Search found 790 matches

by bryanlyon
Wed Jan 18, 2023 5:56 am
Forum: Extract Support
Topic: CRIT ERROR after ~52000 iterations : Can't decrement id ref count
Replies: 3
Views: 1741

Re: CRIT ERROR after ~52000 iterations : Can't decrement id ref count

Unfortunately the model's h5 files have corrupted. Can happen if there is not enough storage space or if the system crashes in the middle of a write. Luckily we do backups, use the "restore" tool to see if the backup is valid, and if so you can continue from there. If not, you may need to ...
by bryanlyon
Mon Jan 16, 2023 12:44 am
Forum: Training Discussion
Topic: Is MP always a VRAM saver / performance enhancer?
Replies: 3
Views: 772

Re: Is MP always a VRAM saver / performance enhancer?

Mixed precision is not a simple 1/2 the vram saved unfortunately. It uses "mixed" precision which means some parts are kept in fp32 and some are in fp16. In fact, recent versions of Tensorflow have even done some stuff into TF32 which is a weird Nvidia-only 19bit float replacement (see htt...
by bryanlyon
Mon Jan 16, 2023 12:37 am
Forum: Extract Support
Topic: OOM killer kills faceswap process because it uses all 32GB of installed RAM (Not VRAM).
Replies: 6
Views: 1861

Re: OOM killer kills faceswap process because it uses all 32GB of installed RAM (Not VRAM).

Thanks, we'll look at those. But know that this is something we probably can't duplicate. I'd suggest giving it a week or so and trying again with a clean install. Most of the libraries we use are very popular and any memory leaks are likely to be fixed pretty quickly.

by bryanlyon
Sun Jan 15, 2023 10:13 pm
Forum: Extract Support
Topic: OOM killer kills faceswap process because it uses all 32GB of installed RAM (Not VRAM).
Replies: 6
Views: 1861

Re: OOM killer kills faceswap process because it uses all 32GB of installed RAM (Not VRAM).

Memory leaks are usually always in the libraries we use (and with this only happening to you it seems exceptionally likely). I'd suggest trying to update FaceSwap using the GUI which also updates the conda environment to the latest versions. If that doesn't work you can try to reinstall so you get t...
by bryanlyon
Fri Jan 13, 2023 6:44 am
Forum: Training Discussion
Topic: Is MP always a VRAM saver / performance enhancer?
Replies: 3
Views: 772

Re: Is MP always a VRAM saver / performance enhancer?

Mixed precision will ALWAYS use less VRAM. It however, will not always be faster. That does depend on your GPU and system. If you're on Windows a LOT of things can use your VRAM, even just opening the start menu takes some. In the end, it's very possible that your web browser reserved some memory an...
by bryanlyon
Thu Jan 12, 2023 8:26 pm
Forum: Training Discussion
Topic: Distribution and Multiple GPUS
Replies: 5
Views: 1147

Re: Distribution and Multiple GPUS

It's probably due to CPU usage and PCIe bandwidth. Most motherboards have full lanes for each slot, but some don't and even with full PCIe lanes available many motherboards have "crosstalk" that limits full speed when multiple slots are being used extensively. Also CPU limitations exist. F...
by bryanlyon
Thu Jan 12, 2023 6:52 pm
Forum: Training Discussion
Topic: Distribution and Multiple GPUS
Replies: 5
Views: 1147

Re: Distribution and Multiple GPUS

Tensorflow has recently combined memory of all GPUs into one pool. By default, TensorFlow maps nearly all of the GPU memory of all GPUs (subject to CUDA_VISIBLE_DEVICES) visible to the process. This is done to more efficiently use the relatively precious GPU memory resources on the devices by reduci...
by bryanlyon
Thu Jan 12, 2023 6:23 am
Forum: General Discussion
Topic: DirectML
Replies: 2
Views: 1848

Re: DirectML

DirectML is replacing PlaidML for AMD support on Windows. If you're using an Nvidia card we HIGHLY recommend you stick with Cuda and the Nvidia installation options (though DirectML should work for Nvidia cards, it will be slower and have more overhead). DirectML will be completely replacing PlaidML...
by bryanlyon
Thu Jan 12, 2023 12:31 am
Forum: Hardware
Topic: odd power supply noise during training
Replies: 3
Views: 3934

Re: odd power supply noise during training

The noise of your GPU is called "Coil whine" and as you said, it's the circuitry. It's just the sound of the board vibrating due to voltage moving through it and is nothing to worry about. It's one of those "good problems" to have, since it means the rest of your system is so qui...
by bryanlyon
Mon Dec 26, 2022 6:13 pm
Forum: Convert Support
Topic: Convert crash
Replies: 2
Views: 4125

Re: Convert crash

Error 32 means that ffmpeg is falling to write it the file. Most common reasons are:

  • Out of disk space

  • Invalid output path

  • Unwriteable path

But it could be anything that prevents ffmpeg from writing out the video.

by bryanlyon
Fri Dec 23, 2022 3:58 pm
Forum: Installation Support
Topic: The program crashes in the installation process
Replies: 2
Views: 1637

Re: The program crashes in the installation process

On the specs: https://support.microsoft.com/en-us/sur ... 43e68ad96b

Either no GPU or a 2gb 1050.

Probably not the best laptop to run Faceswap, but not the reason it fails to install.

by bryanlyon
Fri Dec 23, 2022 12:19 am
Forum: Training Support
Topic: Manual editing of png files without alignment file
Replies: 1
Views: 793

Re: Manual editing of png files without alignment file

The manual tool requires the original video to operate. That's why it needs the alignment file. The individual images contain their own data, but not the information from the original video. Besides, once you edit the alignment data, you'll need to re-extract because the data would change the image.
by bryanlyon
Wed Dec 21, 2022 3:51 am
Forum: Extract Discussion
Topic: separating re-feeds from false positives
Replies: 13
Views: 6595

Re: separating re-feeds from false positives

When you do an additional extract, you have to make sure you do not extract into the same folder. It re-extracts all files from the alignments. This is probably the cause of your issues.

by bryanlyon
Mon Dec 19, 2022 10:42 pm
Forum: Extract Discussion
Topic: separating re-feeds from false positives
Replies: 13
Views: 6595

Re: separating re-feeds from false positives

One time I've seen this sort of thing was when I was working with a .webm file. It turns out that webm has weird functionality and will often duplicate frames (up to 1000 frames per second) meaning that FS sees identical frames over and over. But even then we increment the frame number. If this is w...
by bryanlyon
Mon Dec 12, 2022 6:38 pm
Forum: Hardware
Topic: Hardware best practices
Replies: 83
Views: 150148

Re: Hardware best practices

No. I tend to stick with recommending the 60 level of the GPUs. High end GPUs cost a LOT more for a BIT of better speed. Generally the 60-70 offers the best value. If you just want a top-end GPU then by all means get a 3090 ti over a 4080. The 4080 is a new card and it usually takes a while to get g...
by bryanlyon
Sat Dec 10, 2022 5:35 am
Forum: Training Discussion
Topic: Shorter Training=Smaller Chances of NaNs; A hypothesis by an Uneducated Idiot
Replies: 5
Views: 1422

Re: Shorter Training=Smaller Chances of NaNs; A hypothesis by an Uneducated Idiot

Very interesting. If someone else can try and verify these results I'd appreciate it. I don't typically get many NaNs so cannot test this well. This may justify an option in the training script if others can replicate.

by bryanlyon
Sat Dec 10, 2022 4:03 am
Forum: Hardware
Topic: Hardware best practices
Replies: 83
Views: 150148

Re: Hardware best practices

Thank you for the reply. How are the RAM temps on your 3090? I read that VRAM intensive applications (e.g., ETH mining) caused dangerous temperature (>100 C memory junction temperature) levels on many 3090s, and required additional cooling to the back of the card. I wonder if FS has a similar type ...
by bryanlyon
Sun Dec 04, 2022 11:19 am
Forum: Training Discussion
Topic: [Tip] Why you need more sources (with examples)
Replies: 2
Views: 19663

Re: [Tip] Why you need more sources (with examples)

The length of the video doesn't matter at all. You're training on faces, not a video. A short video doesn't take any less time than a long one since you'll have to train a model with sufficient data before you can use it to convert.

by bryanlyon
Sun Dec 04, 2022 12:16 am
Forum: Convert Support
Topic: Can a model be used on two different clips of the same person?
Replies: 1
Views: 3618

Re: Can a model be used on two different clips of the same person?

When you train a model, you train it on 2 identities. You can then swap ANY videos with those identities. People like to "fine tune" a model on the swap video to make sure it's got the right exposure to that type of content, it's not required though.