🍔😋🍟
Follow me on X/Twitter for exclusive content: https://x.com/Citron_Legacy 😁
Where's the Beef? Meme
Description
Trained on: 199 images
Training Model: Animefull-final-pruned-fp16
flip_aug: False
Num of Repeats: 2
Unit is Epochs or Steps: Epochs
Number of Epochs or Steps: 10
Training Batch Size: 4
Total Steps: 995
Resolution: 512
Network Dim: Dim 32 and Alpha 16
Lora Creation Process took: 00:19:01
FAQ
Comments (13)
Hi, sorry but I need help. I can no longer create Lora models using google colab, because it gives me an error during the process. Can you tell me how you create them? Is there an alternative to Google Colab?
Hi
Which trainer are you using?
Holo's XL trainer is working as of yesterday https://github.com/hollowstrawberry/kohya-colab
I just started training something and it appears to be working.
If XL is working I assume his SD1.5 trainer also works.
@CitronLegacy However, this doesn't work for me either, here's the same error it gives me here too: 04/06 21:23:04 [ERROR] CUID#7 - Download aborted. URI=https://huggingface.co/Linaqruf/stolen/resolve/main/pruned-models/stable_diffusion_1_5-pruned.safetensors
Exception: [AbstractCommand.cc:351] errorCode=24 URI=https://huggingface.co/Linaqruf/stolen/resolve/main/pruned-models/stable_diffusion_1_5-pruned.safetensors
-> [HttpSkipResponseCommand.cc:215] errorCode=24 Authorization failed.
Download Results:
gid |stat|avg speed |path/URI
======+====+===========+============== ==============================
245576|ERR | 0B/s|/content/pretrained_model/Stable-Diffusion-v1-5.safetensors
Status Legend:
(ERR):error occurred.
aria2 will resume download if the transfer is restarted.
If there are any errors, then see the log file. See '-l' option in help/man page for details.
@ApexThunder Dang RIP.
I see why that happened. That error is saying the SD1.5 checkpoint you tried to train with is no longer uploaded.
Linaqruf said he was shutting down his trainer so its no surprise that he'd delete the checkpoint he was hosting.
You won't have that problem with Holo's trainer because Holo is hosting the SD1.5 checkpoint on a different url.
Please confirm if the trainer I linked: https://github.com/hollowstrawberry/kohya-colab is working for you or not.
I just trained a both an XL and a normal SD1.5 since my last message 3 hours ago so I confirmed Holo's code does not produce that error.
If possible please link me the Google Colab you are using?
@CitronLegacy this is what I used https://colab.research.google.com/github/Linaqruf/kohya-trainer/blob/main/kohya-LoRA-dreambooth.ipynb
@CitronLegacy In your link I see 3 options:
1) 📊 Dataset Maker
2) ⭐ Lora Trainer
3) 🌟 XL Lora Trainer
I tried to perform number 2, but it doesn't work for me because it is different from Kohya LoRA Dreambooth. In my opinion the procedure is different from what I did. I'm going crazy believe me :(
@ApexThunder Lol sorry I do believe you. I just need to clarify what you were talking about. "2) ⭐ Lora Trainer" is the right option but you are right it is different from Linaqruf's trainer.
Its 80% the same but there are a few differences that will make it work differently.
You will need to learn how to use Hollowstrawberry's trainer or learn how to fix the issue in Linaqruf's code. (its not a hard issue to to fix but it would be better to learn Hollowstrawberry's process)
Hollowstrawberry has a guide to explain his trainer: https://civitai.com/models/22530?modelVersionId=26903
Like I said the trainer should operate about 80% the same way but there are some differences. See this screenshot for example
https://civitai.com/posts/1840080
See how the unet_lr and text_encoder_lr have different default values in Linaqruf vs Hollowstrawberry.
@CitronLegacy not work for me: CalledProcessError: Command '['/usr/bin/python3', 'train_network.py', '--dataset_config=/content/drive/MyDrive/Loras/charlie_demon/dataset_config.toml', '--config_file=/content/drive/MyDrive/Loras/charlie_demon/training_config.toml']' returned non-zero exit status 1.
@ApexThunder Thats the entire error message? It seems like its missing details.
Also can you verify that the file exists in your google drive "/content/drive/MyDrive/Loras/charlie_demon/dataset_config.toml" ?
@CitronLegacy hi, after I check it better, however sometimes it tells me that I have finished using the GPU for free. Is there a version that can be downloaded onto the PC? because I have an RTX 4070 OC, I think I can do it :)
@CitronLegacy [Dataset 0]
loading image sizes.
100% 22/22 [00:08<00:00, 2.57it/s]
make buckets
number of images (including repeats) / 各bucketの画像枚数（繰り返し回数を含む）
bucket 0: resolution (768, 768), count: 220
mean ar error (without repeats): 0.0
preparing accelerator
╭─────────────────────────────── Traceback (most recent call last) ────────────────────────────────╮
│ /content/kohya-trainer/train_network.py:873 in <module> │
│ │
│ 870 │ args = parser.parse_args() │
│ 871 │ args = train_util.read_config_from_file(args, parser) │
│ 872 │ │
│ ❱ 873 │ train(args) │
│ 874 │
│ │
│ /content/kohya-trainer/train_network.py:161 in train │
│ │
│ 158 │ │
│ 159 │ # acceleratorを準備する │
│ 160 │ print("preparing accelerator") │
│ ❱ 161 │ accelerator, unwrap_model = train_util.prepare_accelerator(args) │
│ 162 │ is_main_process = accelerator.is_main_process │
│ 163 │ │
│ 164 │ # mixed precisionに対応した型を用意しておき適宜castする │
│ │
│ /content/kohya-trainer/library/train_util.py:3069 in prepare_accelerator │
│ │
│ 3066 │ │ │ if args.wandb_api_key is not None: │
│ 3067 │ │ │ │ wandb.login(key=args.wandb_api_key) │
│ 3068 │ │
│ ❱ 3069 │ accelerator = Accelerator( │
│ 3070 │ │ gradient_accumulation_steps=args.gradient_accumulation_steps, │
│ 3071 │ │ mixed_precision=args.mixed_precision, │
│ 3072 │ │ log_with=log_with, │
│ │
│ /usr/local/lib/python3.10/dist-packages/accelerate/accelerator.py:355 in init │
│ │
│ 352 │ │ if self.state.mixed_precision == "fp16" and self.distributed_type != Distributed │
│ 353 │ │ │ self.native_amp = True │
│ 354 │ │ │ if not torch.cuda.is_available() and not parse_flag_from_env("ACCELERATE_USE │
│ ❱ 355 │ │ │ │ raise ValueError(err.format(mode="fp16", requirement="a GPU")) │
│ 356 │ │ │ kwargs = self.scaler_handler.to_kwargs() if self.scaler_handler is not None │
│ 357 │ │ │ if self.distributed_type == DistributedType.FSDP: │
│ 358 │ │ │ │ from torch.distributed.fsdp.sharded_grad_scaler import ShardedGradScaler │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
ValueError: fp16 mixed precision requires a GPU
╭─────────────────────────────── Traceback (most recent call last) ────────────────────────────────╮
│ /usr/local/bin/accelerate:8 in <module> │
│ │
│ 5 from accelerate.commands.accelerate_cli import main │
│ 6 if name == '__main__': │
│ 7 │ sys.argv[0] = re.sub(r'(-script\.pyw|\.exe)?$', '', sys.argv[0]) │
│ ❱ 8 │ sys.exit(main()) │
│ 9 │
│ │
│ /usr/local/lib/python3.10/dist-packages/accelerate/commands/accelerate_cli.py:45 in main │
│ │
│ 42 │ │ exit(1) │
│ 43 │ │
│ 44 │ # Run │
│ ❱ 45 │ args.func(args) │
│ 46 │
│ 47 │
│ 48 if name == "__main__": │
│ │
│ /usr/local/lib/python3.10/dist-packages/accelerate/commands/launch.py:1104 in launch_command │
│ │
│ 1101 │ elif defaults is not None and defaults.compute_environment == ComputeEnvironment.AMA │
│ 1102 │ │ sagemaker_launcher(defaults, args) │
│ 1103 │ else: │
│ ❱ 1104 │ │ simple_launcher(args) │
│ 1105 │
│ 1106 │
│ 1107 def main(): │
│ │
│ /usr/local/lib/python3.10/dist-packages/accelerate/commands/launch.py:567 in simple_launcher │
│ │
│ 564 │ process = subprocess.Popen(cmd, env=current_env) │
│ 565 │ process.wait() │
│ 566 │ if process.returncode != 0: │
│ ❱ 567 │ │ raise subprocess.CalledProcessError(returncode=process.returncode, cmd=cmd) │
│ 568 │
│ 569 │
│ 570 def multi_gpu_launcher(args): │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
CalledProcessError: Command '['/usr/bin/python3', 'train_network.py',
'--dataset_config=/content/drive/MyDrive/Loras/charliedemon/dataset_config.toml',
'--config_file=/content/drive/MyDrive/Loras/charliedemon/training_config.toml']' returned non-zero
exit status 1.
@ApexThunder Yeah this error implies you didn't connect to the GPU. "ValueError: fp16 mixed precision requires a GPU"
Its you just need to select the GPU run time in the top right. And if you ran out of free GPU you just need to wait for 24 hours to try again.
---
If you want to train locally you can use https://github.com/bmaltais/kohya_ss
RTX 4070 sounds good enough. 8GB VRam is enough for SD1.5 training with a Batch size of 2, so if you have more than that (I guess 4070 has 12 or 16 GB VRam?) you'll certainly be able to train offline.
@CitronLegacy because the universe is making things complicated for me XD. I downloaded the version for PC, it opens from browser, as I chose, and from here I don't know what to do, it's a mess. The version I have is v23, all the tutorials I found have an old version, which doesn't match mine. Afterwards I'll try again with the Online one, but I really don't know how to use this one from the PC :(