Feedback needed for Alpha Release

[Linaqruf]

I feel like the quality of my notebook has been getting worse over the past three months, since I declared a hiatus from the Stable Diffusion Community in February. So, I want to make sure that I do the right thing before I release this to the main branch.

The main points about this update are:

• Simplified cells and parameters, so I need your feedback on which cells and parameters I need to simplify further.
• Provided more explanation about certain parameters. I need to know if this is the right thing to do, as it goes against my plan to simplify the notebook.
• Fixed the bug for class_token is undefined.
• Using Wandb instead of Tensorboard as the default logging tool. Please prepare your API key.
• Added more LoRA modules: ["LoRA_LierLa", "LoRA_C3Lier", "DyLoRA_LierLa", "DyLoRA_C3Lier", "LoCon", "LoHa", "IA3", "LoKR"].
  • I have only used LoRA_LierLa and LoRA_C3Lier so far, so I need your help with any recommended parameters.
• Added the Sample Prompt Config back.
• Planned to delete clone/push to Github, as Huggingface Hub is already powerful.
• etc

You can find the alpha release version at these links:

• https://colab.research.google.com/github/Linaqruf/kohya-trainer/blob/dev/kohya-LoRA-dreambooth.ipynb
• https://colab.research.google.com/github/Linaqruf/kohya-trainer/blob/dev/kohya-LoRA-dreambooth.ipynb
• https://colab.research.google.com/github/Linaqruf/kohya-trainer/blob/dev/kohya-LoRA-dreambooth.ipynb

You can also gives suggestion in this Issue. Thank you!

[Bammargiela]

I can't think of anything right now but will hold it in mind as I'm using it. Mainly came to so thanks for creating this notebook. It's the best out there.

[777Starr]

I wanted to try it out, but it had the same issue as the main one a bit earlier,

FileNotFoundError: [Errno 2] No such file or directory: '/usr/local/lib/python3.9/dist-packages/bitsandbytes/cuda_setup/main.py'

• But I glanced over the things, and personally I don't need the 5.5 Sampler, it usually generates crap anyways, and you could edit it regardless. While at that, I have never used anything from the huggingface bracket 7.
• Also, I would really one more like r34 or even some would like e926 or e621. IIRC, you also need to login into boorus to use more than 2 tags there, so You could also display that to avoid confusion.
• Could have a bit of information links on the 5.4 Optimizer section.

I really wanted to try out LoKr, so that's nice

[ride5k]

same result here:

FileNotFoundError: [Errno 2] No such file or directory: '/usr/local/lib/python3.9/dist-packages/bitsandbytes/cuda_setup/main.py'

[roseyyy2022ai]

I wanted to try it out, but it had the same issue as the main one a bit earlier,

FileNotFoundError: [Errno 2] No such file or directory: '/usr/local/lib/python3.9/dist-packages/bitsandbytes/cuda_setup/main.py'

• But I glanced over the things, and personally I don't need the 5.5 Sampler, it usually generates crap anyways, and you could edit it regardless. While at that, I have never used anything from the huggingface bracket 7.
• Also, I would really one more like r34 or even some would like e926 or e621. IIRC, you also need to login into boorus to use more than 2 tags there, so You could also display that to avoid confusion.
• Could have a bit of information links on the 5.4 Optimizer section.

I really wanted to try out LoKr, so that's nice

I'm a noob, so how do I fix this bc I'm getting the same here....

[Particle1904]

Hey! I love the notebook and I have successfully released 20 LoHas using it. But I have a problem, at some point, the resume training option for LoRAs in the fine-tuner notebook broke and I have a big dataset of 960 (multi-concept theme) images that can't be trained in a single run. I need to train it to 20 epochs but after around 4 hours the colab limit is reached and it kicks me at epoch8. When I try to run with resume training it "loads" the LoRA but prints to the console output that no weight was loaded and starts training from scratch.

[anothertal3]

edit: The current alpha seems to have received a couple of pretty big changes and some of my comments are not applicable anymore. I still need to figure out how to configure my previous setup again.

I've only just started a training session and will need to evaluate the results later. Some things I noticed already, though:

• Please don't remove the ZIP-file after extraction. I will often test different settings and reuse the same source ZIP file for that purpose.
• The setting -1 for min_snr_gamma and noise_offset seems to deactivate the corresponding functions. This isn't very transparent. For example, the description for min_snr_gamma tells us that lower numbers have a stronger effect which might confuse users as -1 is pretty low.
• According to the description, noise_offset = 0.1 is recommended but it isn't set as default?
• The following description surprised me: NovelAI and all modern Stable Diffusion model trains at clip_skip = 2. I assume this refers to SD2.0 / SD2.1 as I was under the impression that SD1.5 was using clip_skip = 1? If that's correct then the description should maybe include that recommendation?

On the positive side:

• I was very pleasantly surprised when I found, that my subdirectories were already detected and offered as additional datasets in dataset_config.toml 👍
• The new image browser seems useful
• Being able to generate samples by using existing captions without manual work seems nice

(Some of these issues are probably not new to the Alpha)

I will add items as I find them.

[boginski345]

Hello! The collab issues a warning about the execution of forbidden code and then disconnects. I chose the option: Linaqruf/kohya-trainer (forked repo, stable, optimized for colab use). Apparently Google understands that the Automatic111 code is being executed.