Card: the fixed split is read by fixed_split_eval.py of the code repository
Browse files
README.md
CHANGED
|
@@ -163,11 +163,23 @@ python contrast_experiment/w9/eval/k10_eval.py table scores.json
|
|
| 163 |
```
|
| 164 |
|
| 165 |
`cells` takes the 162 five-fold checkpoints of `rank_ckpt_plan.json` and `i2bce_sel_pool.json`;
|
| 166 |
-
`--only <regex>` keeps a subset, for example `--only "_d512$"` for the 512-sentence runs.
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 171 |
|
| 172 |
## The 512-sentence packs and the `_d512` runs
|
| 173 |
|
|
@@ -214,6 +226,8 @@ dataset keeps the first packs as `wscan_gal_rev_rawprefix.npz`; the workers read
|
|
| 214 |
|
| 215 |
## Changes
|
| 216 |
|
|
|
|
|
|
|
| 217 |
- 2026-11-11, later that day. 62 checkpoints were added, so that every tower the paper reads is
|
| 218 |
here: the DINO-style control (fold 0), eight fixed-split runs at their selected epochs (CE and
|
| 219 |
I-CE trained at 2,048 sentences, the SimCLR-style tower, I + SimCLR, a+u @ views and the three
|
|
|
|
| 163 |
```
|
| 164 |
|
| 165 |
`cells` takes the 162 five-fold checkpoints of `rank_ckpt_plan.json` and `i2bce_sel_pool.json`;
|
| 166 |
+
`--only <regex>` keeps a subset, for example `--only "_d512$"` for the 512-sentence runs. In its
|
| 167 |
+
table the five families of the main comparison go by the paper's names (`I-CE`, `CE`, `SimCLR`,
|
| 168 |
+
`VICReg`, `BYOL`) and every other family by its plan key without the fold; a family with one
|
| 169 |
+
fold, the DINO-style control, is printed with its spread over the ten draws.
|
| 170 |
+
|
| 171 |
+
The fixed split has one command of its own:
|
| 172 |
+
|
| 173 |
+
```bash
|
| 174 |
+
python contrast_experiment/w9/eval/fixed_split_eval.py --ckpt-dir checkpoints
|
| 175 |
+
```
|
| 176 |
+
|
| 177 |
+
It reads every run of `fixed_split_ckpt_plan.json` at its selected checkpoint and prints each
|
| 178 |
+
reading beside the number the paper prints: Table 3, the gallery-aimed settings of Section 6.2
|
| 179 |
+
and Table A7 on the one 4,096-sentence anchor draw of the fixed split, the SimCLR-style test
|
| 180 |
+
curve of Appendix L, the latent queries of Figure A1, and, through `k10_eval.py`, the towers of
|
| 181 |
+
Figure 9 against the ten draws. A near tie can move a cell by one query of 204 (0.005), and the
|
| 182 |
+
collapsed a+u @ gallery row by about 0.01.
|
| 183 |
|
| 184 |
## The 512-sentence packs and the `_d512` runs
|
| 185 |
|
|
|
|
| 226 |
|
| 227 |
## Changes
|
| 228 |
|
| 229 |
+
- 2026-11-11, at night. The code repository gained `fixed_split_eval.py`, which reads the
|
| 230 |
+
fixed-split checkpoints as the paper does.
|
| 231 |
- 2026-11-11, later that day. 62 checkpoints were added, so that every tower the paper reads is
|
| 232 |
here: the DINO-style control (fold 0), eight fixed-split runs at their selected epochs (CE and
|
| 233 |
I-CE trained at 2,048 sentences, the SimCLR-style tower, I + SimCLR, a+u @ views and the three
|