problem with import tensorflow

[Ehplodor]

problem with import tensorflow

[laksjdjf]

1. PFG scale is similar to prompt emphasis by ()[]. PFG scale affects only the image, and cfg scale affects both the image and the prompt.
2. Negative PFG can be implemented, but it is not clear if it will give better results than negative embedding. I will try it as it sounds interesting.
3. You can convert a style by providing prompts or LoRAs that represent the style.

[laksjdjf]

Can you import tensorflow for normal python code?

For example in cmd.

$python3

>>>import tensorflow

[Ehplodor]

OK sorry I finally managed to copy paste a number of libs from an old copy of SD-WEBUI, that I could'nt manage to get done since this morning. I really must have mismanaged my libraries one way or another and shall reinstall everything from scratch... Mhh... Nevertheless, I now get your extension in txt2img. let's try it

[Ehplodor]

Oups

Not enough GPU hardware devices available
2023-02-28 16:02:28.512160: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations: AVX AVX2
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
WARNING:tensorflow:No training configuration found in save file, so the model was not compiled. Compile it manually.
Data shape for DDIM sampling is (1, 4, 64, 64), eta 0.0
Running DDIM Sampling with 31 timesteps
DDIM Sampler: 100%|████████████████████████████████████████████████████████████████████| 31/31 [00:10<00:00, 2.90it/s]

Do you know what does that mean ? Me not really...

[Ehplodor]

I don't seem to get anything related to the input image. My implementation might not be fully functional (what a surprise)

[laksjdjf]

Did you merge it ?

In any case, it is not good that there is no message to confirm that it has been applied. I will fix that.

[Ehplodor]

Yes I merged manually your PR in webui

[Ehplodor]

The problem is probably with my installation I'll try to reinstall everything from scratch later. Don't mind me :-)

[Ehplodor]

fresh install. I had to add
tensorflow==2.12.0rc0
protobuf==3.20.3
to requirements_versions.txt in order for everything to be compatible.
finally...
Not enough GPU hardware devices available
Error running process: C:\AI\AAA\SD2\stable-diffusion-webui\extensions\pfg-webui\scripts\pfg-webui.py
Traceback (most recent call last):
File "C:\AI\AAA\SD2\stable-diffusion-webui\modules\scripts.py", line 386, in process
script.process(p, *script_args)
File "C:\AI\AAA\SD2\stable-diffusion-webui\extensions\pfg-webui\scripts\pfg-webui.py", line 118, in process
self.tagger = load_model(os.path.join(CURRENT_DIRECTORY, TAGGER_DIR))
File "C:\AI\AAA\SD2\stable-diffusion-webui\venv\lib\site-packages\keras\saving\saving_api.py", line 212, in load_model
return legacy_sm_saving_lib.load_model(
File "C:\AI\AAA\SD2\stable-diffusion-webui\venv\lib\site-packages\keras\utils\traceback_utils.py", line 70, in error_handler
raise e.with_traceback(filtered_tb) from None
File "C:\AI\AAA\SD2\stable-diffusion-webui\venv\lib\site-packages\tensorflow\python\trackable\data_structures.py", line 823, in getattribute
return super().getattribute(name)
TypeError: this dict descriptor does not support '_DictWrapper' objects

[laksjdjf]

I recommend to use tensorflow<=2.11.
There seem to be some dependency conflicts, but it works.

[Ehplodor]

Same error after changing tensorflow to 2.11.0, same with 2.9.1. It must be some other module version that is not good.

[laksjdjf]

The problem seems to be confirmed in 2.12rc, but if it still happens after downgrading to 2.11 or lower, it may be due to other dependent libraries.

tensorflow/tensorflow#59869

[laksjdjf]

The version of the wrapt library seems to be relevant.

[Ehplodor]

FYI, I tried downgrading wrapt from 2.15.0 to 2.11.0 as what seems to be suggested from your link but the exact same error is still there. I hope other people trying out your extension will have more luck with their installed libraries and will report success.

[Ehplodor]

After pip install tf-nightly, it seems to go a little further :
Error running process: C:\AI\AAA\SD2\stable-diffusion-webui\extensions\pfg-webui\scripts\pfg-webui.py
Traceback (most recent call last):
File "C:\AI\AAA\SD2\stable-diffusion-webui\modules\scripts.py", line 386, in process
script.process(p, *script_args)
File "C:\AI\AAA\SD2\stable-diffusion-webui\extensions\pfg-webui\scripts\pfg-webui.py", line 104, in process
pfg_weight = torch.load(os.path.join(CURRENT_DIRECTORY, "models/" + pfg_path))
TypeError: can only concatenate str (not "list") to str

EDIT: Actually i forgot to specify the model in the listbox. When done, I get the previous error again...

[laksjdjf]

I added the option to use onnxruntime, it may work without tensorflow.
https://onnxruntime.ai/

[Ehplodor]

Thank you. after updating your extension and restarting completely the webui so as to let onxx model download itself during start, I can confirm this works fine for WD1.4. It works with WD1.5 too.

I notice pfg's scale doesn't seems to have any effect.
Moreover, about number of tokens : only '10' is working. I encounter weird errors using any other number.

[laksjdjf]

Thanks for the report !
I forgot the code to apply pfg_scale and I already added it :)

The number of tokens is determined by one value per pfg file". The last 10 in "pfg-wd14-n10" is the number of tokens.

[Ehplodor]

It works. GG

I tried your WD15 version too. Can't say for the moment if better or not than WD14. Probably better but not enough testing.

The image's embedding seems very strong in comparison to text's embedding. As you said in the readme/Hints, I had to tune up up up the text cfg scale, using cfg treshold extension, and lower image's pfg to 0.65-70 in order to get somehow good results. Also, pfg = 3 destroyed everything :-)

Some questions pop up from those early tests :

• First question : In your understanding, is PFG more comparable to the "strength" parameter of the img2img tab ? or is it more similar to CFG scale ?
• Another question : would it be conceivable to use an image with negative PFG so as to substract it from the latents, similar to negative embedding but as an image ?
• Last question : can your extension be used for style transfer ?