Software Engineering Courses (SECourses)•3y ago

I'm at 4000 iterations and still have `loss_text_ce` values going down by 0.1 per every 800 iteratio

I'm at 4000 iterations and still have

loss_text_ce

loss_text_ce

values going down by 0.1 per every 800 iterations. I have

loss_text_ce: 1.4474e+00

loss_text_ce: 1.4474e+00

loss_mel_ce: 3.7375e-02

loss_mel_ce: 3.7375e-02

, and

loss_gpt_total: 5.1849e-02

loss_gpt_total: 5.1849e-02

.

These are orders of magnitude lower than your demo but I'm guessing since the values are still going down so much it could still benefit from more training? I'm fine leaving it training all day or even 2 days if loss values keep going down. For background, I am training on a large and high quality sound data set with 5 hours of audio.

CClocksmith I'm at 4000 iterations and still have `loss_text_ce` values going down by 0.1 pe...

Furkan Gözükara SECourses•12/4/23, 12:24 AM

well not all values are meaningful

Furkan Gözükara SECourses•12/4/23, 12:24 AM

hopefully i will show with cocui training

Dazzastrous [4090]•12/4/23, 12:40 AM

when will you be uploading the froggy training Prof?

DDazzastrous [4090]when will you be uploading the froggy training Prof?

Furkan Gözükara SECourses•12/4/23, 12:42 AM

hopefully soon. i just finished pixart

Dazzastrous [4090]•12/4/23, 12:42 AM

yeh I installed that

Dazzastrous [4090]•12/4/23, 12:43 AM

You got your haircut inbetween takes ? hehe

DDazzastrous [4090]You got your haircut inbetween takes ? hehe

Furkan Gözükara SECourses•12/4/23, 12:46 AM

yes

Furkan Gözükara SECourses•12/4/23, 12:46 AM

it took a week to make it

Dazzastrous [4090]•12/4/23, 12:46 AM

Dazzastrous [4090]•12/4/23, 12:53 AM

https://twitter.com/StelfieTT/status/1731378027970338946?s=20 what are your thoughts on this from a still image using pose........looks too good to be true

DDazzastrous [4090]https://twitter.com/StelfieTT/status/1731378027970338946?s=20 what are your t...

Furkan Gözükara SECourses•12/4/23, 1:01 AM

ye i find it fake

Furkan Gözükara SECourses•12/4/23, 1:01 AM

we will see

Furkan Gözükara SECourses•12/4/23, 1:01 AM

all woman used. it is like a fake ads haha

Furkan Gözükara SECourses•12/4/23, 1:01 AM

you are a research team but your demo is all attractive woman. major redflag

Dazzastrous [4090]•12/4/23, 1:02 AM

ClocksmithOP•12/4/23, 3:16 PM

I cut training at 5000 iterations. I'm getting this error when running your test command. Doesn't seem to be obviously related to my Python 3.11 installation, but I don't know for sure. Any idea what is going on?

(venv) C:\Users\username\Desktop\tortoise\tortoise-tts-fast\scripts>python tortoise_tts.py --preset fast --ar_checkpoint "C:\Users\username\Desktop\tortoise\DL-Art-School\experiments\voice-dlas\models\5000_gpt.pth" "Welcome to the software engineering courses channel."
Traceback (most recent call last):
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\scripts\tortoise_tts.py", line 240, in <module>
    from tortoise.inference import (
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\tortoise\inference.py", line 167, in <module>
    vfixer = VoiceFixer()
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\base.py", line 13, in __init__
    self._model = voicefixer_fe(channels=2, sample_rate=44100)
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\restorer\model.py", line 180, in __init__
    self.vocoder = Vocoder(sample_rate=44100)
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\vocoder\base.py", line 19, in __init__
    self._load_pretrain(Config.ckpt)
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\vocoder\base.py", line 25, in _load_pretrain
    self.model = Generator(Config.cin_channels)
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\vocoder\model\generator.py", line 34, in __init__
    nn.utils.parametrizations.weight_norm(
AttributeError: module 'torch.nn.utils.parametrizations' has no attribute 'weight_norm'

(venv) C:\Users\username\Desktop\tortoise\tortoise-tts-fast\scripts>python tortoise_tts.py --preset fast --ar_checkpoint "C:\Users\username\Desktop\tortoise\DL-Art-School\experiments\voice-dlas\models\5000_gpt.pth" "Welcome to the software engineering courses channel."
Traceback (most recent call last):
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\scripts\tortoise_tts.py", line 240, in <module>
    from tortoise.inference import (
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\tortoise\inference.py", line 167, in <module>
    vfixer = VoiceFixer()
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\base.py", line 13, in __init__
    self._model = voicefixer_fe(channels=2, sample_rate=44100)
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\restorer\model.py", line 180, in __init__
    self.vocoder = Vocoder(sample_rate=44100)
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\vocoder\base.py", line 19, in __init__
    self._load_pretrain(Config.ckpt)
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\vocoder\base.py", line 25, in _load_pretrain
    self.model = Generator(Config.cin_channels)
  File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\vocoder\model\generator.py", line 34, in __init__
    nn.utils.parametrizations.weight_norm(
AttributeError: module 'torch.nn.utils.parametrizations' has no attribute 'weight_norm'

CClocksmith I cut training at 5000 iterations. I'm getting this error when running your test...

Furkan Gözükara SECourses•12/4/23, 11:27 PM

looks like torch version related error

Digital [Starburst]•12/5/23, 12:52 AM

I really like your Pixart Gradio ui you built. Do you think it would be possible to include more possible styles and models?

Digital [Starburst]•12/5/23, 12:56 AM

The more intricate the txt prompt, the longer it takes to generate the images.

Digital [Starburst]•12/5/23, 12:58 AM

It would be great if you could add txt2vid for up to 240 frame vids. I'm probably dreaming though. I've tried txt2vid with ComfyUI and it's been a dismal failure. I need some good training on creating good prompts. My prompts aren't generating any desired results. I know you must've created a tutorial somewhere, so I will use Ctrl-F to find it on Git.

DDigital [Starburst]It would be great if you could add txt2vid for up to 240 frame vids. I'm probab...

Furkan Gözükara SECourses•12/5/23, 1:45 PM

ye this is a hard task :d

FFurkan Gözükara SECourses looks like torch version related error

BCS/Taylor the Fox-Bear•12/5/23, 5:59 PM

I have this exact same issue. I have the indicated version of torch installed

python -c "import torch; print(torch.__version__)"
1.13.1+cu117

python -c "import torch; print(torch.__version__)"
1.13.1+cu117

Should we be using a newer version of torch?

BBCS/Taylor the Fox-Bear I have this exact same issue. I have the indicated version of torch installed `...

J.•12/5/23, 6:09 PM

Just to guess, probably need to be at 2.1? Lots of stuff seems to need > 2

Just looking at their github repo they do:
&& conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia \

So im assuming it just installs the latest

JJ.Just to guess, probably need to be at 2.1? Lots of stuff seems to need > 2 Just...

BCS/Taylor the Fox-Bear•12/5/23, 6:12 PM

I just upgraded to 2.1.1 just in case and it appears to be running but the problem is that Tortoise itself requests 1.13.1 from requirements.txt:

torch==1.13.1 ; python_version >= "3.8" and python_full_version != "3.9.7" and python_version < "3.12" \
    hashes
torchaudio==0.13.1 ; python_version >= "3.8" and python_full_version != "3.9.7" and python_version < "3.12" \
    hashes
torchdiffeq==0.2.3 ; python_version >= "3.8" and python_version < "3.12" and python_full_version != "3.9.7" \
    hashes
torchlibrosa==0.0.7 ; python_version >= "3.8" and python_full_version != "3.9.7" and python_version < "3.12" \
    hashes
torchsde==0.2.5 ; python_version >= "3.8" and python_version < "3.12" and python_full_version != "3.9.7" \
    hashes
torchvision==0.14.1 ; python_version >= "3.8" and python_full_version != "3.9.7" and python_version < "3.12" \
    hashes

torch==1.13.1 ; python_version >= "3.8" and python_full_version != "3.9.7" and python_version < "3.12" \
    hashes
torchaudio==0.13.1 ; python_version >= "3.8" and python_full_version != "3.9.7" and python_version < "3.12" \
    hashes
torchdiffeq==0.2.3 ; python_version >= "3.8" and python_version < "3.12" and python_full_version != "3.9.7" \
    hashes
torchlibrosa==0.0.7 ; python_version >= "3.8" and python_full_version != "3.9.7" and python_version < "3.12" \
    hashes
torchsde==0.2.5 ; python_version >= "3.8" and python_version < "3.12" and python_full_version != "3.9.7" \
    hashes
torchvision==0.14.1 ; python_version >= "3.8" and python_full_version != "3.9.7" and python_version < "3.12" \
    hashes

Hashes collapsed for brevity

FFurkan Gözükara SECourses hopefully soon. i just finished pixart

Deleted User•12/5/23, 6:20 PM

Thanks for your guide on Pixart. Ran it on RunPod and tested it, and yes, it is way better than SD. Probably because of the better encoder and the higher image quality training. Are you going to try and do a Dreambooth training ?
I will look into https://github.com/PixArt-alpha/PixArt-alpha/blob/master/asset/docs/pixart-dreambooth.md and see if it is any good.

GitHub

PixArt-alpha/asset/docs/pixart-dreambooth.md at master · PixArt-alp...

Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis - PixArt-alpha/PixArt-alpha

BCS/Taylor the Fox-Bear•12/5/23, 7:56 PM

Update: Torch 2.1.1 seems to be working, however it took some effort.

1) I was accessing my machine remotely via RPD. RDP doesn't use the GPU by default. You have to change registry settings to enable the GPU via RDP
2) once the GPU was actually active I had to force reinstall torch and its related packages and then check that Cuda was enabled with the following:

python -c "import torch; print(torch.cuda.is_available())"

python -c "import torch; print(torch.cuda.is_available())"

My CPU is much happier now.

BBCS/Taylor the Fox-Bear I have this exact same issue. I have the indicated version of torch installed `...

Furkan Gözükara SECourses•12/5/23, 10:03 PM

yes

DDeleted User Thanks for your guide on Pixart. Ran it on RunPod and tested it, and yes, it is ...

Furkan Gözükara SECourses•12/5/23, 10:03 PM

yes i am searching for it

BBCS/Taylor the Fox-Bear Update: Torch 2.1.1 seems to be working, however it took some effort. 1) I was ...

Furkan Gözükara SECourses•12/5/23, 10:04 PM

great

BBCS/Taylor the Fox-Bear Update: Torch 2.1.1 seems to be working, however it took some effort. 1) I was ...

ClocksmithOP•12/6/23, 1:20 AM

Can you tell me the command you used to upgrade to Torch 2.1.1? I'm not a regular Python user. The Github doc still only has this instruction and I'm not sure how to convert it and/or if I need to uninstall first

pip3 install torch==1.13.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117

pip3 install torch==1.13.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117

AndyWear•12/6/23, 4:28 AM

finally got around to installing kohya. Does it work for SD 1.5 models?

AndyWear•12/6/23, 4:29 AM

maybe i can just use this from now on instead of a1111

AndyWear•12/6/23, 4:33 AM

with kohya it looks like i can even use dreambooth. Do i even need the A1111 version anymore?

Furkan Gözükara SECourses•12/6/23, 11:49 AM

haha this is so amazing https://www.linkedin.com/posts/furkangozukara_haha-this-suno-generatic-songs-are-really-activity-7138128560824782848-6r1B?utm_source=share&utm_medium=member_desktop

CClocksmith Can you tell me the command you used to upgrade to Torch 2.1.1? I'm not a regula...

Furkan Gözükara SECourses•12/6/23, 11:49 AM

you need to activate venv and run command

Furkan Gözükara SECourses•12/6/23, 11:50 AM

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

AAndyWear finally got around to installing kohya. Does it work for SD 1.5 models?

Furkan Gözükara SECourses•12/6/23, 11:50 AM

yes

Furkan Gözükara SECourses•12/6/23, 11:50 AM

it works both sdxl and sd 1.5

AAndyWear with kohya it looks like i can even use dreambooth. Do i even need the A1111 ve...

Furkan Gözükara SECourses•12/6/23, 11:50 AM

you dont need

Furkan Gözükara SECourses•12/6/23, 11:50 AM

but for sd 1.5 sadly i dont have good parameters yet

Digital [Starburst]•12/6/23, 4:12 PM

I'm generating some pretty awesome images with your Pixart webui.

Digital [Starburst]•12/6/23, 4:13 PM

I've been studying up on writing good prompts. I had no idea you should add open-ended questions. I didn't know curly brackets mean you want it in the image and square brackets mean you don't want it in the image (I thought that's what negative prompts were for).

Digital [Starburst]•12/6/23, 4:14 PM

I have noob tattooed on my forehead.

FFurkan Gözükara SECourses haha this is so amazing https://www.linkedin.com/posts/furkangozukara_haha-this-...

J.•12/6/23, 5:10 PM

woah that is wild

Digital [Starburst]•12/6/23, 9:54 PM

I've messed around with the video generation tools one of which is broken but the one for ComfyUI works. Other than a short GIF file you aren't going to get much, and it won't be very good at that.

Digital [Starburst]•12/6/23, 9:55 PM

I guess I should've thought that if I get some of the results I end up with just with images then the possibility of getting much of a video aren't really all that great.

Digital [Starburst]•12/6/23, 9:57 PM

Someone 15 months ago took over the maintenance of text2video but I did my best to get it to work, and it has an issue with missing modules I know are properly installed with the correct version numbers and everything. I was hoping it would be a little more effective.

Digital [Starburst]•12/6/23, 10:15 PM

Deforum works great

FFurkan Gözükara SECourses you need to activate venv and run command

ClocksmithOP•12/7/23, 2:21 AM

Tried that and got the same error

CClocksmith Tried that and got the same error

Furkan Gözükara SECourses•12/7/23, 8:55 PM

i am helping members minimum gold level by connecting their pc if you need such help

(venv) C:\Users\username\Desktop\tortoise\tortoise-tts-fast\scripts>python tortoise_tts.py --preset fast --ar_checkpoint "C:\Users\username\Desktop\tortoise\DL-Art-School\experiments\voice-dlas\models\5000_gpt.pth" "Welcome to the software engineering courses channel." Traceback (most recent call last): File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\scripts\tortoise_tts.py", line 240, in <module> from tortoise.inference import ( File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\tortoise\inference.py", line 167, in <module> vfixer = VoiceFixer() File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\base.py", line 13, in __init__ self._model = voicefixer_fe(channels=2, sample_rate=44100) File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\restorer\model.py", line 180, in __init__ self.vocoder = Vocoder(sample_rate=44100) File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\vocoder\base.py", line 19, in __init__ self._load_pretrain(Config.ckpt) File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\vocoder\base.py", line 25, in _load_pretrain self.model = Generator(Config.cin_channels) File "C:\Users\username\Desktop\tortoise\tortoise-tts-fast\venv\lib\site-packages\voicefixer\vocoder\model\generator.py", line 34, in __init__ nn.utils.parametrizations.weight_norm( AttributeError: module 'torch.nn.utils.parametrizations' has no attribute 'weight_norm'

torch==1.13.1 ; python_version >= "3.8" and python_full_version != "3.9.7" and python_version < "3.12" \ hashes torchaudio==0.13.1 ; python_version >= "3.8" and python_full_version != "3.9.7" and python_version < "3.12" \ hashes torchdiffeq==0.2.3 ; python_version >= "3.8" and python_version < "3.12" and python_full_version != "3.9.7" \ hashes torchlibrosa==0.0.7 ; python_version >= "3.8" and python_full_version != "3.9.7" and python_version < "3.12" \ hashes torchsde==0.2.5 ; python_version >= "3.8" and python_version < "3.12" and python_full_version != "3.9.7" \ hashes torchvision==0.14.1 ; python_version >= "3.8" and python_full_version != "3.9.7" and python_version < "3.12" \ hashes

I'm at 4000 iterations and still have `loss_text_ce` values going down by 0.1 per every 800 iteratio

Similar Threads