lllyasviel / lllyasviel/ControlNet
custom dataset and gradient exploding
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 34.1k
- Forks
- 3k
- PR merge metrics
- No merged PRs in 30d
Description
Hi, thanks for the great works.
I followed the instructions to make my custom dataset, but the gradient exploded and the weights became NaN after 150-200 steps training.
When I used the tutorial dataset fill50k everything is working fine.
So I'm curious about why my custom dataset leads to gradient exploding. Is this caused by using one single prompt for all inputs? Or do I miss anything? P.S. I'm using sd1.5 and the dataset class is the same as tutorial_dataset.py.
Looking forward to your reply. Thank you.
The following is a sample of my custom dataset.
source
target
prompt
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by comparing the custom dataset and prompt setup with tutorial_dataset.py and the fill50k tutorial dataset, then reproduce the SD1.5 training run to locate when gradients and weights become NaN. Done means identifying the dataset or training difference that causes the failure and documenting a reproducible correction.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 32/100