NoneType' object is not subscriptable' - Training Failed #799
Replies: 4 comments
-
Worth mentioning that Tab "Generate" trigger the same error when clicking: generate sample images. "Exception generating sample(s): 'NoneType' object has no attribute 'device'" |
Beta Was this translation helpful? Give feedback.
-
Setting the value Mixed Precision to "no" skip everything and end with "Training finished. Total lifetime steps: 0". When set to fp16, the error is ''NoneType' object is not subscriptable'. |
Beta Was this translation helpful? Give feedback.
-
Did you manage to fix this issue? |
Beta Was this translation helpful? Give feedback.
-
I'm on Windows and had to revert back to commit 534baa0, which was prior to some "MacOS fixes" (6f87234). It seems like the code change afterward was less accepting of ASCII encoding as suggested here https://github.com/d8ahazard/sd_dreambooth_extension/issues/893 Someone please correct or elaborate on something I may have gotten wrong. |
Beta Was this translation helpful? Give feedback.
-
Greetings!
First incredbile extension. If there is any way I can support your financially for all the work, please let me know. I am running the following JSON format for my concept:
{
"max_steps": -1,
"instance_data_dir": "PATH_TO_IMAGE",
"class_data_dir": "PATH_TO_IMAGE",
"num_class_images": 200,
"instance_prompt": "[filewords]",
"class_prompt": "creatures",
"class_negative_prompt": "woman, human, men, person",
"save_sample_prompt": "a digital painting of a mighty orc warrior creature"
}
This exact JSON (with about 10 concepts) worked perfectly for a while but sadly stopped working when I did the Git Pull today. The rest of my config is as follow (Running on A10 AWS instance)
Steps per image: 150
Pause: 0
Time to pause: 0
Save model Frequency: 15
Save previews: 5
Batch Size: 4
Gradient Acc Steps: 4
Class Batch: 4
Set Gradients: True
Gradient Checkpoint: True
Learning Rate: Constat with Warmup
Learning Rate: 0.0000014
Learning Rate Warmup: 500
Resolution: 512
(No other changes for these until the row Use Copncepts)
Use Concepts: True
Concepts List (The JSON I just posted above)
Use EMA: False
Use 8bit: True
Mixed Precision: fp16
Memory: xformers
(All other setting left as is)
For the Saving Tabs:
Half Model: True
Save in subdirectory: True
Generate .CKPT when saving training: True
Generate .CKPT when training over: True
Beta Was this translation helpful? Give feedback.
All reactions