Closed lweingart closed 2 weeks ago
I found the same issue. Do you have any solution yet?
Hi, no I still don't. I'm stuck for now
@g-jing , have you been able to find a solution on your side ? As for me I still don't.
Hi again, for some reason the CSV created at data processing step 3.2:
# 3.2 Filter by aesthetic scores. This should output ${ROOT_META}/meta_clips_info_fmin1_aes_aesmin5.csv
python -m tools.datasets.datautil ${ROOT_META}/meta_clips_info_fmin1_aes.csv --aesmin 5
has this first line after being generated:
path,id,relpath,num_frames,height,width,aspect_ratio,fps,resolution,aes
but for some reason those headers were lost at step 4.1 and the next csv file, named meta_clips_info_fmin1_aes_aesmin5_caption_part*.csv from step:
# 4.1 Generate caption. This should output ${ROOT_META}/meta_clips_info_fmin1_aes_aesmin5_caption_part*.csv
torchrun --nproc_per_node 8 --standalone -m tools.caption.caption_llava \
${ROOT_META}/meta_clips_info_fmin1_aes_aesmin5.csv \
--dp-size 8 \
--tp-size 1 \
--model-path /path/to/llava-v1.6-mistral-7b \
--prompt video
only has:
path,text,num_frames
I don't know why most columns were lost, but reintegrating the width and height columns should do the trick. I'm rerunning the data processing to check if i missed some errors in the logs, and I'll get back here
So, the code tools.caption.caption_llava
at line 209 has this:
dp_writer.writerow(["path", "text", "num_frames"])
Which, as can be read, only creates the three 'path', 'text' and 'num_frames'
columns.
However, manually adding the height
and width
columns to my csv file fixed the KeyError problem.
Unfortunately, I don't have any automated way to reintroduce these columns based on the previous csv file.
Lucky for me, in this case I specifically resized all my videos to 256x144 so it was easily done, but in a more classic situation where videos can be of multiple resolutions, I don't have any solution to propose.
This issue is stale because it has been open for 7 days with no activity.
This issue was closed because it has been inactive for 7 days since being marked as stale.
I am facing the same issue, is there any official fix yet?
Hello guys,
I just followed your precess to prepare my own dataset as described here and I must admit it went impressively well, no error whatsoever.
Then I went on to check the training part and ran this command:
!torchrun --standalone --nproc_per_node 1 -m scripts.train \ configs/opensora-v1-2/train/stage1.py --data-path {ROOT_META}/meta_clips_caption.csv --ckpt-path {MODEL_OUTPUT}/my_sora.pt
but I end up with a KeyError 'Height'.Could you please help me identify a way to fix this ? Any help would be greatly appreciated.
Thank you very much in advance Cheers
Here is the full log trace: