>> TALKS, BUGS, SUGGESTIONS #3 about dfl-colab

[thefakereport]

yess!!! you are the best!

[iperov]

заходи в телеграм

[chervonij]

та я есть же в телеге )
такой же ник и такая же аватара

[salgado]

Hi, where can I find the workspace.zip file requested in your colab file (#Manage workspace)?

[chervonij]

@salgado
Hey. "workspace.zip" is the name of the archive file with your workspace.

It means that if you import from google drive it will be the name of the archive that will be imported from the drive. If you export, then an archive with that filename will be created on your drive.

In addition, workspace.zip is the default filename. but you can change it at any convenient for you. this is useful if you store several workspaces on a drive.

[3mlZ68s]

del

[jcisz]

Hi, while running the program, can i pause it like the desktop version and then resume the process? Thanks

[3mlZ68s]

->фейк голоса
есть подозрения, что можно обойтись обычным автоэнкодером. (паперы по теме подтверждают). главное маску правильно подобрать. Всмысле, разбивать речь на элементарные фонемы.

Хотя, наверное, нечестно было отвечать на такое старое сообщение, всмысле, что теперь это и так очевидно.

П.С. Справедливости ради, я ещё летом пытался прям с помощью DFL-моделей звук фейчить, но мне тогда скилла тем более не хватило...

[3mlZ68s]

Ещё с прошлого треда хочу похвастаться(лучше поздно). Мне Colab ,короче, выдал новую машину с HBМ памятью. Для ТПУ. Такой:"Бро, вижу твой датасет не помещается в память, давай мы тебе новый ПК выдадим". Я такой:"Ок". Вначале дали 25 гигабайт оперативки, сейчас смотрю - 35 уже. Больше, чем жесткий диск(свободное место). Если так пойдёт дальше, то я смогу какой-нибудь ImageNet сразу в память грузить и потом на ТПУ-ху перекидывать.

Я к тому, что да, соглашусь, всё-таки забавно со стороны выглядит. То у некоторых T4 отбирают , то такие "подгоны" делают.

[chervonij]

@jcisz
Hey. No, but you can stop training and manually export your workspace to google drive.
So, you can continue training later.

[chervonij]

@3mlZ68s
Если честно, то уже не помню о чем была речь по фейку голоса )
Могу ошибаться, но вроде сейчас нет достойных решений в speech-to-speech

Та какой там Т4. Мне последнюю неделю Р100 постоянно дают
Пусть оставят Т4 себе ))

[3mlZ68s]

Ну я P100 тоже имел ввиду. Хорошо, если её не забирают.

Могу ошибаться, но вроде сейчас нет достойных решений в speech-to-speech

Upd:

я ещё летом пытался прям с помощью DFL-моделей звук фейчить, но мне тогда скилла тем более не хватило...

[3mlZ68s]

:(

[chervonij]

@3mlZ68s
по фейку звука там RNN походу нужна
какой-нибудь LSTM Autoencoder

[chervonij]

основная проблема в том, что бы правильно разбить сэмплы
чем мельче, тем лучше
разбивать на простейшие фонемы может быть недостаточно

[3mlZ68s]

разбивать на простейшие фонемы может быть недостаточно

да, у меня похожие мысли были. Ну тут лучше договориться что считать фонемой.

Насчет LSTM...

Мне надо немного подумать над всем этим :-)

[MrFunny6969]

Please fix this bug

[MrFunny6969]

Mounted at /content/drive
/content
[workspace.zip]
End-of-central-directory signature not found. Either this file is not
a zipfile, or it constitutes one disk of a multi-part archive. In the
latter case the central directory and zipfile comment will be found on
the last disk(s) of this archive.
unzip: cannot find zipfile directory in one of workspace.zip or
workspace.zip.zip, and cannot find workspace.zip.ZIP, period.
Done!

[MrFunny6969]

I had the workspace.zip tho in my Drive

[chervonij]

@MrFunny6969
It's not a bug.
Your archive was corrupted.

[MrFunny6969]

Seems like Colab or Drive divided my zip file to a multi-part one, Thanks for the info tho! 😊😊

[3mlZ68s]

Посмотрел я фейк-аудио проекты и репы (с чего, наверное, и надо было начать), почитал описание. Смотрел всё до кучи - и text-to-speech, и speech-to-speech. И если проект может и то и другое, то подходы обычно не сильно отличаются.

Tacotron и WaveNet - оба проекта активно используют LSTM. И сейчас WaveNet объективно
показывает лучшие результаты. Особенно после последних обновлений.
DeepVoice - почти не смотрел. Из того, что смотрел - активно используется RNN.
DC-TTS - а вот dctts использует энкодеры и декодеры. И потом на выходе прикручен ХайРез. Очень приличный результат, несмотря на относительную простоту сети, но только t2s.

[iperov]

@3mlZ68s фейчить можно любой язык? ты в телеграм канале есть?

[3mlZ68s]

@3mlZ68s фейчить можно любой язык? ты в телеграм канале есть?

ну на русском по-крайней мере можно. Я в телеграмме read-only.

[jcisz]

Hi, im getting this error

Time to backup: 11 hours
Traceback (most recent call last):
  File "DeepFaceLab/main.py", line 19, in <module>
    multiprocessing.set_start_method("spawn")
  File "/usr/lib/python3.6/multiprocessing/context.py", line 242, in set_start_method
    raise RuntimeError('context has already been set')
RuntimeError: context has already been set

Its reported in the repo but idk if it's a notebook problem
iperov/issues/528

[MrFunny6969]

Same bro

[chervonij]

It seems that problem is in DFL
I wrote to iperov about this

[chervonij]

@MrFunny6969 @jcisz
Fixed.

[3mlZ68s]

@iperov
ну приличное качество - это тоже относительно. И две минуты аудио будет недостаточно. Короче проблем и минусов тоже хватает.

[rlkiran]

/content
Duplicate filenames are not allowed, skipping: IMG_2514-gigapixel.PNG
Performing 1st pass...
Running on Tesla P4.
2019-12-24 10:13:42.309390: E tensorflow/stream_executor/cuda/cuda_driver.cc:300] failed call to cuInit: CUDA_ERROR_NOT_INITIALIZED: initialization error
Exception: Traceback (most recent call last):
File "/content/DeepFaceLab/joblib/SubprocessorBase.py", line 59, in _subprocess_run
self.on_initialize(client_dict)
File "/content/DeepFaceLab/mainscripts/Extractor.py", line 77, in on_initialize
nnlib.import_all (device_config)
File "/content/DeepFaceLab/nnlib/nnlib.py", line 1410, in import_all
nnlib.import_keras(device_config)
File "/content/DeepFaceLab/nnlib/nnlib.py", line 208, in import_keras
nnlib._import_tf(device_config)
File "/content/DeepFaceLab/nnlib/nnlib.py", line 196, in _import_tf
nnlib.tf_sess = tf.Session(config=config)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/client/session.py", line 1551, in init
super(Session, self).init(target, graph, config=config)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/client/session.py", line 676, in init
self._session = tf_session.TF_NewSessionRef(self._graph._c_graph, opts)
tensorflow.python.framework.errors_impl.InvalidArgumentError: 'visible_device_list' listed an invalid GPU id '0' but visible device count is -1

[rlkiran]

How to solve this error?

[thefakereport]

hey there. I cannot convert using FAN-DST

[rlkiran]

ANYBODY PLEASE EXPLAIN HOW TO USE HFS WITH GOOGLE COLAB?

[thefakereport]

this is the error message for converting with fan - dst

converting: 9% 148/1662 [00:03<00:39, 38.01it/s]2019-12-24 14:00:31.420531: E tensorflow/core/grappler/clusters/utils.cc:83] Failed to get device properties, error code: 3
2019-12-24 14:00:31.421019: E tensorflow/core/grappler/clusters/utils.cc:83] Failed to get device properties, error code: 3
CPU0 doesnt response, terminating it.
CPU1 doesnt response, terminating it.

[chervonij]

@rlkiran
Fixed in last DFL commit.

[chervonij]

About the HFS. Is it really so difficult to google, find the official HFS website and read the instructions?

[chervonij]

@thefakereport
Is this error only with FAN masks?

[rlkiran]

About the HFS. Is it really so difficult to google, find the official HFS website and read the instructions?

I tried but whenever a connection is established it disconnects immediately..
I tried it with ngrok but still it refuses to connect

[chervonij]

@rlkiran
Maybe problem in your AV settings.
Also, maybe it's need to port forward in your router settings

[thefakereport]

@chervonij yes, the problem is only with FAN masks, learned works just fine. everything else works perfectly since yesterday

[chervonij]

@thefakereport
To be honest, I can’t understand what could be the problem.
I can not reproduce the error. I checked on several videos and FAN masks worked without errors.

Which model are you using?

[thefakereport]

@chervonij
the problem is while converting. I use options 1. Overlay / 2. Mask: Fan - DST. but the fan dst option give me an error. My model is SAEHD

"
converting: 9% 148/1662 [00:03<00:39, 38.01it/s]

**2019-12-24 14:00:31.420531: E tensorflow/core/grappler/clusters/utils.cc:83] Failed to get device properties, error code: 3

2019-12-24 14:00:31.421019: E tensorflow/core/grappler/clusters/utils.cc:83] Failed to get device properties, error code: 3**

CPU0 doesnt response, terminating it.
CPU1 doesnt response, terminating it.
"

[thefakereport]

and it is working again! Thanks @chervonij and happy chritsmass!

[thefakereport]

are we getting this on colab?
please! thanks!!!
DeepFaceLab : released fully automatic multi-gpu faceset enhancer.

[chervonij]

@thefakereport
30 minutes ago DFL-Colab was updated. Faceset Enhancer was added
Merry Christmas ))

[chervonij]

@thefakereport
Not me. I just added this for the colab. iperov did a great job to make this possible.
If you can, then thank him with a donate. Because, he makes these amazing things free for you.

[thefakereport]

how it works? it is automatic?

where can I donate? he is truly something

[chervonij]

@thefakereport
Yes, it's automatic. In module "Extract" you can see "Faceset Enhancer" block with choice data_src/data_dst. In the end of process you will have the same directory with aligned, but enhanced images.

This is official DFL donation link by PayPal: Link
You can find it on main page of DFL: Link

[thefakereport]

this error pop up_

Traceback (most recent call last):
File "/content/DeepFaceLab/mainscripts/Trainer.py", line 50, in trainerThread
device_args=device_args)
File "/content/DeepFaceLab/models/ModelBase.py", line 145, in init
self.onInitialize()
File "/content/DeepFaceLab/models/Model_SAEHD/Model.py", line 429, in onInitialize
loaded, not_loaded = self.load_weights_safe(not_loaded)
File "/content/DeepFaceLab/models/ModelBase.py", line 423, in load_weights_safe
model.load_weights(filename)
File "/usr/local/lib/python3.6/dist-packages/keras/engine/network.py", line 1166, in load_weights
f, self.layers, reshape=reshape)
File "/usr/local/lib/python3.6/dist-packages/keras/engine/saving.py", line 1058, in load_weights_from_hdf5_group
K.batch_set_value(weight_value_tuples)
File "/usr/local/lib/python3.6/dist-packages/keras/backend/tensorflow_backend.py", line 2465, in batch_set_value
assign_op = x.assign(assign_placeholder)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/variables.py", line 1762, in assign
name=name)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/state_ops.py", line 223, in assign
validate_shape=validate_shape)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/gen_state_ops.py", line 64, in assign
use_locking=use_locking, name=name)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/op_def_library.py", line 788, in _apply_op_helper
op_def=op_def)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/util/deprecation.py", line 507, in new_func
return func(*args, **kwargs)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/ops.py", line 3300, in create_op
op_def=op_def)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/ops.py", line 1823, in init
control_input_ops)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/ops.py", line 1662, in _create_c_op
raise ValueError(str(e))
ValueError: Dimension 0 in both shapes must be equal, but are 1 and 3. Shapes are [1,1,128,3] and [3,128,5,5]. for 'Assign_74' (op: 'Assign') with input shapes: [1,1,128,3], [3,128,5,5].

[chervonij]

@thefakereport
Beacause SAEHD architecture was updated.

[thefakereport]

so.... i have to restart a new model??

[chervonij]

@thefakereport
Last commit in DFL repository.

[3mlZ68s]

С НГ!

[chervonij]

@3mlZ68s

И тебя )
Добрых градиентов в этом году и поменьше локальных минимумов ))

[xstrauss]

I trained it 7+5 hours but it is still blurry. What do you suggest?

[xstrauss]

how can I train one workspace with two accounts?

[IEWbgfnYDwHRoRRSKtkdyMDUzgdwuBYgDKtDJWd]

I would not push your luck using so many accounts/resources for deepfakes on colab. ^

[xstrauss]

Can I use OneDrive instead Google Drive?

[xstrauss]

I would not push your luck using so many accounts/resources for deepfakes on colab. ^

I just asking :)

[xstrauss]

How can I change the training options while it is training?

[3mlZ68s]

И тебя )
Добрых градиентов в этом году и поменьше локальных минимумов ))

👍

[xstrauss]

which sharpen mode do you sugggest?

[chervonij]

@xstrauss

I trained it 7+5 hours but it is still blurry. What do you suggest?

which sharpen mode do you sugggest?

How can I change the training options while it is training?

I suggest you read manual.
If it's you wrote same questions to DeepFaceLab chat on Telegram, then I already (few times) answer to you, and you still ask stupid questions.

This thread is called "TALKS, BUGS, SUGGESTIONS" and not "Help desk" or "Consulting".
If you want to understand process of creating deepfakes, then read manuals.

If you spam here, then I will block you in this repository.

Can I use OneDrive instead Google Drive?

Nope. Oddly enough, but you cannot use the service from Microsoft in a service from Google.
This is just logical.

[chervonij]

I do not want to be rude, but I am not a consultant.

[chervonij]

@IEWbgfnYDwHRoRRSKtkdyMDUzgdwuBYgDKtDJWd

I would not push your luck using so many accounts/resources for deepfakes on colab. ^

Not much needed. Only two accounts.
Otherwise, if you use the Colab 24/7, then Google will block your access to the GPU backend.

[3mlZ68s]

Let It Be.
dc_tts

https://arxiv.org/pdf/1710.08969.pdf

[3mlZ68s]

Добрые люди уже нафейчили - грех не воспользоваться.

https://soundcloud.com/user-694498525/sets/tts-examples

Авторство Ю. Петрова. По этой фамилии и ключевым словам гуглится
пошаговая инструкция по dc_tts.

[thefakereport]

@chervonij quick question, is it possible to have manual masking and manual extraction on colab?
Thanks!!!!

[chervonij]

@thefakereport
Hey. Unfortunately, due to interface limitations, this is not possible.

[chervonij]

@3mlZ68s
Так это же text-to-speech
По этому направлению уже давно наработки есть

[HeeraSinghKang]

Can we move trained model from colab to local system and resume training?

[chervonij]

@HeeraSinghKang
Yes, but you need a latest DFL version on your local system.

[3mlZ68s]

@chervonij
Да, я написал раньше.

но только t2s.

Короче тут дальше экспериментировать надо, чем я и пытаюсь сейчас заниматься.
Хотя может я не вижу очевидных причин, почему не получится dc_tts в
voice-to-voice переделать.

[3mlZ68s]

Повторюсь, опубликовав старое сообщение. Я его написал, но че-то удалил, чтоб профаном не выглядеть.

"Я ещё летом пытался прям с помощью DFL-моделей звук фейчить.
Мое видение было простым(на бумаге).
Оставить всё то же самое, но сделать/найти SampleProcessor и mtcnn для аудио.(ну не совсем так, но ,думаю, понятно что имеется ввиду)
Т.е. вместо row/pitch головы будет высота звука, например, и т.д.

это если совсем на пальцах объяснять."

С тех пор видение принципиально не поменялась. Ключевое слово - принципиально.

[thefakereport]

@chervonij DFL 2.0 please, you are my man

[chervonij]

@3mlZ68s
ну, переделать модель t2s в s2s, это уже поменять модель
я думаю, что будет большая разница, даже в принципах работы сети
текст можно достаточно точно переводить в речь потому что там есть контекст
прикол s2s, что бы не зависить от контекста

хотел заняться VoiceCloning
но пока нет времени
я сейчас в отпуске и + DFL 2.0 в релиз вышла
буду пока этим заниматься )

[chervonij]

@thefakereport
Tomorrow = )

[thefakereport]

@chervonij i am happy for the new model and sad for loosing my actual one... thanks anyway, can't wait!

[thefakereport]

@chervonij it is possible to mantain both DFL1.0 and DFL2.0 online for a few days? I have 3 projects on the run i am going to loose them... if not i will start over but just asking, thanks!

[chervonij]

@thefakereport
After upgrade there will be two versions.
DFL-Colab which uses DFL 1.0
And which uses DFL 2.0

[vfxy]

on main page github.com/chervonij/DFL-Colab
links are broken in here:
DeepFaceLab Official Manual: English / Russian
thank you!

[chervonij]

@vfxy
thanks, I will remove it
official DFL manuals not available anymore

[vfxy]

@chervonij no worries, can i still get it somewhere to learn, please? any language...

[chervonij]

@vfxy
on main page you can find guides links

[chervonij]

@thefakereport
Updated.
New links to notebooks on repo main page.

[thefakereport]

@chervonij you are santa claus! thanks!

[zakaria47fs]

Hi @chervonij I am a Deepfacelab user with colab, I am working on AVATAR mode, but I couldn't move to the stage 2, I pressed "Enter" doesn t worked, when I interrupt the cell and relaunches it again so it gives me the choice to choose stage 2, it continues instead on the training of the stage 1,
Any suggestions pllzz

[chervonij]

@zakaria47fs
Replied in Telegram

[zakaria47fs]

Hi @chervonij
Why we can t choose the AVATAR TYPE ! it goes direcctly for "head" type, I m running it on colab

[zakaria47fs]

@chervonij in response to my question just above, I checked the deepfacelab code and I have found that @iperov has comment the 'avatar_type' section out so that the model takes directly 'head' as avaar type

[chervonij]

@zakaria47fs
And what?
If iperov did it, then it was necessary.

[zakaria47fs]

I don't think so, I have uncommented it in colab and it worked for me fine

…

On Fri, Jan 24, 2020, 14:15 chervonij ***@***.***> wrote: @zakaria47fs <https://github.com/zakaria47fs> And what? If iperov did it, then it was necessary. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#14?email_source=notifications&email_token=ANFZZCOBRBCAMPFOSIFFNA3Q7LSXHA5CNFSM4JXMUH72YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEJ2X34Q#issuecomment-578125298>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANFZZCNHALYFRJNEX2ZKISTQ7LSXHANCNFSM4JXMUH7Q> .

[iperov]

@3mlZ68s тут вот описание real time voice cloning https://proglib.io/p/golosovoy-deepfake-ili-kak-rabotaet-tehnologiya-klonirovaniya-golosa-2019-12-11

т.е. если настругать из ютуба путинских фраз с субтитрами к сотням видео (субтитры выкачиваются отдельно)
то можно и путина зафейчить?

Я ещё летом пытался прям с помощью DFL-моделей звук фейчить.

прочекай архитектуру дфл 2.0 и мою либу leras - надстройка над сырым tf.
Теперь полная свобода работы с тензорами.

[3mlZ68s]

@iperov

т.е. если настругать из ютуба путинских фраз с субтитрами к сотням видео (субтитры выкачиваются отдельно) то можно и путина зафейчить?

Вообще я смотрел в сторону стенограмм. Но да, почему нет. Если качество соответствует.

прочекай архитектуру дфл 2.0 и мою либу leras

ок

[chervonij]

@juliuselsh
everything is working fine.
check the name of archive you are importing.
also, keep in mind that there is a slight delay between Colab and Drive, so Colab "sees" your archive some time after you upload it to the Drive.

[pisces300300]

small issue, and im sorry i cannot remember the exact details, but i noticed i had an "aligned" and also "aliged" folder created during the process, so somewhere in the code is a typo i think. also i cannot remember if that is the exact misspelling.

one suggestion is to change the preview from updating every 10 iterations to maybe 20 or a customizable number as it would increase training speeds by a fairly large amount, possibly up to around 20-30% with a 16gb collab card if my staring at the iteration counter and doing quick maths in my head is anywhere close to accurate

also thanks you guys for all that you do, really

[chervonij]

@pisces300300

If you don’t remember, then why write?
If there was a typo, then nothing would work.
Believe it when I see it.

It does not increase speed.
Do you have any idea how neural networks work?
This is a preview. It does not need as many resources as for training.

[pisces300300]

no i dont know how neural networks work, but i know that training pauses for a couple seconds at every 10 iterations and i presume that is to create the preview jpg, but maybe that is incorrect.

also per the typo, it was this folder: aliged_debug created using dfl v1 sae, so disregard as i know you no longer support v1 and thus it is a non issue

[chervonij]

@pisces300300
I have never seen a training session stop for a couple of seconds.
maybe a couple of milliseconds?

[pisces300300]

yeah idk man. mine stops for about 4 seconds every 10 iterations on collab with a 16gb card. if the preview is created by the cpu instead of the gpu then it would make sense because i use an old laptop.

anyways, ill stop annoying you with this.