OpenFaceSwap model

Got questions or tips about the Conversion process? This is the place to discuss them.


Forum rules

Read the FAQs and search the forum before posting a new topic.

This forum is for discussing tips and understanding the process involved with Converting faces from your trained model.

If you are having issues with the Convert process not working as you would expect, then you should post in the Convert Support forum.

Please mark any answers that fixed your problems so others can find the solutions.

Locked
User avatar
sykym
Posts: 6
Joined: Thu Jul 23, 2020 9:51 pm
Location: Italy

OpenFaceSwap model

Post by sykym »

I have trained a model with OpenFaceSwap. Is it possible to use it with/convert it with Faceswap?

User avatar
torzdf
Posts: 2649
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 129 times
Been thanked: 622 times

Re: OpenFaceSwap model

Post by torzdf »

The answer to that is "possibly".....

OpenFaceSwap has been dead for a long while, but IIRC it was based on an old version of Faceswap....

We have tried to maintain backwards compatibility, so running it in Faceswap should update the model to the latest code.

That said, if you can screengrab your model folder, I should be able to tell you for sure.

My word is final

User avatar
sykym
Posts: 6
Joined: Thu Jul 23, 2020 9:51 pm
Location: Italy

Re: OpenFaceSwap model

Post by sykym »

Model files: (Name)(Last modified)(Type)(Size)(Created)

model files
model files
Annotazione 2020-07-24 180943.jpg (20.93 KiB) Viewed 20482 times

Train command used on OFS:
python\scripts\python.bat faceswap_lowmem\faceswap.py train -A "C:\Users\*\A\faces" -B "C:\Users\*\B\faces" -m "C:\Users\***\model" -t LowMem -bs 16

I trained for 2-4 hours a day for almost 2 weeks until the weight loss got stuck at 0.02xx

I have NVIDIA GeForce 920M with 2GB

I tried renaming the files to lightweight_* and run the training, but it gave me this error:

Code: Select all

07/24/2020 19:45:40 MainProcess     _training_0     multithreading  run                       DEBUG    Error in thread (_training_0): Tensor("Placeholder:0", shape=(1024, 512, 3, 3), dtype=float32) must be from the same graph as Tensor("upscale_8_0_conv2d/kernel:0", shape=(3, 3, 256, 2048), dtype=float32_ref).
07/24/2020 19:45:41 MainProcess     MainThread      train           _monitor                  DEBUG    Thread error detected
07/24/2020 19:45:41 MainProcess     MainThread      train           _monitor                  DEBUG    Closed Monitor
07/24/2020 19:45:41 MainProcess     MainThread      train           _end_thread               DEBUG    Ending Training thread
07/24/2020 19:45:41 MainProcess     MainThread      train           _end_thread               CRITICAL Error caught! Exiting...
07/24/2020 19:45:41 MainProcess     MainThread      multithreading  join                      DEBUG    Joining Threads: '_training'
07/24/2020 19:45:41 MainProcess     MainThread      multithreading  join                      DEBUG    Joining Thread: '_training_0'
07/24/2020 19:45:41 MainProcess     MainThread      multithreading  join                      ERROR    Caught exception in thread: '_training_0'
Traceback (most recent call last):
  File "C:\Users\***\faceswap\plugins\train\model\_base.py", line 834, in load
    network = load_model(self.filename, custom_objects=get_custom_objects())
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\keras\engine\saving.py", line 419, in load_model
    model = _deserialize_model(f, custom_objects, compile)
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\keras\engine\saving.py", line 221, in _deserialize_model
    model_config = f['model_config']
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\keras\utils\io_utils.py", line 302, in __getitem__
    raise ValueError('Cannot create group in read only mode.')
ValueError: Cannot create group in read only mode.

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "C:\Users\***\faceswap\lib\cli\launcher.py", line 155, in execute_script
    process.process()
  File "C:\Users\***\faceswap\scripts\train.py", line 161, in process
    self._end_thread(thread, err)
  File "C:\Users\***\faceswap\scripts\train.py", line 201, in _end_thread
    thread.join()
  File "C:\Users\***\faceswap\lib\multithreading.py", line 121, in join
    raise thread.err[1].with_traceback(thread.err[2])
  File "C:\Users\***\faceswap\lib\multithreading.py", line 37, in run
    self._target(*self._args, **self._kwargs)
  File "C:\Users\***\faceswap\scripts\train.py", line 226, in _training
    raise err
  File "C:\Users\***\faceswap\scripts\train.py", line 214, in _training
    model = self._load_model()
  File "C:\Users\***\faceswap\scripts\train.py", line 255, in _load_model
    predict=False)
  File "C:\Users\***\faceswap\plugins\train\model\lightweight.py", line 20, in __init__
    super().__init__(*args, **kwargs)
  File "C:\Users\***\faceswap\plugins\train\model\original.py", line 25, in __init__
    super().__init__(*args, **kwargs)
  File "C:\Users\***\faceswap\plugins\train\model\_base.py", line 125, in __init__
    self.build()
  File "C:\Users\***\faceswap\plugins\train\model\_base.py", line 244, in build
    self.load_models(swapped=False)
  File "C:\Users\***\faceswap\plugins\train\model\_base.py", line 456, in load_models
    is_loaded = network.load(fullpath=model_mapping[network.side][network.type])
  File "C:\Users\***\faceswap\plugins\train\model\_base.py", line 837, in load
    self.convert_legacy_weights()
  File "C:\Users\***\faceswap\plugins\train\model\_base.py", line 863, in convert_legacy_weights
    self.network.load_weights(self.filename)
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\keras\engine\network.py", line 1166, in load_weights
    f, self.layers, reshape=reshape)
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\keras\engine\saving.py", line 1058, in load_weights_from_hdf5_group
    K.batch_set_value(weight_value_tuples)
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\keras\backend\tensorflow_backend.py", line 2465, in batch_set_value
    assign_op = x.assign(assign_placeholder)
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\tensorflow_core\python\ops\variables.py", line 2067, in assign
    self._variable, value, use_locking=use_locking, name=name)
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\tensorflow_core\python\ops\state_ops.py", line 227, in assign
    validate_shape=validate_shape)
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\tensorflow_core\python\ops\gen_state_ops.py", line 66, in assign
    use_locking=use_locking, name=name)
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\tensorflow_core\python\framework\op_def_library.py", line 367, in _apply_op_helper
    g = ops._get_graph_from_inputs(_Flatten(keywords.values()))
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\tensorflow_core\python\framework\ops.py", line 5979, in _get_graph_from_inputs
    _assert_same_graph(original_graph_element, graph_element)
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\tensorflow_core\python\framework\ops.py", line 5914, in _assert_same_graph
    (item, original_item))
ValueError: Tensor("Placeholder:0", shape=(1024, 512, 3, 3), dtype=float32) must be from the same graph as Tensor("upscale_8_0_conv2d/kernel:0", shape=(3, 3, 256, 2048), dtype=float32_ref).

EDIT: Forgot to share the settings

Code: Select all

{
  "name": "lightweight",
  "sessions": {
    "1": {
      "timestamp": 1595608900.6572073,
      "no_logs": false,
      "pingpong": false,
      "loss_names": {
        "a": [
          "face_loss"
        ],
        "b": [
          "face_loss"
        ]
      },
      "batchsize": 16,
      "iterations": 9,
      "config": {
        "learning_rate": 5e-05
      }
    }
  },
  "lowest_avg_loss": {
    "a": 0.20435872860252857,
    "b": 0.15313835069537163
  },
  "iterations": 9,
  "inputs": {
    "face_in:0": [
      64,
      64,
      3
    ]
  },
  "training_size": 256,
  "config": {
    "coverage": 68.75,
    "mask_type": null,
    "mask_blur_kernel": 3,
    "mask_threshold": 4,
    "learn_mask": false,
    "icnr_init": false,
    "conv_aware_init": false,
    "reflect_padding": false,
    "penalized_mask_loss": true,
    "loss_function": "mae",
    "learning_rate": 5e-05
  }
}
Last edited by sykym on Sat Jul 25, 2020 12:47 am, edited 1 time in total.
User avatar
sykym
Posts: 6
Joined: Thu Jul 23, 2020 9:51 pm
Location: Italy

Re: OpenFaceSwap model

Post by sykym »

Encoder View: (all the conv2d and dense folder groups contained data)

Encoder view
Encoder view
Annotazione 2020-07-24 encoder.jpg (37.8 KiB) Viewed 20482 times

Decoder A View:

Decoder A view
Decoder A view
Annotazione 2020-07-24 decoder_a.jpg (39.16 KiB) Viewed 20482 times

Decoder B View:

Decoder B view
Decoder B view
Annotazione 2020-07-24 decoder_b.jpg (40 KiB) Viewed 20482 times

I tried to edit the h5 file, putting everything inside a model_weights folder group and gave me this error:

Code: Select all

07/24/2020 20:15:54 MainProcess     _training_0     multithreading  run                       DEBUG    Error in thread (_training_0): "Unable to open object (object 'input_2' doesn't exist)"
07/24/2020 20:15:55 MainProcess     MainThread      train           _monitor                  DEBUG    Thread error detected
07/24/2020 20:15:55 MainProcess     MainThread      train           _monitor                  DEBUG    Closed Monitor
07/24/2020 20:15:55 MainProcess     MainThread      train           _end_thread               DEBUG    Ending Training thread
07/24/2020 20:15:55 MainProcess     MainThread      train           _end_thread               CRITICAL Error caught! Exiting...
07/24/2020 20:15:55 MainProcess     MainThread      multithreading  join                      DEBUG    Joining Threads: '_training'
07/24/2020 20:15:55 MainProcess     MainThread      multithreading  join                      DEBUG    Joining Thread: '_training_0'
07/24/2020 20:15:55 MainProcess     MainThread      multithreading  join                      ERROR    Caught exception in thread: '_training_0'
Traceback (most recent call last):
  File "C:\Users\***\faceswap\plugins\train\model\_base.py", line 834, in load
    network = load_model(self.filename, custom_objects=get_custom_objects())
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\keras\engine\saving.py", line 419, in load_model
    model = _deserialize_model(f, custom_objects, compile)
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\keras\engine\saving.py", line 221, in _deserialize_model
    model_config = f['model_config']
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\keras\utils\io_utils.py", line 302, in __getitem__
    raise ValueError('Cannot create group in read only mode.')
ValueError: Cannot create group in read only mode.

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "C:\Users\***\faceswap\lib\cli\launcher.py", line 155, in execute_script
    process.process()
  File "C:\Users\***\faceswap\scripts\train.py", line 161, in process
    self._end_thread(thread, err)
  File "C:\Users\***\faceswap\scripts\train.py", line 201, in _end_thread
    thread.join()
  File "C:\Users\***\faceswap\lib\multithreading.py", line 121, in join
    raise thread.err[1].with_traceback(thread.err[2])
  File "C:\Users\***\faceswap\lib\multithreading.py", line 37, in run
    self._target(*self._args, **self._kwargs)
  File "C:\Users\***\faceswap\scripts\train.py", line 226, in _training
    raise err
  File "C:\Users\***\faceswap\scripts\train.py", line 214, in _training
    model = self._load_model()
  File "C:\Users\***\faceswap\scripts\train.py", line 255, in _load_model
    predict=False)
  File "C:\Users\***\faceswap\plugins\train\model\lightweight.py", line 20, in __init__
    super().__init__(*args, **kwargs)
  File "C:\Users\***\faceswap\plugins\train\model\original.py", line 25, in __init__
    super().__init__(*args, **kwargs)
  File "C:\Users\***\faceswap\plugins\train\model\_base.py", line 125, in __init__
    self.build()
  File "C:\Users\***\faceswap\plugins\train\model\_base.py", line 244, in build
    self.load_models(swapped=False)
  File "C:\Users\***\faceswap\plugins\train\model\_base.py", line 456, in load_models
    is_loaded = network.load(fullpath=model_mapping[network.side][network.type])
  File "C:\Users\***\faceswap\plugins\train\model\_base.py", line 837, in load
    self.convert_legacy_weights()
  File "C:\Users\***\faceswap\plugins\train\model\_base.py", line 863, in convert_legacy_weights
    self.network.load_weights(self.filename)
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\keras\engine\network.py", line 1166, in load_weights
    f, self.layers, reshape=reshape)
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\keras\engine\saving.py", line 1021, in load_weights_from_hdf5_group
    g = f[name]
  File "h5py\_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
  File "h5py\_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
  File "C:\Users\***\MiniConda3\envs\faceswap\lib\site-packages\h5py\_hl\group.py", line 264, in __getitem__
    oid = h5o.open(self.id, self._e(name), lapl=self._lapl)
  File "h5py\_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
  File "h5py\_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
  File "h5py\h5o.pyx", line 190, in h5py.h5o.open
KeyError: "Unable to open object (object 'input_2' doesn't exist)"

On converting I got more or less the same error.

User avatar
torzdf
Posts: 2649
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 129 times
Been thanked: 622 times

Re: OpenFaceSwap model

Post by torzdf »

Ok, assuming you're using the Faceswap GUI (and you should it's the easiest way to use faceswap), then these models should auto-update when you put them into Faceswap either to train or convert....

It's been a long time since we changed our structure from the format you have, but they should update. The old LowMem model is now part of the Original model, so select the "Original" model in faceswap.

Your model files should be named:

Code: Select all

lowmem_encoder.h5
lowmem_decoder_A.h5
lowmem_decoder_B.h5

NB: Lightweight model and LowMem are not the same thing.

Also do not manually edit the weights files! If you have already edited the files and don't have a backup, you may have ruined any chance you had for updating your model.

My word is final

User avatar
sykym
Posts: 6
Joined: Thu Jul 23, 2020 9:51 pm
Location: Italy

Re: OpenFaceSwap model

Post by sykym »

Firstly I want to tank you already for your help, I didn't know this forum was so active :D

NB: Lightweight model and LowMem are not the same thing.

Also do not manually edit the weights files! If you have already edited the files and don't have a backup, you may have ruined any chance you had for updating your model.

Thanks for clarifying, and I yes, I there's no way I wouldn't make a back up of a model I trained for days. 😅

But, it gives me ValueError: Cannot create group in read only mode. and ValueError: Tensor("Placeholder:0", shape=(3, 3, 512, 1024), dtype=float32) must be from the same graph as Tensor("upscale_8_0_conv2d/kernel:0", shape=(3, 3, 512, 1024), dtype=float32_ref).
I've checked lowmem on Original in the train settings, but the model files keep getting called original_* .
On the train.ini file, lowmem is set to lowmem = True .
This is the current original_state.json file:

Code: Select all

{
  "name": "original",
  "sessions": {
    "1": {
      "timestamp": 1595697018.5932138,
      "no_logs": false,
      "pingpong": false,
      "loss_names": {},
      "batchsize": 0,
      "iterations": 0,
      "config": {
        "learning_rate": 5e-05
      }
    }
  },
  "lowest_avg_loss": {},
  "iterations": 0,
  "inputs": {
    "face:0": [
      64,
      64,
      3
    ]
  },
  "training_size": 256,
  "config": {
    "coverage": 62.5,
    "reflect_padding": false,
    "mask_type": null,
    "mask_blur_kernel": 3,
    "mask_threshold": 4,
    "learn_mask": false,
    "lowmem": false,
    "icnr_init": false,
    "conv_aware_init": false,
    "penalized_mask_loss": true,
    "loss_function": "mae",
    "learning_rate": 5e-05
  }
}

EDIT:
Training settings:

settings
settings
Annotazione 2020-07-25 settjpg.jpg (32.55 KiB) Viewed 20424 times

Original lowmem check:

low mem check
low mem check
Annotazione 2020-07-25 lowmem.jpg (32.27 KiB) Viewed 20424 times
Last edited by sykym on Sat Jul 25, 2020 6:41 pm, edited 1 time in total.
User avatar
torzdf
Posts: 2649
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 129 times
Been thanked: 622 times

Re: OpenFaceSwap model

Post by torzdf »

Can you zip up your model files (the ones generated from OpenFaceSwap) and post them please. A Personal Message with the link is fine.

My word is final

User avatar
torzdf
Posts: 2649
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 129 times
Been thanked: 622 times

Re: OpenFaceSwap model

Post by torzdf »

Ok, I've looked into this.

Legacy support appears to no longer work. This isn't a huge issue from perspective... The fact that this has only just been discovered means that people generally don't use it anyway.

I rolled back to a known working version, and whilst I was able to convert the decoders, there appears to be an issue with your encoder. Ultimately, it is the wrong size (it should be about double the size of the file you provided), so when it tries to convert, it throws up errors about layer count mismatch.

At this point, all you can really do is make sure you provided me the right files. If you did, then I can only assume that OpenFaceSwap made some changes to this model which is not compatible with Faceswap.

Either way, there have been many developments in Faceswap since the version you are using, so you're probably best off starting again (not the answer you wanted to hear, I'm sure, but probably the best approach).

My word is final

User avatar
sykym
Posts: 6
Joined: Thu Jul 23, 2020 9:51 pm
Location: Italy

Re: OpenFaceSwap model

Post by sykym »

The reason for me to move from OFS to Faceswap is that on conversion with OFS it gives me those warnings :

Code: Select all

Failed to convert image: C:\Users\***\Desktop\Nuova cartella (2)\faceswap\m2\p2,1\images\imageA*.jpg. 
Reason: __init__() got an unexpected keyword argument 'r'

and it returns only the images that did not detect a face.

Does it matter that, in Faceswap, the lowmem value in the original_state.json is set to false, even though I turned it on?
Also, could I get this old working version to try myself?

User avatar
torzdf
Posts: 2649
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 129 times
Been thanked: 622 times

Re: OpenFaceSwap model

Post by torzdf »

You can check out the pre-snapshot branch of Faceswap, which should be compatible with your model (if it hasn't been changed by openfaceswap) by entering your faceswap folder and entering

Code: Select all

git checkout pre-refactor-snapshot

This is very old though, and no way near as user friendly, so you may be best off creating a new environment.

You can get back to the master branch by entering:

Code: Select all

git checkout master

The first compatible commit that auto-updates legacy files is:

Code: Select all

git checkout 9114262234e1bf5ca52ed200e6a757f9e6456143

Same caveats as above apply.

My word is final

User avatar
sykym
Posts: 6
Joined: Thu Jul 23, 2020 9:51 pm
Location: Italy

Re: OpenFaceSwap model

Post by sykym »

Sorry for bothering, can't open faceswap on pre-refactor-snapshot. It give me :
ModuleNotFoundError: No module named 'dlib'
And when trying to get git checkout 9114262234e1bf5ca52ed200e6a757f9e6456143
it gives me fatal: reference is not a tree: 9114262234e1bf5ca52ed200e6a757f9e6456143.

And sorry for asking too much, do you know if is it possible to get a version of OFS that you are sure is compatible? Thanks and sorry again🙇

User avatar
bryanlyon
Site Admin
Posts: 793
Joined: Fri Jul 12, 2019 12:49 am
Answers: 44
Location: San Francisco
Has thanked: 4 times
Been thanked: 218 times
Contact:

Re: OpenFaceSwap model

Post by bryanlyon »

Pre-refactor is quite old and before we had a uniform installation. You're not going to be able to simply "pip" install dlib. It requires you to compile it. When you checked out the pre-refactor it included an install.md file. You'll need to open that (preferrably in a tool that reads markup) and follow the instructions there to get it running.

User avatar
torzdf
Posts: 2649
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 129 times
Been thanked: 622 times

Re: OpenFaceSwap model

Post by torzdf »

You can get dlib with

Code: Select all

conda install -c conda-forge dlib

However, from this point, I'm afraid you're on you're own. It's just not a great use of time providing support for a seriously outdated version of our software.

As for OpenFaceSwap. I don't know, I never used it.

My word is final

User avatar
torzdf
Posts: 2649
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 129 times
Been thanked: 622 times

Re: OpenFaceSwap model

Post by torzdf »

My word is final

Locked