diff options
| -rw-r--r-- | RANK_RESULTS.md | 31 | ||||
| -rw-r--r-- | logs/obj_large_gpu.log | 2 | ||||
| -rw-r--r-- | logs/omit_rest.log | 0 |
3 files changed, 32 insertions, 1 deletions
diff --git a/RANK_RESULTS.md b/RANK_RESULTS.md index 0e481b9..2998e47 100644 --- a/RANK_RESULTS.md +++ b/RANK_RESULTS.md @@ -118,6 +118,37 @@ directions and then stops. Enlarging the scene population does not help either scenes, so the ceiling is a property of what the two corpora are about rather than of how many scenes are sampled. +## Corpus overlap sets the width, by intervention + +The invariance results say the width does not come from the encoders. A +controlled intervention says where it does come from. In the synthetic world +the caption is encoded into separate factor blocks, so a block can be +suppressed — the caption then never states that property while vision +continues to see it. Nothing else changes: same images, same segmentation, +same kernel. + +| Caption content | ρ | shared | recovery | +|---|---|---|---| +| states everything | 0.928 | 20 | **96.4%** | +| never states size | 0.830 | 15 | **5.6%** | + +Deleting one word class costs five shared directions and the entire result. +The correlation is still 0.830 — higher than anything we have achieved on +photographs — and recovery is at chance. **The synthetic world degraded to +Visual Genome's shared width fails exactly the way Visual Genome fails**, at a +correlation Visual Genome never reaches. + +This closes the causal chain. How much of the same world the two corpora +describe sets the width of the shared spectrum; the width sets recovery; and +the correlation reports on neither reliably. It also reframes the natural-data +deficit as a statement about Visual Genome rather than about our pipeline: a +region description simply does not say as much about a photograph as the +synthetic captions say about their scenes. + +(Widths in this table are computed on standardised fields and are not +comparable entry-for-entry with the tables above, which use raw fields; within +the table the convention is identical, which is what the comparison needs.) + ## Two smaller results, one refuted hypothesis **The projection improved.** Scaling each discriminant direction by its own diff --git a/logs/obj_large_gpu.log b/logs/obj_large_gpu.log index 748f6bd..8dfb4fe 100644 --- a/logs/obj_large_gpu.log +++ b/logs/obj_large_gpu.log @@ -1,3 +1,3 @@ Using a slow image processor as `use_fast` is unset and a slow processor was saved with this model. `use_fast=True` will be the default behavior in v4.52, even if the model was saved with a slow processor. This will result in minor differences in outputs. You'll still be able to use a slow processor with `use_fast=False`. `torch_dtype` is deprecated! Use `dtype` instead! -
segment: 0%| | 0/250 [00:00<?, ?it/s]
segment: 0%| | 1/250 [00:03<14:05, 3.39s/it]
segment: 1%| | 2/250 [00:06<12:26, 3.01s/it]
segment: 1%| | 3/250 [00:08<11:50, 2.88s/it]
segment: 2%|▏ | 4/250 [00:11<11:34, 2.82s/it]
segment: 2%|▏ | 5/250 [00:14<11:25, 2.80s/it]
segment: 2%|▏ | 6/250 [00:17<11:16, 2.77s/it]
segment: 3%|▎ | 7/250 [00:19<11:09, 2.76s/it]
segment: 3%|▎ | 8/250 [00:22<11:08, 2.76s/it]
segment: 4%|▎ | 9/250 [00:25<11:02, 2.75s/it]
segment: 4%|▍ | 10/250 [00:28<10:59, 2.75s/it]
segment: 4%|▍ | 11/250 [00:30<10:55, 2.74s/it]
segment: 5%|▍ | 12/250 [00:33<10:52, 2.74s/it]
segment: 5%|▌ | 13/250 [00:36<10:49, 2.74s/it]
segment: 6%|▌ | 14/250 [00:38<10:47, 2.74s/it]
segment: 6%|▌ | 15/250 [00:41<10:43, 2.74s/it]
segment: 6%|▋ | 16/250 [00:44<10:40, 2.74s/it]
segment: 7%|▋ | 17/250 [00:47<10:39, 2.74s/it]
segment: 7%|▋ | 18/250 [00:49<10:35, 2.74s/it]
segment: 8%|▊ | 19/250 [00:52<10:31, 2.73s/it]
segment: 8%|▊ | 20/250 [00:55<10:29, 2.74s/it]
segment: 8%|▊ | 21/250 [00:58<10:26, 2.74s/it]
segment: 9%|▉ | 22/250 [01:00<10:24, 2.74s/it]
segment: 9%|▉ | 23/250 [01:03<10:21, 2.74s/it]
segment: 10%|▉ | 24/250 [01:06<10:20, 2.75s/it]
segment: 10%|█ | 25/250 [01:09<10:16, 2.74s/it]
segment: 10%|█ | 26/250 [01:11<10:11, 2.73s/it]
segment: 11%|█ | 27/250 [01:14<10:09, 2.73s/it]
segment: 11%|█ | 28/250 [01:17<10:07, 2.74s/it]
segment: 12%|█▏ | 29/250 [01:20<10:05, 2.74s/it]
segment: 12%|█▏ | 30/250 [01:22<10:02, 2.74s/it]
segment: 12%|█▏ | 31/250 [01:25<10:00, 2.74s/it]
segment: 13%|█▎ | 32/250 [01:28<09:56, 2.74s/it]
segment: 13%|█▎ | 33/250 [01:30<09:53, 2.74s/it]
segment: 14%|█▎ | 34/250 [01:33<09:50, 2.73s/it]
segment: 14%|█▍ | 35/250 [01:36<09:47, 2.73s/it]
segment: 14%|█▍ | 36/250 [01:39<10:36, 2.97s/it]
segment: 15%|█▍ | 37/250 [01:43<11:11, 3.15s/it]
segment: 15%|█▌ | 38/250 [01:47<11:35, 3.28s/it]
segment: 16%|█▌ | 39/250 [01:50<11:50, 3.37s/it]
segment: 16%|█▌ | 40/250 [01:54<11:58, 3.42s/it]
segment: 16%|█▋ | 41/250 [01:57<12:03, 3.46s/it]
segment: 17%|█▋ | 42/250 [02:01<12:07, 3.50s/it]
segment: 17%|█▋ | 43/250 [02:04<12:06, 3.51s/it]
segment: 18%|█▊ | 44/250 [02:08<12:05, 3.52s/it]
segment: 18%|█▊ | 45/250 [02:12<12:05, 3.54s/it]
segment: 18%|█▊ | 46/250 [02:15<12:00, 3.53s/it]
segment: 19%|█▉ | 47/250 [02:19<11:58, 3.54s/it]
segment: 19%|█▉ | 48/250 [02:22<11:56, 3.55s/it]
segment: 20%|█▉ | 49/250 [02:26<11:54, 3.55s/it]
segment: 20%|██ | 50/250 [02:29<11:51, 3.56s/it]
segment: 20%|██ | 51/250 [02:33<11:48, 3.56s/it]
segment: 21%|██ | 52/250 [02:36<11:45, 3.56s/it]
segment: 21%|██ | 53/250 [02:40<11:42, 3.57s/it]
segment: 22%|██▏ | 54/250 [02:44<11:38, 3.57s/it]
segment: 22%|██▏ | 55/250 [02:47<11:35, 3.56s/it]
segment: 22%|██▏ | 56/250 [02:51<11:31, 3.56s/it]
segment: 23%|██▎ | 57/250 [02:54<11:27, 3.56s/it]
segment: 23%|██▎ | 58/250 [02:58<11:23, 3.56s/it]
segment: 24%|██▎ | 59/250 [03:01<11:23, 3.58s/it]
segment: 24%|██▍ | 60/250 [03:05<11:20, 3.58s/it]
segment: 24%|██▍ | 61/250 [03:09<11:17, 3.58s/it]
segment: 25%|██▍ | 62/250 [03:12<11:11, 3.57s/it]
\ No newline at end of file +
segment: 0%| | 0/250 [00:00<?, ?it/s]
segment: 0%| | 1/250 [00:03<14:05, 3.39s/it]
segment: 1%| | 2/250 [00:06<12:26, 3.01s/it]
segment: 1%| | 3/250 [00:08<11:50, 2.88s/it]
segment: 2%|▏ | 4/250 [00:11<11:34, 2.82s/it]
segment: 2%|▏ | 5/250 [00:14<11:25, 2.80s/it]
segment: 2%|▏ | 6/250 [00:17<11:16, 2.77s/it]
segment: 3%|▎ | 7/250 [00:19<11:09, 2.76s/it]
segment: 3%|▎ | 8/250 [00:22<11:08, 2.76s/it]
segment: 4%|▎ | 9/250 [00:25<11:02, 2.75s/it]
segment: 4%|▍ | 10/250 [00:28<10:59, 2.75s/it]
segment: 4%|▍ | 11/250 [00:30<10:55, 2.74s/it]
segment: 5%|▍ | 12/250 [00:33<10:52, 2.74s/it]
segment: 5%|▌ | 13/250 [00:36<10:49, 2.74s/it]
segment: 6%|▌ | 14/250 [00:38<10:47, 2.74s/it]
segment: 6%|▌ | 15/250 [00:41<10:43, 2.74s/it]
segment: 6%|▋ | 16/250 [00:44<10:40, 2.74s/it]
segment: 7%|▋ | 17/250 [00:47<10:39, 2.74s/it]
segment: 7%|▋ | 18/250 [00:49<10:35, 2.74s/it]
segment: 8%|▊ | 19/250 [00:52<10:31, 2.73s/it]
segment: 8%|▊ | 20/250 [00:55<10:29, 2.74s/it]
segment: 8%|▊ | 21/250 [00:58<10:26, 2.74s/it]
segment: 9%|▉ | 22/250 [01:00<10:24, 2.74s/it]
segment: 9%|▉ | 23/250 [01:03<10:21, 2.74s/it]
segment: 10%|▉ | 24/250 [01:06<10:20, 2.75s/it]
segment: 10%|█ | 25/250 [01:09<10:16, 2.74s/it]
segment: 10%|█ | 26/250 [01:11<10:11, 2.73s/it]
segment: 11%|█ | 27/250 [01:14<10:09, 2.73s/it]
segment: 11%|█ | 28/250 [01:17<10:07, 2.74s/it]
segment: 12%|█▏ | 29/250 [01:20<10:05, 2.74s/it]
segment: 12%|█▏ | 30/250 [01:22<10:02, 2.74s/it]
segment: 12%|█▏ | 31/250 [01:25<10:00, 2.74s/it]
segment: 13%|█▎ | 32/250 [01:28<09:56, 2.74s/it]
segment: 13%|█▎ | 33/250 [01:30<09:53, 2.74s/it]
segment: 14%|█▎ | 34/250 [01:33<09:50, 2.73s/it]
segment: 14%|█▍ | 35/250 [01:36<09:47, 2.73s/it]
segment: 14%|█▍ | 36/250 [01:39<10:36, 2.97s/it]
segment: 15%|█▍ | 37/250 [01:43<11:11, 3.15s/it]
segment: 15%|█▌ | 38/250 [01:47<11:35, 3.28s/it]
segment: 16%|█▌ | 39/250 [01:50<11:50, 3.37s/it]
segment: 16%|█▌ | 40/250 [01:54<11:58, 3.42s/it]
segment: 16%|█▋ | 41/250 [01:57<12:03, 3.46s/it]
segment: 17%|█▋ | 42/250 [02:01<12:07, 3.50s/it]
segment: 17%|█▋ | 43/250 [02:04<12:06, 3.51s/it]
segment: 18%|█▊ | 44/250 [02:08<12:05, 3.52s/it]
segment: 18%|█▊ | 45/250 [02:12<12:05, 3.54s/it]
segment: 18%|█▊ | 46/250 [02:15<12:00, 3.53s/it]
segment: 19%|█▉ | 47/250 [02:19<11:58, 3.54s/it]
segment: 19%|█▉ | 48/250 [02:22<11:56, 3.55s/it]
segment: 20%|█▉ | 49/250 [02:26<11:54, 3.55s/it]
segment: 20%|██ | 50/250 [02:29<11:51, 3.56s/it]
segment: 20%|██ | 51/250 [02:33<11:48, 3.56s/it]
segment: 21%|██ | 52/250 [02:36<11:45, 3.56s/it]
segment: 21%|██ | 53/250 [02:40<11:42, 3.57s/it]
segment: 22%|██▏ | 54/250 [02:44<11:38, 3.57s/it]
segment: 22%|██▏ | 55/250 [02:47<11:35, 3.56s/it]
segment: 22%|██▏ | 56/250 [02:51<11:31, 3.56s/it]
segment: 23%|██▎ | 57/250 [02:54<11:27, 3.56s/it]
segment: 23%|██▎ | 58/250 [02:58<11:23, 3.56s/it]
segment: 24%|██▎ | 59/250 [03:01<11:23, 3.58s/it]
segment: 24%|██▍ | 60/250 [03:05<11:20, 3.58s/it]
segment: 24%|██▍ | 61/250 [03:09<11:17, 3.58s/it]
segment: 25%|██▍ | 62/250 [03:12<11:11, 3.57s/it]
segment: 25%|██▌ | 63/250 [03:16<11:07, 3.57s/it]
segment: 26%|██▌ | 64/250 [03:19<11:03, 3.57s/it]
segment: 26%|██▌ | 65/250 [03:23<11:00, 3.57s/it]
segment: 26%|██▋ | 66/250 [03:26<10:54, 3.56s/it]
segment: 27%|██▋ | 67/250 [03:30<10:50, 3.56s/it]
segment: 27%|██▋ | 68/250 [03:34<10:46, 3.55s/it]
segment: 28%|██▊ | 69/250 [03:37<10:43, 3.56s/it]
segment: 28%|██▊ | 70/250 [03:41<10:40, 3.56s/it]
segment: 28%|██▊ | 71/250 [03:44<10:38, 3.57s/it]
segment: 29%|██▉ | 72/250 [03:48<10:33, 3.56s/it]
segment: 29%|██▉ | 73/250 [03:51<10:29, 3.56s/it]
segment: 30%|██▉ | 74/250 [03:55<10:27, 3.56s/it]
segment: 30%|███ | 75/250 [03:58<10:22, 3.56s/it]
segment: 30%|███ | 76/250 [04:02<10:18, 3.56s/it]
segment: 31%|███ | 77/250 [04:06<10:14, 3.55s/it]
segment: 31%|███ | 78/250 [04:09<10:11, 3.55s/it]
segment: 32%|███▏ | 79/250 [04:13<10:08, 3.56s/it]
segment: 32%|███▏ | 80/250 [04:16<10:04, 3.56s/it]
segment: 32%|███▏ | 81/250 [04:20<10:01, 3.56s/it]
segment: 33%|███▎ | 82/250 [04:23<09:58, 3.56s/it]
segment: 33%|███▎ | 83/250 [04:27<09:53, 3.56s/it]
segment: 34%|███▎ | 84/250 [04:30<09:50, 3.56s/it]
segment: 34%|███▍ | 85/250 [04:34<09:46, 3.56s/it]
segment: 34%|███▍ | 86/250 [04:38<09:42, 3.55s/it]
segment: 35%|███▍ | 87/250 [04:41<09:39, 3.55s/it]
segment: 35%|███▌ | 88/250 [04:45<09:36, 3.56s/it]
segment: 36%|███▌ | 89/250 [04:48<09:33, 3.56s/it]
segment: 36%|███▌ | 90/250 [04:52<09:31, 3.57s/it]
segment: 36%|███▋ | 91/250 [04:55<09:28, 3.58s/it]
segment: 37%|███▋ | 92/250 [04:59<09:25, 3.58s/it]
segment: 37%|███▋ | 93/250 [05:03<09:22, 3.58s/it]
segment: 38%|███▊ | 94/250 [05:06<09:18, 3.58s/it]
segment: 38%|███▊ | 95/250 [05:10<09:13, 3.57s/it]
segment: 38%|███▊ | 96/250 [05:13<09:10, 3.57s/it]
segment: 39%|███▉ | 97/250 [05:17<09:06, 3.57s/it]
segment: 39%|███▉ | 98/250 [05:20<09:03, 3.57s/it]
segment: 40%|███▉ | 99/250 [05:24<08:59, 3.57s/it]
segment: 40%|████ | 100/250 [05:28<08:55, 3.57s/it]
segment: 40%|████ | 101/250 [05:31<08:52, 3.58s/it]
segment: 41%|████ | 102/250 [05:35<08:48, 3.57s/it]
segment: 41%|████ | 103/250 [05:38<08:45, 3.57s/it]
segment: 42%|████▏ | 104/250 [05:42<08:40, 3.57s/it]
segment: 42%|████▏ | 105/250 [05:45<08:38, 3.57s/it]
segment: 42%|████▏ | 106/250 [05:49<08:34, 3.58s/it]
segment: 43%|████▎ | 107/250 [05:53<08:31, 3.58s/it]
segment: 43%|████▎ | 108/250 [05:56<08:27, 3.57s/it]
segment: 44%|████▎ | 109/250 [06:00<08:23, 3.57s/it]
segment: 44%|████▍ | 110/250 [06:03<08:21, 3.58s/it]
segment: 44%|████▍ | 111/250 [06:07<08:17, 3.58s/it]
segment: 45%|████▍ | 112/250 [06:11<08:13, 3.58s/it]
segment: 45%|████▌ | 113/250 [06:14<08:10, 3.58s/it]
segment: 46%|████▌ | 114/250 [06:17<07:53, 3.48s/it]
segment: 46%|████▌ | 115/250 [06:20<07:19, 3.26s/it]
segment: 46%|████▋ | 116/250 [06:23<06:55, 3.10s/it]
segment: 47%|████▋ | 117/250 [06:26<06:38, 3.00s/it]
segment: 47%|████▋ | 118/250 [06:28<06:25, 2.92s/it]
segment: 48%|████▊ | 119/250 [06:31<06:14, 2.86s/it]
segment: 48%|████▊ | 120/250 [06:34<06:07, 2.83s/it]
segment: 48%|████▊ | 121/250 [06:37<06:01, 2.80s/it]
segment: 49%|████▉ | 122/250 [06:39<05:56, 2.78s/it]
segment: 49%|████▉ | 123/250 [06:42<05:52, 2.77s/it]
segment: 50%|████▉ | 124/250 [06:45<05:48, 2.77s/it]
segment: 50%|█████ | 125/250 [06:47<05:44, 2.75s/it]
segment: 50%|█████ | 126/250 [06:50<05:41, 2.75s/it]
segment: 51%|█████ | 127/250 [06:53<05:38, 2.75s/it]
segment: 51%|█████ | 128/250 [06:56<05:35, 2.75s/it]
segment: 52%|█████▏ | 129/250 [06:58<05:33, 2.76s/it]
segment: 52%|█████▏ | 130/250 [07:01<05:30, 2.76s/it]
segment: 52%|█████▏ | 131/250 [07:04<05:27, 2.75s/it]
segment: 53%|█████▎ | 132/250 [07:07<05:24, 2.75s/it]
segment: 53%|█████▎ | 133/250 [07:09<05:20, 2.74s/it]
segment: 54%|█████▎ | 134/250 [07:12<05:18, 2.75s/it]
segment: 54%|█████▍ | 135/250 [07:15<05:15, 2.75s/it]
segment: 54%|█████▍ | 136/250 [07:18<05:13, 2.75s/it]
segment: 55%|█████▍ | 137/250 [07:20<05:10, 2.75s/it]
segment: 55%|█████▌ | 138/250 [07:23<05:07, 2.74s/it]
segment: 56%|█████▌ | 139/250 [07:26<05:05, 2.75s/it]
segment: 56%|█████▌ | 140/250 [07:29<05:01, 2.74s/it]
segment: 56%|█████▋ | 141/250 [07:31<04:59, 2.75s/it]
segment: 57%|█████▋ | 142/250 [07:34<04:56, 2.74s/it]
segment: 57%|█████▋ | 143/250 [07:37<04:53, 2.74s/it]
segment: 58%|█████▊ | 144/250 [07:40<05:00, 2.84s/it]
segment: 58%|█████▊ | 145/250 [07:44<05:20, 3.06s/it]
segment: 58%|█████▊ | 146/250 [07:47<05:33, 3.21s/it]
segment: 59%|█████▉ | 147/250 [07:51<05:41, 3.32s/it]
segment: 59%|█████▉ | 148/250 [07:54<05:48, 3.42s/it]
segment: 60%|█████▉ | 149/250 [07:58<05:49, 3.46s/it]
segment: 60%|██████ | 150/250 [08:01<05:49, 3.49s/it]
segment: 60%|██████ | 151/250 [08:05<05:48, 3.52s/it]
segment: 61%|██████ | 152/250 [08:09<05:46, 3.53s/it]
segment: 61%|██████ | 153/250 [08:12<05:43, 3.54s/it]
segment: 62%|██████▏ | 154/250 [08:16<05:41, 3.56s/it]
segment: 62%|██████▏ | 155/250 [08:19<05:37, 3.56s/it]
segment: 62%|██████▏ | 156/250 [08:23<05:34, 3.56s/it]
segment: 63%|██████▎ | 157/250 [08:26<05:31, 3.56s/it]
segment: 63%|██████▎ | 158/250 [08:30<05:28, 3.57s/it]
segment: 64%|██████▎ | 159/250 [08:34<05:24, 3.57s/it]
segment: 64%|██████▍ | 160/250 [08:37<05:21, 3.57s/it]
segment: 64%|██████▍ | 161/250 [08:41<05:17, 3.57s/it]
\ No newline at end of file diff --git a/logs/omit_rest.log b/logs/omit_rest.log new file mode 100644 index 0000000..e69de29 --- /dev/null +++ b/logs/omit_rest.log |