118
+ 219,
119
+ 72,
120
+ 66,
121
+ 33,
122
+ 326,
123
+ 8,
124
+ 8
125
+ ],
126
+ [
127
+ 41,
128
+ 5,
129
+ 22,
130
+ 28,
131
+ 14,
132
+ 19,
133
+ 3,
134
+ 5,
135
+ 586,
136
+ 77
137
+ ],
138
+ [
139
+ 20,
140
+ 9,
141
+ 161,
142
+ 22,
143
+ 11,
144
+ 16,
145
+ 15,
146
+ 11,
147
+ 161,
148
+ 374
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }
downstream/C128_l0.01_clean/finetune/f0.1/s2/test_predictions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1e9a925c714ab9717c9ac08a0b096b55db92ec67019b1b3b8ed48d9e85069fdb
3
+ size 8128
downstream/C128_l0.01_clean/finetune/f0.25/s0/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/finetune/f0.25/s0",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "finetune",
8
+ "fraction": 0.25,
9
+ "seed": 0,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.05,
13
+ "weight_decay": 0.0005,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 1120,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 112,
26
+ 112,
27
+ 112,
28
+ 112,
29
+ 112,
30
+ 112,
31
+ 112,
32
+ 112,
33
+ 112,
34
+ 112
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.598,
38
+ "val_macro_f1": 0.5998760388379384,
39
+ "val_loss": 2.135194091796875,
40
+ "step": 900
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 1.3392980098724365,
46
+ "val_loss": 1.75057177734375,
47
+ "val_top1": 0.398,
48
+ "val_macro_f1": 0.3794102903491342,
49
+ "lr": 0.05,
50
+ "train_time_s": 13.875871758999892
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 1.141497015953064,
55
+ "val_loss": 2.776034423828125,
56
+ "val_top1": 0.382,
57
+ "val_macro_f1": 0.3463041778154979,
58
+ "lr": 0.04852201922385564,
59
+ "train_time_s": 25.42991890399935
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 0.6150950789451599,
64
+ "val_loss": 1.634187255859375,
65
+ "val_top1": 0.506,
66
+ "val_macro_f1": 0.5032138941121254,
67
+ "lr": 0.04420708808983809,
68
+ "train_time_s": 40.74091223599953
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 0.4763893485069275,
73
+ "val_loss": 1.925493896484375,
74
+ "val_top1": 0.526,
75
+ "val_macro_f1": 0.5218058229986063,
76
+ "lr": 0.03757549866575588,
77
+ "train_time_s": 56.03826060099982
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 0.1384853571653366,
82
+ "val_loss": 2.185093505859375,
83
+ "val_top1": 0.534,
84
+ "val_macro_f1": 0.5198918068124869,
85
+ "lr": 0.029427118507989587,
86
+ "train_time_s": 71.33629051300068
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 0.2182987630367279,
91
+ "val_loss": 2.2829892578125,
92
+ "val_top1": 0.542,
93
+ "val_macro_f1": 0.5419744408880669,
94
+ "lr": 0.020744762520849194,
95
+ "train_time_s": 83.29302232500049
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.054394371807575226,
100
+ "val_loss": 2.20001708984375,
101
+ "val_top1": 0.55,
102
+ "val_macro_f1": 0.5496483877837706,
103
+ "lr": 0.012575650974311119,
104
+ "train_time_s": 92.19104595200042
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.03407266363501549,
109
+ "val_loss": 2.110118896484375,
110
+ "val_top1": 0.596,
111
+ "val_macro_f1": 0.5987605662397861,
112
+ "lr": 0.005905099284133953,
113
+ "train_time_s": 107.32956836800076
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.07010406255722046,
118
+ "val_loss": 2.135194091796875,
119
+ "val_top1": 0.598,
120
+ "val_macro_f1": 0.5998760388379384,
121
+ "lr": 0.001537674431011102,
122
+ "train_time_s": 122.61886956000126
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.030658934265375137,
127
+ "val_loss": 2.15734375,
128
+ "val_top1": 0.596,
129
+ "val_macro_f1": 0.597947050769364,
130
+ "lr": 1.5230855524017706e-07,
131
+ "train_time_s": 137.88322177400187
132
+ }
133
+ ],
134
+ "train_time_s": 137.88325717200223,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 919.21826171875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 11173962,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/finetune/f0.25/s0/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.580125,
3
+ "macro_f1": 0.5805694562484712,
4
+ "balanced_accuracy": 0.580125,
5
+ "per_class_accuracy": [
6
+ 0.6425,
7
+ 0.54,
8
+ 0.715,
9
+ 0.43625,
10
+ 0.52875,
11
+ 0.40875,
12
+ 0.63125,
13
+ 0.49,
14
+ 0.72875,
15
+ 0.68
16
+ ],
17
+ "per_class_f1": [
18
+ 0.6844207723035952,
19
+ 0.5356478611283324,
20
+ 0.7333333333333333,
21
+ 0.43327126008690253,
22
+ 0.5121065375302664,
23
+ 0.41603053435114506,
24
+ 0.630461922596754,
25
+ 0.501920614596671,
26
+ 0.7255756067205975,
27
+ 0.6329261198371147
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 514,
32
+ 50,
33
+ 46,
34
+ 16,
35
+ 6,
36
+ 13,
37
+ 6,
38
+ 4,
39
+ 83,
40
+ 62
41
+ ],
42
+ [
43
+ 36,
44
+ 432,
45
+ 13,
46
+ 101,
47
+ 51,
48
+ 55,
49
+ 23,
50
+ 68,
51
+ 15,
52
+ 6
53
+ ],
54
+ [
55
+ 22,
56
+ 7,
57
+ 572,
58
+ 11,
59
+ 5,
60
+ 4,
61
+ 4,
62
+ 6,
63
+ 17,
64
+ 152
65
+ ],
66
+ [
67
+ 5,
68
+ 65,
69
+ 7,
70
+ 349,
71
+ 113,
72
+ 95,
73
+ 36,
74
+ 100,
75
+ 22,
76
+ 8
77
+ ],
78
+ [
79
+ 7,
80
+ 53,
81
+ 5,
82
+ 108,
83
+ 423,
84
+ 78,
85
+ 71,
86
+ 38,
87
+ 3,
88
+ 14
89
+ ],
90
+ [
91
+ 5,
92
+ 63,
93
+ 7,
94
+ 89,
95
+ 86,
96
+ 327,
97
+ 104,
98
+ 101,
99
+ 3,
100
+ 15
101
+ ],
102
+ [
103
+ 5,
104
+ 24,
105
+ 11,
106
+ 13,
107
+ 87,
108
+ 84,
109
+ 505,
110
+ 46,
111
+ 0,
112
+ 25
113
+ ],
114
+ [
115
+ 1,
116
+ 105,
117
+ 3,
118
+ 104,
119
+ 64,
120
+ 91,
121
+ 32,
122
+ 392,
123
+ 2,
124
+ 6
125
+ ],
126
+ [
127
+ 82,
128
+ 4,
129
+ 13,
130
+ 7,
131
+ 10,
132
+ 10,
133
+ 3,
134
+ 1,
135
+ 583,
136
+ 87
137
+ ],
138
+ [
139
+ 25,
140
+ 10,
141
+ 83,
142
+ 13,
143
+ 7,
144
+ 15,
145
+ 18,
146
+ 6,
147
+ 79,
148
+ 544
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }
downstream/C128_l0.01_clean/finetune/f0.25/s0/test_predictions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:938ede000ef42c2bdd5a7bcba9a4e9be708361a332732c69947db49dd42d4ba7
3
+ size 8128
downstream/C128_l0.01_clean/finetune/f0.25/s1/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/finetune/f0.25/s1",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "finetune",
8
+ "fraction": 0.25,
9
+ "seed": 1,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.05,
13
+ "weight_decay": 0.0005,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 1120,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 112,
26
+ 112,
27
+ 112,
28
+ 112,
29
+ 112,
30
+ 112,
31
+ 112,
32
+ 112,
33
+ 112,
34
+ 112
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.602,
38
+ "val_macro_f1": 0.6017251026935571,
39
+ "val_loss": 2.01970458984375,
40
+ "step": 600
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 1.4362108707427979,
46
+ "val_loss": 2.222216796875,
47
+ "val_top1": 0.374,
48
+ "val_macro_f1": 0.3358232953356814,
49
+ "lr": 0.05,
50
+ "train_time_s": 19.609342612000546
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 1.1777256727218628,
55
+ "val_loss": 2.21171240234375,
56
+ "val_top1": 0.394,
57
+ "val_macro_f1": 0.40102444484093347,
58
+ "lr": 0.04852201922385564,
59
+ "train_time_s": 35.735606671000824
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 0.7664592266082764,
64
+ "val_loss": 2.16381298828125,
65
+ "val_top1": 0.47,
66
+ "val_macro_f1": 0.4447068897383433,
67
+ "lr": 0.04420708808983809,
68
+ "train_time_s": 51.64684962100091
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 0.5216617584228516,
73
+ "val_loss": 2.150870361328125,
74
+ "val_top1": 0.512,
75
+ "val_macro_f1": 0.5091161011180149,
76
+ "lr": 0.03757549866575588,
77
+ "train_time_s": 64.02580361200125
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 0.21346740424633026,
82
+ "val_loss": 2.360287109375,
83
+ "val_top1": 0.534,
84
+ "val_macro_f1": 0.5242415770259976,
85
+ "lr": 0.029427118507989587,
86
+ "train_time_s": 72.45979360600177
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 0.0608561709523201,
91
+ "val_loss": 2.01970458984375,
92
+ "val_top1": 0.602,
93
+ "val_macro_f1": 0.6017251026935571,
94
+ "lr": 0.020744762520849194,
95
+ "train_time_s": 88.30734562800171
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.05216239392757416,
100
+ "val_loss": 2.07039013671875,
101
+ "val_top1": 0.6,
102
+ "val_macro_f1": 0.5994901439895515,
103
+ "lr": 0.012575650974311119,
104
+ "train_time_s": 104.1786659820018
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.035959575325250626,
109
+ "val_loss": 2.081227294921875,
110
+ "val_top1": 0.592,
111
+ "val_macro_f1": 0.592498126589047,
112
+ "lr": 0.005905099284133953,
113
+ "train_time_s": 120.04553574000147
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.02596161514520645,
118
+ "val_loss": 2.0930361328125,
119
+ "val_top1": 0.596,
120
+ "val_macro_f1": 0.597218874943765,
121
+ "lr": 0.001537674431011102,
122
+ "train_time_s": 135.95104919500136
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.034508608281612396,
127
+ "val_loss": 2.121373291015625,
128
+ "val_top1": 0.592,
129
+ "val_macro_f1": 0.5928969586654399,
130
+ "lr": 1.5230855524017706e-07,
131
+ "train_time_s": 151.83161901300082
132
+ }
133
+ ],
134
+ "train_time_s": 151.83164037400093,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 919.21826171875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 11173962,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/finetune/f0.25/s1/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.56825,
3
+ "macro_f1": 0.5683876344620022,
4
+ "balanced_accuracy": 0.56825,
5
+ "per_class_accuracy": [
6
+ 0.68,
7
+ 0.51625,
8
+ 0.75875,
9
+ 0.4475,
10
+ 0.52625,
11
+ 0.3275,
12
+ 0.625,
13
+ 0.53875,
14
+ 0.665,
15
+ 0.5975
16
+ ],
17
+ "per_class_f1": [
18
+ 0.6695384615384616,
19
+ 0.5452145214521453,
20
+ 0.74845869297164,
21
+ 0.4264443120905301,
22
+ 0.4915353181552831,
23
+ 0.3692741367159972,
24
+ 0.6134969325153374,
25
+ 0.49341728677733254,
26
+ 0.711705685618729,
27
+ 0.614790996784566
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 544,
32
+ 35,
33
+ 47,
34
+ 16,
35
+ 10,
36
+ 6,
37
+ 9,
38
+ 7,
39
+ 63,
40
+ 63
41
+ ],
42
+ [
43
+ 49,
44
+ 413,
45
+ 7,
46
+ 103,
47
+ 65,
48
+ 43,
49
+ 23,
50
+ 86,
51
+ 9,
52
+ 2
53
+ ],
54
+ [
55
+ 32,
56
+ 7,
57
+ 607,
58
+ 13,
59
+ 11,
60
+ 8,
61
+ 12,
62
+ 11,
63
+ 13,
64
+ 86
65
+ ],
66
+ [
67
+ 7,
68
+ 52,
69
+ 6,
70
+ 358,
71
+ 122,
72
+ 89,
73
+ 26,
74
+ 121,
75
+ 9,
76
+ 10
77
+ ],
78
+ [
79
+ 17,
80
+ 46,
81
+ 4,
82
+ 122,
83
+ 421,
84
+ 47,
85
+ 71,
86
+ 65,
87
+ 1,
88
+ 6
89
+ ],
90
+ [
91
+ 10,
92
+ 64,
93
+ 8,
94
+ 106,
95
+ 73,
96
+ 262,
97
+ 121,
98
+ 148,
99
+ 1,
100
+ 7
101
+ ],
102
+ [
103
+ 3,
104
+ 16,
105
+ 6,
106
+ 24,
107
+ 109,
108
+ 70,
109
+ 500,
110
+ 57,
111
+ 0,
112
+ 15
113
+ ],
114
+ [
115
+ 2,
116
+ 72,
117
+ 2,
118
+ 105,
119
+ 71,
120
+ 74,
121
+ 40,
122
+ 431,
123
+ 2,
124
+ 1
125
+ ],
126
+ [
127
+ 122,
128
+ 3,
129
+ 14,
130
+ 13,
131
+ 13,
132
+ 12,
133
+ 1,
134
+ 3,
135
+ 532,
136
+ 87
137
+ ],
138
+ [
139
+ 39,
140
+ 7,
141
+ 121,
142
+ 19,
143
+ 18,
144
+ 8,
145
+ 27,
146
+ 18,
147
+ 65,
148
+ 478
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }
downstream/C128_l0.01_clean/finetune/f0.25/s1/test_predictions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:07c91a4d4f39297ee8bf9eee884c42d8cf78a49c1ac220aff6ff612b69b25823
3
+ size 8128
downstream/C128_l0.01_clean/finetune/f0.25/s2/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/finetune/f0.25/s2",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "finetune",
8
+ "fraction": 0.25,
9
+ "seed": 2,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.05,
13
+ "weight_decay": 0.0005,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 1120,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 112,
26
+ 112,
27
+ 112,
28
+ 112,
29
+ 112,
30
+ 112,
31
+ 112,
32
+ 112,
33
+ 112,
34
+ 112
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.576,
38
+ "val_macro_f1": 0.5786520195765992,
39
+ "val_loss": 2.20154296875,
40
+ "step": 900
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 1.5442334413528442,
46
+ "val_loss": 2.61542822265625,
47
+ "val_top1": 0.372,
48
+ "val_macro_f1": 0.3621421685829033,
49
+ "lr": 0.05,
50
+ "train_time_s": 19.220177342999705
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 1.0383472442626953,
55
+ "val_loss": 1.725662109375,
56
+ "val_top1": 0.47,
57
+ "val_macro_f1": 0.45914562498902856,
58
+ "lr": 0.04852201922385564,
59
+ "train_time_s": 34.93146596000042
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 0.6199432611465454,
64
+ "val_loss": 1.765404541015625,
65
+ "val_top1": 0.516,
66
+ "val_macro_f1": 0.5139317854460306,
67
+ "lr": 0.04420708808983809,
68
+ "train_time_s": 50.825228142000014
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 0.43570783734321594,
73
+ "val_loss": 2.000652587890625,
74
+ "val_top1": 0.492,
75
+ "val_macro_f1": 0.4891188032747644,
76
+ "lr": 0.03757549866575588,
77
+ "train_time_s": 60.61319253200054
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 0.16606317460536957,
82
+ "val_loss": 2.2829853515625,
83
+ "val_top1": 0.524,
84
+ "val_macro_f1": 0.5210168060562163,
85
+ "lr": 0.029427118507989587,
86
+ "train_time_s": 72.15283780900063
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 0.12014079093933105,
91
+ "val_loss": 2.1225771484375,
92
+ "val_top1": 0.558,
93
+ "val_macro_f1": 0.5623607846461238,
94
+ "lr": 0.020744762520849194,
95
+ "train_time_s": 87.47775533500135
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.060131266713142395,
100
+ "val_loss": 2.18406494140625,
101
+ "val_top1": 0.552,
102
+ "val_macro_f1": 0.55113508985014,
103
+ "lr": 0.012575650974311119,
104
+ "train_time_s": 103.10129001400128
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.03902437537908554,
109
+ "val_loss": 2.187182373046875,
110
+ "val_top1": 0.574,
111
+ "val_macro_f1": 0.5763493399156965,
112
+ "lr": 0.005905099284133953,
113
+ "train_time_s": 118.72207301100116
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.03008311055600643,
118
+ "val_loss": 2.20154296875,
119
+ "val_top1": 0.576,
120
+ "val_macro_f1": 0.5786520195765992,
121
+ "lr": 0.001537674431011102,
122
+ "train_time_s": 134.41589049400136
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.030429497361183167,
127
+ "val_loss": 2.2016474609375,
128
+ "val_top1": 0.574,
129
+ "val_macro_f1": 0.5762061342129027,
130
+ "lr": 1.5230855524017706e-07,
131
+ "train_time_s": 150.2329307300015
132
+ }
133
+ ],
134
+ "train_time_s": 150.23295239800154,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 919.21826171875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 11173962,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/finetune/f0.25/s2/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.578125,
3
+ "macro_f1": 0.5772926766495662,
4
+ "balanced_accuracy": 0.578125,
5
+ "per_class_accuracy": [
6
+ 0.68,
7
+ 0.5,
8
+ 0.7375,
9
+ 0.4325,
10
+ 0.55875,
11
+ 0.35375,
12
+ 0.62125,
13
+ 0.505,
14
+ 0.725,
15
+ 0.6675
16
+ ],
17
+ "per_class_f1": [
18
+ 0.6992287917737788,
19
+ 0.5298013245033112,
20
+ 0.73888541014402,
21
+ 0.43277048155096937,
22
+ 0.5293072824156305,
23
+ 0.36516129032258066,
24
+ 0.6132017273288094,
25
+ 0.5091367359798361,
26
+ 0.7178217821782179,
27
+ 0.6376119402985074
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 544,
32
+ 36,
33
+ 54,
34
+ 13,
35
+ 9,
36
+ 7,
37
+ 4,
38
+ 2,
39
+ 71,
40
+ 60
41
+ ],
42
+ [
43
+ 49,
44
+ 400,
45
+ 14,
46
+ 92,
47
+ 61,
48
+ 66,
49
+ 25,
50
+ 71,
51
+ 14,
52
+ 8
53
+ ],
54
+ [
55
+ 23,
56
+ 6,
57
+ 590,
58
+ 5,
59
+ 7,
60
+ 7,
61
+ 13,
62
+ 4,
63
+ 21,
64
+ 124
65
+ ],
66
+ [
67
+ 4,
68
+ 52,
69
+ 11,
70
+ 346,
71
+ 114,
72
+ 102,
73
+ 29,
74
+ 107,
75
+ 18,
76
+ 17
77
+ ],
78
+ [
79
+ 10,
80
+ 37,
81
+ 7,
82
+ 86,
83
+ 447,
84
+ 75,
85
+ 77,
86
+ 49,
87
+ 3,
88
+ 9
89
+ ],
90
+ [
91
+ 9,
92
+ 60,
93
+ 6,
94
+ 123,
95
+ 88,
96
+ 283,
97
+ 120,
98
+ 95,
99
+ 6,
100
+ 10
101
+ ],
102
+ [
103
+ 5,
104
+ 23,
105
+ 7,
106
+ 19,
107
+ 74,
108
+ 101,
109
+ 497,
110
+ 44,
111
+ 6,
112
+ 24
113
+ ],
114
+ [
115
+ 2,
116
+ 88,
117
+ 3,
118
+ 97,
119
+ 72,
120
+ 84,
121
+ 41,
122
+ 404,
123
+ 3,
124
+ 6
125
+ ],
126
+ [
127
+ 95,
128
+ 1,
129
+ 11,
130
+ 6,
131
+ 9,
132
+ 13,
133
+ 0,
134
+ 2,
135
+ 580,
136
+ 83
137
+ ],
138
+ [
139
+ 15,
140
+ 7,
141
+ 94,
142
+ 12,
143
+ 8,
144
+ 12,
145
+ 15,
146
+ 9,
147
+ 94,
148
+ 534
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }
downstream/C128_l0.01_clean/finetune/f0.25/s2/test_predictions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:38c54a64fee7f29fcd31d7a982d80f0bd47f1181aa4ebb5f7238fd820af507d3
3
+ size 8128
downstream/C128_l0.01_clean/finetune/f0.5/s0/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/finetune/f0.5/s0",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "finetune",
8
+ "fraction": 0.5,
9
+ "seed": 0,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.05,
13
+ "weight_decay": 0.0005,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 2250,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 225,
26
+ 225,
27
+ 225,
28
+ 225,
29
+ 225,
30
+ 225,
31
+ 225,
32
+ 225,
33
+ 225,
34
+ 225
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.622,
38
+ "val_macro_f1": 0.6221782460581855,
39
+ "val_loss": 1.4661429443359375,
40
+ "step": 900
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 1.664527177810669,
46
+ "val_loss": 2.424224853515625,
47
+ "val_top1": 0.356,
48
+ "val_macro_f1": 0.3157327227925132,
49
+ "lr": 0.05,
50
+ "train_time_s": 13.696915329999683
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 1.5986084938049316,
55
+ "val_loss": 2.0173961181640623,
56
+ "val_top1": 0.462,
57
+ "val_macro_f1": 0.45025898459899505,
58
+ "lr": 0.04852201922385564,
59
+ "train_time_s": 26.329687538000144
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 1.1938261985778809,
64
+ "val_loss": 1.327635986328125,
65
+ "val_top1": 0.566,
66
+ "val_macro_f1": 0.5651982959513828,
67
+ "lr": 0.04420708808983809,
68
+ "train_time_s": 41.73735346399917
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 0.82810378074646,
73
+ "val_loss": 1.659606689453125,
74
+ "val_top1": 0.558,
75
+ "val_macro_f1": 0.5471666373938,
76
+ "lr": 0.03757549866575588,
77
+ "train_time_s": 57.16610283099908
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 0.6535406112670898,
82
+ "val_loss": 1.462862548828125,
83
+ "val_top1": 0.564,
84
+ "val_macro_f1": 0.5654383710369502,
85
+ "lr": 0.029427118507989587,
86
+ "train_time_s": 72.5963499319987
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 0.5393353700637817,
91
+ "val_loss": 1.3466689453125,
92
+ "val_top1": 0.604,
93
+ "val_macro_f1": 0.6025902562435809,
94
+ "lr": 0.020744762520849194,
95
+ "train_time_s": 83.5285831809997
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.3642338514328003,
100
+ "val_loss": 1.458321533203125,
101
+ "val_top1": 0.604,
102
+ "val_macro_f1": 0.6021349547539999,
103
+ "lr": 0.012575650974311119,
104
+ "train_time_s": 94.85576707600012
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.2942779064178467,
109
+ "val_loss": 1.4757919921875,
110
+ "val_top1": 0.608,
111
+ "val_macro_f1": 0.608161160305464,
112
+ "lr": 0.005905099284133953,
113
+ "train_time_s": 110.15946771900053
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.22133029997348785,
118
+ "val_loss": 1.4661429443359375,
119
+ "val_top1": 0.622,
120
+ "val_macro_f1": 0.6221782460581855,
121
+ "lr": 0.001537674431011102,
122
+ "train_time_s": 125.53320325899949
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.27174630761146545,
127
+ "val_loss": 1.4711282958984375,
128
+ "val_top1": 0.616,
129
+ "val_macro_f1": 0.6158800037264107,
130
+ "lr": 1.5230855524017706e-07,
131
+ "train_time_s": 140.93479338399993
132
+ }
133
+ ],
134
+ "train_time_s": 140.9348156470005,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 919.21826171875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 11173962,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/finetune/f0.5/s0/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.637625,
3
+ "macro_f1": 0.6380527603646357,
4
+ "balanced_accuracy": 0.637625,
5
+ "per_class_accuracy": [
6
+ 0.76,
7
+ 0.59875,
8
+ 0.8225,
9
+ 0.48125,
10
+ 0.56125,
11
+ 0.46,
12
+ 0.65375,
13
+ 0.55,
14
+ 0.7875,
15
+ 0.70125
16
+ ],
17
+ "per_class_f1": [
18
+ 0.7755102040816326,
19
+ 0.5998747651847213,
20
+ 0.8058787507654622,
21
+ 0.47648514851485146,
22
+ 0.5647798742138365,
23
+ 0.45885286783042395,
24
+ 0.6492861576660459,
25
+ 0.5422057917436846,
26
+ 0.7875,
27
+ 0.7201540436456997
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 608,
32
+ 41,
33
+ 34,
34
+ 11,
35
+ 5,
36
+ 7,
37
+ 5,
38
+ 4,
39
+ 51,
40
+ 34
41
+ ],
42
+ [
43
+ 37,
44
+ 479,
45
+ 5,
46
+ 70,
47
+ 42,
48
+ 59,
49
+ 8,
50
+ 88,
51
+ 10,
52
+ 2
53
+ ],
54
+ [
55
+ 27,
56
+ 5,
57
+ 658,
58
+ 8,
59
+ 4,
60
+ 2,
61
+ 11,
62
+ 3,
63
+ 12,
64
+ 70
65
+ ],
66
+ [
67
+ 0,
68
+ 65,
69
+ 5,
70
+ 385,
71
+ 98,
72
+ 114,
73
+ 29,
74
+ 85,
75
+ 13,
76
+ 6
77
+ ],
78
+ [
79
+ 4,
80
+ 45,
81
+ 6,
82
+ 84,
83
+ 449,
84
+ 56,
85
+ 96,
86
+ 54,
87
+ 2,
88
+ 4
89
+ ],
90
+ [
91
+ 5,
92
+ 56,
93
+ 5,
94
+ 96,
95
+ 69,
96
+ 368,
97
+ 95,
98
+ 102,
99
+ 0,
100
+ 4
101
+ ],
102
+ [
103
+ 7,
104
+ 18,
105
+ 11,
106
+ 33,
107
+ 58,
108
+ 95,
109
+ 523,
110
+ 41,
111
+ 1,
112
+ 13
113
+ ],
114
+ [
115
+ 1,
116
+ 77,
117
+ 1,
118
+ 99,
119
+ 51,
120
+ 93,
121
+ 30,
122
+ 440,
123
+ 5,
124
+ 3
125
+ ],
126
+ [
127
+ 57,
128
+ 3,
129
+ 24,
130
+ 11,
131
+ 6,
132
+ 6,
133
+ 2,
134
+ 0,
135
+ 630,
136
+ 61
137
+ ],
138
+ [
139
+ 22,
140
+ 8,
141
+ 84,
142
+ 19,
143
+ 8,
144
+ 4,
145
+ 12,
146
+ 6,
147
+ 76,
148
+ 561
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }
downstream/C128_l0.01_clean/finetune/f0.5/s0/test_predictions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:31f206cf3c6b09fa712902c87a6948e3d41231bda09a60b0c9f8a5bea1327582
3
+ size 8128
downstream/C128_l0.01_clean/finetune/f0.5/s1/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/finetune/f0.5/s1",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "finetune",
8
+ "fraction": 0.5,
9
+ "seed": 1,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.05,
13
+ "weight_decay": 0.0005,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 2250,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 225,
26
+ 225,
27
+ 225,
28
+ 225,
29
+ 225,
30
+ 225,
31
+ 225,
32
+ 225,
33
+ 225,
34
+ 225
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.628,
38
+ "val_macro_f1": 0.6290187976498306,
39
+ "val_loss": 1.4078359375,
40
+ "step": 1000
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 1.5371099710464478,
46
+ "val_loss": 1.9343111572265625,
47
+ "val_top1": 0.38,
48
+ "val_macro_f1": 0.34719261175023763,
49
+ "lr": 0.05,
50
+ "train_time_s": 17.724598741999216
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 1.4742543697357178,
55
+ "val_loss": 2.013494384765625,
56
+ "val_top1": 0.398,
57
+ "val_macro_f1": 0.38683685463967044,
58
+ "lr": 0.04852201922385564,
59
+ "train_time_s": 33.16876373800005
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 1.1212689876556396,
64
+ "val_loss": 2.0129852294921875,
65
+ "val_top1": 0.496,
66
+ "val_macro_f1": 0.4750149901623363,
67
+ "lr": 0.04420708808983809,
68
+ "train_time_s": 48.674035547999665
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 1.0422449111938477,
73
+ "val_loss": 1.465039306640625,
74
+ "val_top1": 0.55,
75
+ "val_macro_f1": 0.5492188153620395,
76
+ "lr": 0.03757549866575588,
77
+ "train_time_s": 63.14872954599923
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 0.6011922359466553,
82
+ "val_loss": 1.606311279296875,
83
+ "val_top1": 0.556,
84
+ "val_macro_f1": 0.5525847675468278,
85
+ "lr": 0.029427118507989587,
86
+ "train_time_s": 71.42782198299938
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 0.4771849811077118,
91
+ "val_loss": 1.453379150390625,
92
+ "val_top1": 0.614,
93
+ "val_macro_f1": 0.6147878781222331,
94
+ "lr": 0.020744762520849194,
95
+ "train_time_s": 86.53574524399846
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.31056541204452515,
100
+ "val_loss": 1.4402371826171876,
101
+ "val_top1": 0.61,
102
+ "val_macro_f1": 0.6097286753865624,
103
+ "lr": 0.012575650974311119,
104
+ "train_time_s": 101.8722211159984
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.3110702633857727,
109
+ "val_loss": 1.4650087890625,
110
+ "val_top1": 0.614,
111
+ "val_macro_f1": 0.6126294338684682,
112
+ "lr": 0.005905099284133953,
113
+ "train_time_s": 117.23309545699885
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.19435670971870422,
118
+ "val_loss": 1.410713134765625,
119
+ "val_top1": 0.626,
120
+ "val_macro_f1": 0.6278351080382449,
121
+ "lr": 0.001537674431011102,
122
+ "train_time_s": 132.58969802799948
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.2845712900161743,
127
+ "val_loss": 1.4078359375,
128
+ "val_top1": 0.628,
129
+ "val_macro_f1": 0.6290187976498306,
130
+ "lr": 1.5230855524017706e-07,
131
+ "train_time_s": 139.95889154199904
132
+ }
133
+ ],
134
+ "train_time_s": 139.95896696899945,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 919.21826171875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 11173962,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/finetune/f0.5/s1/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.6485,
3
+ "macro_f1": 0.6498532055749948,
4
+ "balanced_accuracy": 0.6485000000000001,
5
+ "per_class_accuracy": [
6
+ 0.75,
7
+ 0.595,
8
+ 0.79625,
9
+ 0.4925,
10
+ 0.59875,
11
+ 0.5075,
12
+ 0.65625,
13
+ 0.54875,
14
+ 0.79375,
15
+ 0.74625
16
+ ],
17
+ "per_class_f1": [
18
+ 0.7766990291262137,
19
+ 0.6002522068095838,
20
+ 0.8125000000000001,
21
+ 0.4876237623762376,
22
+ 0.59875,
23
+ 0.4839094159713945,
24
+ 0.6649778340721976,
25
+ 0.5567533291058973,
26
+ 0.7947434292866082,
27
+ 0.7223230490018149
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 600,
32
+ 45,
33
+ 29,
34
+ 10,
35
+ 6,
36
+ 3,
37
+ 5,
38
+ 1,
39
+ 54,
40
+ 47
41
+ ],
42
+ [
43
+ 38,
44
+ 476,
45
+ 4,
46
+ 75,
47
+ 44,
48
+ 56,
49
+ 18,
50
+ 78,
51
+ 5,
52
+ 6
53
+ ],
54
+ [
55
+ 20,
56
+ 8,
57
+ 637,
58
+ 9,
59
+ 6,
60
+ 5,
61
+ 7,
62
+ 2,
63
+ 18,
64
+ 88
65
+ ],
66
+ [
67
+ 2,
68
+ 57,
69
+ 3,
70
+ 394,
71
+ 95,
72
+ 117,
73
+ 24,
74
+ 90,
75
+ 8,
76
+ 10
77
+ ],
78
+ [
79
+ 5,
80
+ 43,
81
+ 4,
82
+ 77,
83
+ 479,
84
+ 62,
85
+ 74,
86
+ 45,
87
+ 4,
88
+ 7
89
+ ],
90
+ [
91
+ 4,
92
+ 52,
93
+ 3,
94
+ 104,
95
+ 54,
96
+ 406,
97
+ 81,
98
+ 88,
99
+ 1,
100
+ 7
101
+ ],
102
+ [
103
+ 4,
104
+ 13,
105
+ 10,
106
+ 34,
107
+ 58,
108
+ 111,
109
+ 525,
110
+ 30,
111
+ 2,
112
+ 13
113
+ ],
114
+ [
115
+ 1,
116
+ 82,
117
+ 0,
118
+ 86,
119
+ 42,
120
+ 106,
121
+ 34,
122
+ 439,
123
+ 3,
124
+ 7
125
+ ],
126
+ [
127
+ 49,
128
+ 5,
129
+ 15,
130
+ 12,
131
+ 6,
132
+ 5,
133
+ 2,
134
+ 0,
135
+ 635,
136
+ 71
137
+ ],
138
+ [
139
+ 22,
140
+ 5,
141
+ 63,
142
+ 15,
143
+ 10,
144
+ 7,
145
+ 9,
146
+ 4,
147
+ 68,
148
+ 597
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }
downstream/C128_l0.01_clean/finetune/f0.5/s1/test_predictions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0cb71928ae9b9001a2cf580177d03b08c11439b91f6a2ed48ab3876e10a7cb43
3
+ size 8128
downstream/C128_l0.01_clean/finetune/f0.5/s2/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/finetune/f0.5/s2",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "finetune",
8
+ "fraction": 0.5,
9
+ "seed": 2,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.05,
13
+ "weight_decay": 0.0005,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 2250,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 225,
26
+ 225,
27
+ 225,
28
+ 225,
29
+ 225,
30
+ 225,
31
+ 225,
32
+ 225,
33
+ 225,
34
+ 225
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.626,
38
+ "val_macro_f1": 0.6276225599970873,
39
+ "val_loss": 1.3873631591796876,
40
+ "step": 900
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 1.3954614400863647,
46
+ "val_loss": 11.521220703125,
47
+ "val_top1": 0.228,
48
+ "val_macro_f1": 0.1732283469704522,
49
+ "lr": 0.05,
50
+ "train_time_s": 19.041560211000615
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 1.3203586339950562,
55
+ "val_loss": 1.5385164794921875,
56
+ "val_top1": 0.472,
57
+ "val_macro_f1": 0.4589835647364133,
58
+ "lr": 0.04852201922385564,
59
+ "train_time_s": 34.76120481599901
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 1.2003684043884277,
64
+ "val_loss": 1.5034833984375,
65
+ "val_top1": 0.506,
66
+ "val_macro_f1": 0.5040322670482625,
67
+ "lr": 0.04420708808983809,
68
+ "train_time_s": 50.47000436499911
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 0.8662247061729431,
73
+ "val_loss": 1.551482177734375,
74
+ "val_top1": 0.546,
75
+ "val_macro_f1": 0.5403759746792641,
76
+ "lr": 0.03757549866575588,
77
+ "train_time_s": 66.18782085699786
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 0.7043923139572144,
82
+ "val_loss": 1.3412440185546874,
83
+ "val_top1": 0.602,
84
+ "val_macro_f1": 0.6002351284074914,
85
+ "lr": 0.029427118507989587,
86
+ "train_time_s": 81.75911541899768
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 0.5357675552368164,
91
+ "val_loss": 1.3644307861328124,
92
+ "val_top1": 0.604,
93
+ "val_macro_f1": 0.6020021276453285,
94
+ "lr": 0.020744762520849194,
95
+ "train_time_s": 97.51658064399817
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.4202096164226532,
100
+ "val_loss": 1.408474365234375,
101
+ "val_top1": 0.618,
102
+ "val_macro_f1": 0.6235743879997069,
103
+ "lr": 0.012575650974311119,
104
+ "train_time_s": 113.302075976997
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.3027457594871521,
109
+ "val_loss": 1.46721044921875,
110
+ "val_top1": 0.594,
111
+ "val_macro_f1": 0.5913408386008079,
112
+ "lr": 0.005905099284133953,
113
+ "train_time_s": 129.11843697699624
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.18171077966690063,
118
+ "val_loss": 1.3873631591796876,
119
+ "val_top1": 0.626,
120
+ "val_macro_f1": 0.6276225599970873,
121
+ "lr": 0.001537674431011102,
122
+ "train_time_s": 144.9325570669953
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.2305491864681244,
127
+ "val_loss": 1.37533984375,
128
+ "val_top1": 0.622,
129
+ "val_macro_f1": 0.6238061395799843,
130
+ "lr": 1.5230855524017706e-07,
131
+ "train_time_s": 160.6571598759947
132
+ }
133
+ ],
134
+ "train_time_s": 160.65719096399516,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 919.21826171875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 11173962,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/finetune/f0.5/s2/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.6375,
3
+ "macro_f1": 0.6388758991094365,
4
+ "balanced_accuracy": 0.6375000000000001,
5
+ "per_class_accuracy": [
6
+ 0.7425,
7
+ 0.59,
8
+ 0.7875,
9
+ 0.505,
10
+ 0.545,
11
+ 0.48625,
12
+ 0.66,
13
+ 0.5425,
14
+ 0.78875,
15
+ 0.7275
16
+ ],
17
+ "per_class_f1": [
18
+ 0.7754569190600522,
19
+ 0.5940843297671492,
20
+ 0.7989854153455929,
21
+ 0.4867469879518072,
22
+ 0.5647668393782385,
23
+ 0.4669867947178871,
24
+ 0.6579439252336449,
25
+ 0.5521628498727735,
26
+ 0.7853142501555693,
27
+ 0.7063106796116505
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 594,
32
+ 35,
33
+ 32,
34
+ 12,
35
+ 7,
36
+ 4,
37
+ 1,
38
+ 4,
39
+ 61,
40
+ 50
41
+ ],
42
+ [
43
+ 34,
44
+ 472,
45
+ 5,
46
+ 79,
47
+ 39,
48
+ 61,
49
+ 17,
50
+ 78,
51
+ 12,
52
+ 3
53
+ ],
54
+ [
55
+ 21,
56
+ 9,
57
+ 630,
58
+ 8,
59
+ 3,
60
+ 7,
61
+ 7,
62
+ 5,
63
+ 21,
64
+ 89
65
+ ],
66
+ [
67
+ 1,
68
+ 59,
69
+ 3,
70
+ 404,
71
+ 89,
72
+ 121,
73
+ 29,
74
+ 77,
75
+ 7,
76
+ 10
77
+ ],
78
+ [
79
+ 4,
80
+ 54,
81
+ 5,
82
+ 87,
83
+ 436,
84
+ 65,
85
+ 95,
86
+ 43,
87
+ 1,
88
+ 10
89
+ ],
90
+ [
91
+ 3,
92
+ 48,
93
+ 5,
94
+ 120,
95
+ 58,
96
+ 389,
97
+ 85,
98
+ 89,
99
+ 0,
100
+ 3
101
+ ],
102
+ [
103
+ 2,
104
+ 16,
105
+ 4,
106
+ 33,
107
+ 50,
108
+ 110,
109
+ 528,
110
+ 35,
111
+ 3,
112
+ 19
113
+ ],
114
+ [
115
+ 1,
116
+ 83,
117
+ 4,
118
+ 95,
119
+ 46,
120
+ 92,
121
+ 36,
122
+ 434,
123
+ 1,
124
+ 8
125
+ ],
126
+ [
127
+ 58,
128
+ 4,
129
+ 12,
130
+ 8,
131
+ 5,
132
+ 7,
133
+ 1,
134
+ 0,
135
+ 631,
136
+ 74
137
+ ],
138
+ [
139
+ 14,
140
+ 9,
141
+ 77,
142
+ 14,
143
+ 11,
144
+ 10,
145
+ 6,
146
+ 7,
147
+ 70,
148
+ 582
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }
downstream/C128_l0.01_clean/finetune/f0.5/s2/test_predictions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1cc424b0ec4e3abe939358fae155cb32455b99473538ef78d9ddd8d5f1da8ca2
3
+ size 8128
downstream/C128_l0.01_clean/finetune/f1/s0/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/finetune/f1/s0",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "finetune",
8
+ "fraction": 1.0,
9
+ "seed": 0,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.05,
13
+ "weight_decay": 0.0005,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 4500,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 450,
26
+ 450,
27
+ 450,
28
+ 450,
29
+ 450,
30
+ 450,
31
+ 450,
32
+ 450,
33
+ 450,
34
+ 450
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.646,
38
+ "val_macro_f1": 0.6431995595043855,
39
+ "val_loss": 1.113109619140625,
40
+ "step": 800
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 1.4402191638946533,
46
+ "val_loss": 48.434125,
47
+ "val_top1": 0.262,
48
+ "val_macro_f1": 0.231486794513879,
49
+ "lr": 0.05,
50
+ "train_time_s": 19.595666684999742
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 1.2260222434997559,
55
+ "val_loss": 1.630965087890625,
56
+ "val_top1": 0.466,
57
+ "val_macro_f1": 0.44737279290425763,
58
+ "lr": 0.04852201922385564,
59
+ "train_time_s": 35.5338824570008
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 1.4344639778137207,
64
+ "val_loss": 2.4450634765625,
65
+ "val_top1": 0.444,
66
+ "val_macro_f1": 0.44565782839920426,
67
+ "lr": 0.04420708808983809,
68
+ "train_time_s": 48.08675563500037
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 1.1334424018859863,
73
+ "val_loss": 1.74398583984375,
74
+ "val_top1": 0.464,
75
+ "val_macro_f1": 0.4501137889652932,
76
+ "lr": 0.03757549866575588,
77
+ "train_time_s": 57.57589853300124
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 1.0889031887054443,
82
+ "val_loss": 1.6507138671875,
83
+ "val_top1": 0.55,
84
+ "val_macro_f1": 0.5460326036460827,
85
+ "lr": 0.029427118507989587,
86
+ "train_time_s": 73.21386474400242
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 0.842443585395813,
91
+ "val_loss": 1.354664306640625,
92
+ "val_top1": 0.582,
93
+ "val_macro_f1": 0.579455241868269,
94
+ "lr": 0.020744762520849194,
95
+ "train_time_s": 89.11956499400185
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.7784000039100647,
100
+ "val_loss": 1.1626893310546875,
101
+ "val_top1": 0.624,
102
+ "val_macro_f1": 0.6217312015547114,
103
+ "lr": 0.012575650974311119,
104
+ "train_time_s": 105.07884173900129
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.503559410572052,
109
+ "val_loss": 1.113109619140625,
110
+ "val_top1": 0.646,
111
+ "val_macro_f1": 0.6431995595043855,
112
+ "lr": 0.005905099284133953,
113
+ "train_time_s": 121.17437448900091
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.6135572195053101,
118
+ "val_loss": 1.0885562744140624,
119
+ "val_top1": 0.646,
120
+ "val_macro_f1": 0.6467807014854703,
121
+ "lr": 0.001537674431011102,
122
+ "train_time_s": 137.24749545600025
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.6150004863739014,
127
+ "val_loss": 1.0763377685546875,
128
+ "val_top1": 0.636,
129
+ "val_macro_f1": 0.6373144693143017,
130
+ "lr": 1.5230855524017706e-07,
131
+ "train_time_s": 153.18861528900015
132
+ }
133
+ ],
134
+ "train_time_s": 153.18864026600022,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 919.21826171875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 11173962,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/finetune/f1/s0/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.670125,
3
+ "macro_f1": 0.6650833752104608,
4
+ "balanced_accuracy": 0.6701250000000001,
5
+ "per_class_accuracy": [
6
+ 0.81375,
7
+ 0.6575,
8
+ 0.85,
9
+ 0.4375,
10
+ 0.62625,
11
+ 0.4325,
12
+ 0.75875,
13
+ 0.605,
14
+ 0.8475,
15
+ 0.6725
16
+ ],
17
+ "per_class_f1": [
18
+ 0.7871825876662635,
19
+ 0.6162858816637375,
20
+ 0.8104886769964244,
21
+ 0.4946996466431095,
22
+ 0.6317780580075663,
23
+ 0.4890459363957597,
24
+ 0.6949055523755009,
25
+ 0.587378640776699,
26
+ 0.8085867620751341,
27
+ 0.7304820095044127
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 651,
32
+ 27,
33
+ 41,
34
+ 2,
35
+ 3,
36
+ 0,
37
+ 4,
38
+ 1,
39
+ 57,
40
+ 14
41
+ ],
42
+ [
43
+ 53,
44
+ 526,
45
+ 10,
46
+ 54,
47
+ 36,
48
+ 25,
49
+ 17,
50
+ 72,
51
+ 5,
52
+ 2
53
+ ],
54
+ [
55
+ 30,
56
+ 5,
57
+ 680,
58
+ 4,
59
+ 5,
60
+ 1,
61
+ 4,
62
+ 3,
63
+ 11,
64
+ 57
65
+ ],
66
+ [
67
+ 3,
68
+ 96,
69
+ 8,
70
+ 350,
71
+ 107,
72
+ 75,
73
+ 37,
74
+ 104,
75
+ 15,
76
+ 5
77
+ ],
78
+ [
79
+ 5,
80
+ 46,
81
+ 5,
82
+ 50,
83
+ 501,
84
+ 34,
85
+ 100,
86
+ 52,
87
+ 4,
88
+ 3
89
+ ],
90
+ [
91
+ 9,
92
+ 87,
93
+ 4,
94
+ 69,
95
+ 50,
96
+ 346,
97
+ 133,
98
+ 95,
99
+ 2,
100
+ 5
101
+ ],
102
+ [
103
+ 10,
104
+ 15,
105
+ 15,
106
+ 11,
107
+ 33,
108
+ 59,
109
+ 607,
110
+ 33,
111
+ 4,
112
+ 13
113
+ ],
114
+ [
115
+ 2,
116
+ 91,
117
+ 3,
118
+ 58,
119
+ 47,
120
+ 71,
121
+ 35,
122
+ 484,
123
+ 4,
124
+ 5
125
+ ],
126
+ [
127
+ 58,
128
+ 4,
129
+ 20,
130
+ 6,
131
+ 1,
132
+ 1,
133
+ 0,
134
+ 1,
135
+ 678,
136
+ 31
137
+ ],
138
+ [
139
+ 33,
140
+ 10,
141
+ 92,
142
+ 11,
143
+ 3,
144
+ 3,
145
+ 10,
146
+ 3,
147
+ 97,
148
+ 538
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }
downstream/C128_l0.01_clean/finetune/f1/s0/test_predictions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3e047749c7aa2d64fa91ff4972086f796c9bdab14faf90e170aca429f37a6371
3
+ size 8128
downstream/C128_l0.01_clean/finetune/f1/s1/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/finetune/f1/s1",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "finetune",
8
+ "fraction": 1.0,
9
+ "seed": 1,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.05,
13
+ "weight_decay": 0.0005,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 4500,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 450,
26
+ 450,
27
+ 450,
28
+ 450,
29
+ 450,
30
+ 450,
31
+ 450,
32
+ 450,
33
+ 450,
34
+ 450
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.642,
38
+ "val_macro_f1": 0.6426350724197716,
39
+ "val_loss": 1.2152886962890626,
40
+ "step": 1000
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 1.6809144020080566,
46
+ "val_loss": 3.02781201171875,
47
+ "val_top1": 0.432,
48
+ "val_macro_f1": 0.4200789625529183,
49
+ "lr": 0.05,
50
+ "train_time_s": 13.070318083999155
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 1.3151838779449463,
55
+ "val_loss": 1.76627294921875,
56
+ "val_top1": 0.418,
57
+ "val_macro_f1": 0.40380787884380676,
58
+ "lr": 0.04852201922385564,
59
+ "train_time_s": 28.248113009000008
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 1.5031670331954956,
64
+ "val_loss": 1.421893798828125,
65
+ "val_top1": 0.514,
66
+ "val_macro_f1": 0.512205624776953,
67
+ "lr": 0.04420708808983809,
68
+ "train_time_s": 43.45445638899946
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 1.0106605291366577,
73
+ "val_loss": 1.6025562744140625,
74
+ "val_top1": 0.492,
75
+ "val_macro_f1": 0.48392156936379677,
76
+ "lr": 0.03757549866575588,
77
+ "train_time_s": 58.65733413899943
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 0.9204467535018921,
82
+ "val_loss": 1.7309908447265625,
83
+ "val_top1": 0.558,
84
+ "val_macro_f1": 0.554980999984315,
85
+ "lr": 0.029427118507989587,
86
+ "train_time_s": 69.13762343000053
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 0.821257472038269,
91
+ "val_loss": 1.3000107421875,
92
+ "val_top1": 0.602,
93
+ "val_macro_f1": 0.598506191312312,
94
+ "lr": 0.020744762520849194,
95
+ "train_time_s": 80.36585809700046
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.8413910269737244,
100
+ "val_loss": 1.399740234375,
101
+ "val_top1": 0.6,
102
+ "val_macro_f1": 0.5977736028782317,
103
+ "lr": 0.012575650974311119,
104
+ "train_time_s": 95.57113237200065
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.6839640736579895,
109
+ "val_loss": 1.21922119140625,
110
+ "val_top1": 0.632,
111
+ "val_macro_f1": 0.6320365787142193,
112
+ "lr": 0.005905099284133953,
113
+ "train_time_s": 110.78489923300003
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.5734846591949463,
118
+ "val_loss": 1.227885009765625,
119
+ "val_top1": 0.634,
120
+ "val_macro_f1": 0.6355198963290737,
121
+ "lr": 0.001537674431011102,
122
+ "train_time_s": 125.99263506399984
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.4888577461242676,
127
+ "val_loss": 1.2152886962890626,
128
+ "val_top1": 0.642,
129
+ "val_macro_f1": 0.6426350724197716,
130
+ "lr": 1.5230855524017706e-07,
131
+ "train_time_s": 137.19366119000006
132
+ }
133
+ ],
134
+ "train_time_s": 137.19369314100004,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 919.21826171875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 11173962,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/finetune/f1/s1/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.68525,
3
+ "macro_f1": 0.6856048905682238,
4
+ "balanced_accuracy": 0.6852500000000001,
5
+ "per_class_accuracy": [
6
+ 0.79375,
7
+ 0.63875,
8
+ 0.80625,
9
+ 0.52125,
10
+ 0.6425,
11
+ 0.50625,
12
+ 0.7125,
13
+ 0.63,
14
+ 0.8375,
15
+ 0.76375
16
+ ],
17
+ "per_class_f1": [
18
+ 0.8017676767676769,
19
+ 0.6666666666666667,
20
+ 0.8258642765685019,
21
+ 0.5258511979823455,
22
+ 0.638112973308504,
23
+ 0.5049875311720697,
24
+ 0.7201516108654453,
25
+ 0.6021505376344086,
26
+ 0.8230958230958232,
27
+ 0.7474006116207952
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 635,
32
+ 27,
33
+ 28,
34
+ 6,
35
+ 8,
36
+ 2,
37
+ 4,
38
+ 1,
39
+ 58,
40
+ 31
41
+ ],
42
+ [
43
+ 38,
44
+ 511,
45
+ 6,
46
+ 69,
47
+ 33,
48
+ 46,
49
+ 14,
50
+ 74,
51
+ 4,
52
+ 5
53
+ ],
54
+ [
55
+ 19,
56
+ 10,
57
+ 645,
58
+ 8,
59
+ 5,
60
+ 9,
61
+ 3,
62
+ 4,
63
+ 10,
64
+ 87
65
+ ],
66
+ [
67
+ 4,
68
+ 47,
69
+ 2,
70
+ 417,
71
+ 98,
72
+ 97,
73
+ 15,
74
+ 101,
75
+ 11,
76
+ 8
77
+ ],
78
+ [
79
+ 7,
80
+ 29,
81
+ 3,
82
+ 64,
83
+ 514,
84
+ 52,
85
+ 69,
86
+ 49,
87
+ 5,
88
+ 8
89
+ ],
90
+ [
91
+ 4,
92
+ 41,
93
+ 2,
94
+ 99,
95
+ 55,
96
+ 405,
97
+ 84,
98
+ 103,
99
+ 0,
100
+ 7
101
+ ],
102
+ [
103
+ 5,
104
+ 5,
105
+ 5,
106
+ 24,
107
+ 47,
108
+ 90,
109
+ 570,
110
+ 34,
111
+ 1,
112
+ 19
113
+ ],
114
+ [
115
+ 1,
116
+ 51,
117
+ 0,
118
+ 84,
119
+ 42,
120
+ 94,
121
+ 16,
122
+ 504,
123
+ 1,
124
+ 7
125
+ ],
126
+ [
127
+ 45,
128
+ 5,
129
+ 13,
130
+ 8,
131
+ 3,
132
+ 4,
133
+ 0,
134
+ 0,
135
+ 670,
136
+ 52
137
+ ],
138
+ [
139
+ 26,
140
+ 7,
141
+ 58,
142
+ 7,
143
+ 6,
144
+ 5,
145
+ 8,
146
+ 4,
147
+ 68,
148
+ 611
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }
downstream/C128_l0.01_clean/finetune/f1/s1/test_predictions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ff52c6215405eec245b1a676867ca5650220925db76ca941b7459aa1508963dd
3
+ size 8128
downstream/C128_l0.01_clean/finetune/f1/s2/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/finetune/f1/s2",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "finetune",
8
+ "fraction": 1.0,
9
+ "seed": 2,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.05,
13
+ "weight_decay": 0.0005,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 4500,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 450,
26
+ 450,
27
+ 450,
28
+ 450,
29
+ 450,
30
+ 450,
31
+ 450,
32
+ 450,
33
+ 450,
34
+ 450
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.654,
38
+ "val_macro_f1": 0.6561774190023232,
39
+ "val_loss": 1.059621337890625,
40
+ "step": 900
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 1.9473944902420044,
46
+ "val_loss": 1.667942138671875,
47
+ "val_top1": 0.404,
48
+ "val_macro_f1": 0.3820966213343172,
49
+ "lr": 0.05,
50
+ "train_time_s": 19.184310279000783
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 1.2435941696166992,
55
+ "val_loss": 1.682259765625,
56
+ "val_top1": 0.446,
57
+ "val_macro_f1": 0.43524970170722366,
58
+ "lr": 0.04852201922385564,
59
+ "train_time_s": 34.93623902900072
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 1.1772289276123047,
64
+ "val_loss": 1.382130126953125,
65
+ "val_top1": 0.522,
66
+ "val_macro_f1": 0.5141435742222427,
67
+ "lr": 0.04420708808983809,
68
+ "train_time_s": 50.70281876199988
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 1.0206830501556396,
73
+ "val_loss": 1.425110595703125,
74
+ "val_top1": 0.52,
75
+ "val_macro_f1": 0.5193152477283001,
76
+ "lr": 0.03757549866575588,
77
+ "train_time_s": 66.33444605299883
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 0.8035917282104492,
82
+ "val_loss": 1.31982373046875,
83
+ "val_top1": 0.546,
84
+ "val_macro_f1": 0.5512592005581908,
85
+ "lr": 0.029427118507989587,
86
+ "train_time_s": 82.10266869699808
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 0.9900924563407898,
91
+ "val_loss": 1.175665283203125,
92
+ "val_top1": 0.606,
93
+ "val_macro_f1": 0.6087787723079152,
94
+ "lr": 0.020744762520849194,
95
+ "train_time_s": 97.8688987209971
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.731718897819519,
100
+ "val_loss": 1.1156395263671874,
101
+ "val_top1": 0.63,
102
+ "val_macro_f1": 0.6317331233146192,
103
+ "lr": 0.012575650974311119,
104
+ "train_time_s": 113.62670028299726
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.8999506831169128,
109
+ "val_loss": 1.0977762451171875,
110
+ "val_top1": 0.65,
111
+ "val_macro_f1": 0.6514253995334529,
112
+ "lr": 0.005905099284133953,
113
+ "train_time_s": 129.24092872699657
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.5546733140945435,
118
+ "val_loss": 1.059621337890625,
119
+ "val_top1": 0.654,
120
+ "val_macro_f1": 0.6561774190023232,
121
+ "lr": 0.001537674431011102,
122
+ "train_time_s": 144.99802787099725
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.5416405200958252,
127
+ "val_loss": 1.062933837890625,
128
+ "val_top1": 0.65,
129
+ "val_macro_f1": 0.6515320532329333,
130
+ "lr": 1.5230855524017706e-07,
131
+ "train_time_s": 160.61895952599662
132
+ }
133
+ ],
134
+ "train_time_s": 160.6189987589969,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 919.21826171875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 11173962,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/finetune/f1/s2/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.681625,
3
+ "macro_f1": 0.6818435286504656,
4
+ "balanced_accuracy": 0.6816249999999999,
5
+ "per_class_accuracy": [
6
+ 0.77375,
7
+ 0.655,
8
+ 0.84125,
9
+ 0.54375,
10
+ 0.5825,
11
+ 0.51625,
12
+ 0.69875,
13
+ 0.63375,
14
+ 0.8375,
15
+ 0.73375
16
+ ],
17
+ "per_class_f1": [
18
+ 0.7951188182402055,
19
+ 0.6562304320601127,
20
+ 0.8303516347933374,
21
+ 0.5256797583081572,
22
+ 0.6188579017264276,
23
+ 0.5247776365946634,
24
+ 0.7089410272669626,
25
+ 0.6007109004739336,
26
+ 0.8170731707317073,
27
+ 0.7406940063091482
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 619,
32
+ 28,
33
+ 34,
34
+ 12,
35
+ 4,
36
+ 1,
37
+ 6,
38
+ 1,
39
+ 62,
40
+ 33
41
+ ],
42
+ [
43
+ 39,
44
+ 524,
45
+ 7,
46
+ 73,
47
+ 29,
48
+ 35,
49
+ 11,
50
+ 73,
51
+ 5,
52
+ 4
53
+ ],
54
+ [
55
+ 25,
56
+ 4,
57
+ 673,
58
+ 7,
59
+ 3,
60
+ 6,
61
+ 3,
62
+ 3,
63
+ 11,
64
+ 65
65
+ ],
66
+ [
67
+ 2,
68
+ 65,
69
+ 5,
70
+ 435,
71
+ 75,
72
+ 95,
73
+ 14,
74
+ 98,
75
+ 8,
76
+ 3
77
+ ],
78
+ [
79
+ 6,
80
+ 41,
81
+ 4,
82
+ 93,
83
+ 466,
84
+ 43,
85
+ 87,
86
+ 52,
87
+ 2,
88
+ 6
89
+ ],
90
+ [
91
+ 1,
92
+ 48,
93
+ 3,
94
+ 101,
95
+ 45,
96
+ 413,
97
+ 78,
98
+ 106,
99
+ 3,
100
+ 2
101
+ ],
102
+ [
103
+ 4,
104
+ 10,
105
+ 7,
106
+ 27,
107
+ 41,
108
+ 91,
109
+ 559,
110
+ 44,
111
+ 0,
112
+ 17
113
+ ],
114
+ [
115
+ 1,
116
+ 65,
117
+ 2,
118
+ 82,
119
+ 36,
120
+ 85,
121
+ 13,
122
+ 507,
123
+ 1,
124
+ 8
125
+ ],
126
+ [
127
+ 38,
128
+ 4,
129
+ 15,
130
+ 8,
131
+ 3,
132
+ 2,
133
+ 0,
134
+ 0,
135
+ 670,
136
+ 60
137
+ ],
138
+ [
139
+ 22,
140
+ 8,
141
+ 71,
142
+ 17,
143
+ 4,
144
+ 3,
145
+ 6,
146
+ 4,
147
+ 78,
148
+ 587
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }
downstream/C128_l0.01_clean/finetune/f1/s2/test_predictions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d806bd66ce8780ef9f61164d9856c732aecd0e8cee2011e42ad4b56cd2b58a23
3
+ size 8128
downstream/C128_l0.01_clean/frozen/f0.1/s0/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/frozen/f0.1/s0",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "frozen",
8
+ "fraction": 0.1,
9
+ "seed": 0,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.1,
13
+ "weight_decay": 0.0,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 450,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 45,
26
+ 45,
27
+ 45,
28
+ 45,
29
+ 45,
30
+ 45,
31
+ 45,
32
+ 45,
33
+ 45,
34
+ 45
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.44,
38
+ "val_macro_f1": 0.44009913036814596,
39
+ "val_loss": 1.845666015625,
40
+ "step": 1000
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 1.3843863010406494,
46
+ "val_loss": 1.85146240234375,
47
+ "val_top1": 0.416,
48
+ "val_macro_f1": 0.41094460579587794,
49
+ "lr": 0.1,
50
+ "train_time_s": 6.155195332000403
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 1.1762166023254395,
55
+ "val_loss": 1.85185205078125,
56
+ "val_top1": 0.432,
57
+ "val_macro_f1": 0.4283816465787057,
58
+ "lr": 0.09704403844771128,
59
+ "train_time_s": 12.091626259000805
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 1.4299083948135376,
64
+ "val_loss": 2.0896416015625,
65
+ "val_top1": 0.42,
66
+ "val_macro_f1": 0.4143021564356443,
67
+ "lr": 0.08841417617967617,
68
+ "train_time_s": 17.748211211001035
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 1.2334506511688232,
73
+ "val_loss": 1.940521484375,
74
+ "val_top1": 0.41,
75
+ "val_macro_f1": 0.40913836255529956,
76
+ "lr": 0.07515099733151176,
77
+ "train_time_s": 23.817515388001084
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 1.1678426265716553,
82
+ "val_loss": 1.970739990234375,
83
+ "val_top1": 0.42,
84
+ "val_macro_f1": 0.41561868648255046,
85
+ "lr": 0.05885423701597917,
86
+ "train_time_s": 29.884542739001517
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 1.0756964683532715,
91
+ "val_loss": 1.8877965087890625,
92
+ "val_top1": 0.428,
93
+ "val_macro_f1": 0.4264721259716658,
94
+ "lr": 0.04148952504169839,
95
+ "train_time_s": 35.79857658900164
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.8500015735626221,
100
+ "val_loss": 1.874067626953125,
101
+ "val_top1": 0.426,
102
+ "val_macro_f1": 0.42714003184264404,
103
+ "lr": 0.025151301948622237,
104
+ "train_time_s": 41.87412414300161
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.9025719165802002,
109
+ "val_loss": 1.8612298583984375,
110
+ "val_top1": 0.424,
111
+ "val_macro_f1": 0.42130096464073963,
112
+ "lr": 0.011810198568267906,
113
+ "train_time_s": 47.843332691001706
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.8374037146568298,
118
+ "val_loss": 1.831461669921875,
119
+ "val_top1": 0.434,
120
+ "val_macro_f1": 0.4353035632787007,
121
+ "lr": 0.003075348862022204,
122
+ "train_time_s": 53.8075837710021
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.8359435796737671,
127
+ "val_loss": 1.845666015625,
128
+ "val_top1": 0.44,
129
+ "val_macro_f1": 0.44009913036814596,
130
+ "lr": 3.046171104803541e-07,
131
+ "train_time_s": 59.79376875100206
132
+ }
133
+ ],
134
+ "train_time_s": 59.79379628400238,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 832.591796875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 5130,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/frozen/f0.1/s0/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.44375,
3
+ "macro_f1": 0.4421897116384993,
4
+ "balanced_accuracy": 0.44375,
5
+ "per_class_accuracy": [
6
+ 0.54375,
7
+ 0.385,
8
+ 0.60125,
9
+ 0.37,
10
+ 0.37875,
11
+ 0.21,
12
+ 0.47375,
13
+ 0.43,
14
+ 0.59375,
15
+ 0.45125
16
+ ],
17
+ "per_class_f1": [
18
+ 0.5788423153692615,
19
+ 0.3786109403810694,
20
+ 0.5909090909090908,
21
+ 0.3400344629523262,
22
+ 0.3997361477572559,
23
+ 0.24068767908309452,
24
+ 0.45552884615384615,
25
+ 0.41001191895113226,
26
+ 0.5740181268882175,
27
+ 0.45351758793969854
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 435,
32
+ 62,
33
+ 44,
34
+ 9,
35
+ 6,
36
+ 3,
37
+ 14,
38
+ 10,
39
+ 138,
40
+ 79
41
+ ],
42
+ [
43
+ 59,
44
+ 308,
45
+ 24,
46
+ 118,
47
+ 59,
48
+ 51,
49
+ 50,
50
+ 101,
51
+ 19,
52
+ 11
53
+ ],
54
+ [
55
+ 31,
56
+ 10,
57
+ 481,
58
+ 10,
59
+ 0,
60
+ 18,
61
+ 19,
62
+ 10,
63
+ 47,
64
+ 174
65
+ ],
66
+ [
67
+ 12,
68
+ 79,
69
+ 24,
70
+ 296,
71
+ 92,
72
+ 86,
73
+ 69,
74
+ 117,
75
+ 18,
76
+ 7
77
+ ],
78
+ [
79
+ 7,
80
+ 82,
81
+ 9,
82
+ 176,
83
+ 303,
84
+ 54,
85
+ 79,
86
+ 69,
87
+ 5,
88
+ 16
89
+ ],
90
+ [
91
+ 16,
92
+ 131,
93
+ 14,
94
+ 107,
95
+ 87,
96
+ 168,
97
+ 139,
98
+ 123,
99
+ 5,
100
+ 10
101
+ ],
102
+ [
103
+ 8,
104
+ 45,
105
+ 15,
106
+ 54,
107
+ 87,
108
+ 110,
109
+ 379,
110
+ 72,
111
+ 5,
112
+ 25
113
+ ],
114
+ [
115
+ 3,
116
+ 75,
117
+ 7,
118
+ 144,
119
+ 68,
120
+ 54,
121
+ 96,
122
+ 344,
123
+ 3,
124
+ 6
125
+ ],
126
+ [
127
+ 102,
128
+ 18,
129
+ 45,
130
+ 15,
131
+ 8,
132
+ 20,
133
+ 3,
134
+ 11,
135
+ 475,
136
+ 103
137
+ ],
138
+ [
139
+ 30,
140
+ 17,
141
+ 165,
142
+ 12,
143
+ 6,
144
+ 32,
145
+ 16,
146
+ 21,
147
+ 140,
148
+ 361
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }
downstream/C128_l0.01_clean/frozen/f0.1/s0/test_predictions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fbe5f5f73f0dfebe7a1744f37d8a5968432a46fd6130b6f620eaf4119a9f5909
3
+ size 8128
downstream/C128_l0.01_clean/frozen/f0.1/s1/metrics.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "out_dir": "/kaggle/working/OSAE/runs_fast/downstream/C128_l0.01_clean/frozen/f0.1/s1",
4
+ "data_dir": "/tmp/stl10",
5
+ "init": "pretrained",
6
+ "pretrained": "/kaggle/working/OSAE/runs_fast/pretrain/C128_l0.01_clean_ps0/encoder.pt",
7
+ "protocol": "frozen",
8
+ "fraction": 0.1,
9
+ "seed": 1,
10
+ "steps": 1000,
11
+ "batch_size": 128,
12
+ "lr": 0.1,
13
+ "weight_decay": 0.0,
14
+ "momentum": 0.9,
15
+ "warmup_steps": 100,
16
+ "label_smoothing": 0.0,
17
+ "eval_every": 100,
18
+ "patience": 0,
19
+ "amp": "auto",
20
+ "deterministic": false
21
+ },
22
+ "n_train": 450,
23
+ "n_val": 500,
24
+ "train_class_counts": [
25
+ 45,
26
+ 45,
27
+ 45,
28
+ 45,
29
+ 45,
30
+ 45,
31
+ 45,
32
+ 45,
33
+ 45,
34
+ 45
35
+ ],
36
+ "best": {
37
+ "val_top1": 0.446,
38
+ "val_macro_f1": 0.4453593226150893,
39
+ "val_loss": 1.8388778076171874,
40
+ "step": 800
41
+ },
42
+ "history": [
43
+ {
44
+ "step": 100,
45
+ "train_loss": 2.4660844802856445,
46
+ "val_loss": 1.86965625,
47
+ "val_top1": 0.414,
48
+ "val_macro_f1": 0.4079415310510881,
49
+ "lr": 0.1,
50
+ "train_time_s": 7.266183196999918
51
+ },
52
+ {
53
+ "step": 200,
54
+ "train_loss": 1.7698860168457031,
55
+ "val_loss": 1.946380615234375,
56
+ "val_top1": 0.408,
57
+ "val_macro_f1": 0.40676453298966575,
58
+ "lr": 0.09704403844771128,
59
+ "train_time_s": 13.105111814999873
60
+ },
61
+ {
62
+ "step": 300,
63
+ "train_loss": 1.4437757730484009,
64
+ "val_loss": 1.9902896728515624,
65
+ "val_top1": 0.416,
66
+ "val_macro_f1": 0.4106488661073076,
67
+ "lr": 0.08841417617967617,
68
+ "train_time_s": 18.812408579000476
69
+ },
70
+ {
71
+ "step": 400,
72
+ "train_loss": 1.1257678270339966,
73
+ "val_loss": 1.956866455078125,
74
+ "val_top1": 0.438,
75
+ "val_macro_f1": 0.437109893095986,
76
+ "lr": 0.07515099733151176,
77
+ "train_time_s": 24.661498603000837
78
+ },
79
+ {
80
+ "step": 500,
81
+ "train_loss": 0.95513516664505,
82
+ "val_loss": 2.0018597412109376,
83
+ "val_top1": 0.424,
84
+ "val_macro_f1": 0.4172086181453885,
85
+ "lr": 0.05885423701597917,
86
+ "train_time_s": 30.518781046000186
87
+ },
88
+ {
89
+ "step": 600,
90
+ "train_loss": 0.9779168367385864,
91
+ "val_loss": 1.9543453369140624,
92
+ "val_top1": 0.428,
93
+ "val_macro_f1": 0.42845451438749854,
94
+ "lr": 0.04148952504169839,
95
+ "train_time_s": 36.381413778000024
96
+ },
97
+ {
98
+ "step": 700,
99
+ "train_loss": 0.982588529586792,
100
+ "val_loss": 1.9029217529296876,
101
+ "val_top1": 0.432,
102
+ "val_macro_f1": 0.4296990464012257,
103
+ "lr": 0.025151301948622237,
104
+ "train_time_s": 42.23041101600029
105
+ },
106
+ {
107
+ "step": 800,
108
+ "train_loss": 0.9052144289016724,
109
+ "val_loss": 1.8388778076171874,
110
+ "val_top1": 0.446,
111
+ "val_macro_f1": 0.4453593226150893,
112
+ "lr": 0.011810198568267906,
113
+ "train_time_s": 47.970417442000326
114
+ },
115
+ {
116
+ "step": 900,
117
+ "train_loss": 0.8832823038101196,
118
+ "val_loss": 1.860576171875,
119
+ "val_top1": 0.444,
120
+ "val_macro_f1": 0.44384931618871476,
121
+ "lr": 0.003075348862022204,
122
+ "train_time_s": 53.84162593600013
123
+ },
124
+ {
125
+ "step": 1000,
126
+ "train_loss": 0.8680331707000732,
127
+ "val_loss": 1.84340185546875,
128
+ "val_top1": 0.44,
129
+ "val_macro_f1": 0.43927184072709924,
130
+ "lr": 3.046171104803541e-07,
131
+ "train_time_s": 59.67688776700015
132
+ }
133
+ ],
134
+ "train_time_s": 59.676914024000325,
135
+ "amp_resolved": "fp16",
136
+ "steps_run": 1000,
137
+ "peak_gpu_mem_mb": 832.591796875,
138
+ "params_total": 11173962,
139
+ "params_trainable": 5130,
140
+ "env": {
141
+ "python": "3.13.16",
142
+ "platform": "Linux-6.18.55+-x86_64-with-glibc2.39",
143
+ "torch": "2.11.0+cu130",
144
+ "cuda_available": true,
145
+ "cudnn": 92700,
146
+ "gpu": "Tesla T4",
147
+ "cuda": "13.0"
148
+ }
149
+ }
downstream/C128_l0.01_clean/frozen/f0.1/s1/test_metrics.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "top1": 0.43975,
3
+ "macro_f1": 0.4375152966060824,
4
+ "balanced_accuracy": 0.43975,
5
+ "per_class_accuracy": [
6
+ 0.57,
7
+ 0.3575,
8
+ 0.60375,
9
+ 0.355,
10
+ 0.37125,
11
+ 0.1875,
12
+ 0.46875,
13
+ 0.4425,
14
+ 0.58625,
15
+ 0.455
16
+ ],
17
+ "per_class_f1": [
18
+ 0.6003949967083607,
19
+ 0.3620253164556962,
20
+ 0.5926380368098159,
21
+ 0.3317757009345795,
22
+ 0.38101347017318793,
23
+ 0.21961932650073207,
24
+ 0.44937088076692633,
25
+ 0.40877598152424943,
26
+ 0.5751072961373391,
27
+ 0.4544319600499376
28
+ ],
29
+ "confusion_matrix": [
30
+ [
31
+ 456,
32
+ 47,
33
+ 41,
34
+ 11,
35
+ 10,
36
+ 8,
37
+ 13,
38
+ 5,
39
+ 127,
40
+ 82
41
+ ],
42
+ [
43
+ 66,
44
+ 286,
45
+ 20,
46
+ 115,
47
+ 61,
48
+ 52,
49
+ 52,
50
+ 120,
51
+ 16,
52
+ 12
53
+ ],
54
+ [
55
+ 31,
56
+ 8,
57
+ 483,
58
+ 12,
59
+ 2,
60
+ 14,
61
+ 24,
62
+ 8,
63
+ 48,
64
+ 170
65
+ ],
66
+ [
67
+ 12,
68
+ 80,
69
+ 23,
70
+ 284,
71
+ 100,
72
+ 78,
73
+ 74,
74
+ 117,
75
+ 20,
76
+ 12
77
+ ],
78
+ [
79
+ 10,
80
+ 85,
81
+ 10,
82
+ 164,
83
+ 297,
84
+ 51,
85
+ 78,
86
+ 84,
87
+ 5,
88
+ 16
89
+ ],
90
+ [
91
+ 10,
92
+ 124,
93
+ 12,
94
+ 114,
95
+ 107,
96
+ 150,
97
+ 140,
98
+ 129,
99
+ 5,
100
+ 9
101
+ ],
102
+ [
103
+ 5,
104
+ 41,
105
+ 13,
106
+ 51,
107
+ 99,
108
+ 104,
109
+ 375,
110
+ 81,
111
+ 5,
112
+ 26
113
+ ],
114
+ [
115
+ 0,
116
+ 78,
117
+ 8,
118
+ 129,
119
+ 70,
120
+ 54,
121
+ 95,
122
+ 354,
123
+ 3,
124
+ 9
125
+ ],
126
+ [
127
+ 101,
128
+ 18,
129
+ 45,
130
+ 17,
131
+ 7,
132
+ 29,
133
+ 2,
134
+ 10,
135
+ 469,
136
+ 102
137
+ ],
138
+ [
139
+ 28,
140
+ 13,
141
+ 175,
142
+ 15,
143
+ 6,
144
+ 26,
145
+ 16,
146
+ 24,
147
+ 133,
148
+ 364
149
+ ]
150
+ ],
151
+ "n": 8000,
152
+ "inference": {
153
+ "params_total": 11173962,
154
+ "amp": "auto",
155
+ "benchmark": {
156
+ "bs1": {
157
+ "latency_ms_median": 3.0639289998362074,
158
+ "latency_ms_p90": 3.2141113002580823,
159
+ "throughput_img_s": 326.3783201417064,
160
+ "peak_gpu_mem_mb": 76.3740234375
161
+ },
162
+ "bs256": {
163
+ "latency_ms_median": 44.95981049967668,
164
+ "latency_ms_p90": 45.17436420137528,
165
+ "throughput_img_s": 5693.974177267517,
166
+ "peak_gpu_mem_mb": 441.84228515625
167
+ }
168
+ }
169
+ }
170
+ }