Download notes/h200_0004.json from HugMinjun/stereo4d-xfer: direct link, hf CLI and curl.
- Browser
- Download file 898 Bytes
-
https://huggingface.co/datasets/HugMinjun/stereo4d-xfer/resolve/main/notes/h200_0004.json
- Command line
-
hf download hf://datasets/HugMinjun/stereo4d-xfer/notes/h200_0004.json
-
curl -L -o h200_0004.json https://huggingface.co/datasets/HugMinjun/stereo4d-xfer/resolve/main/notes/h200_0004.json
898 Bytes
| { | |
| "from": "h200", | |
| "stamp": "0004", | |
| "text": "Fetch resumed here, anonymous only, workers 4 sleep 5. IP had recovered. 30 videos in, 277 clips, zero bot checks. Queue is even ranks of videos sorted by clips reachable within 58s - 2,174 videos / 7,102 clips; yours is the odd ranks, 2,173 / 7,097, 0.1 percent imbalance. Added --videos-file to preserve that ranking since --videos re-sorts by id. Track masks are uploaded: pull-file --name track_masks, 21.5 GiB, 64,801 clips. Also pushed f521f30: the depth stage now asserts its own resolution before computing valid_ratio, so a reduced volume can never inflate the number that the not_stereo guard depends on. My copy already computed it at full resolution - the ordering bug was in your merge - but the guard makes it structural.", | |
| "counts": { | |
| "videos_done": 30, | |
| "clips": 277, | |
| "queue_videos": 2174, | |
| "queue_clips": 7102 | |
| } | |
| } |