cd00859c40
* model_refactor (#571) * original model to new structure * IAE model to new structure * OriginalHiRes to new structure * Fix trainer for different resolutions * Initial config implementation * Configparse library added * improved training data loader * dfaker model working * Add logging to training functions * Non blocking input for cli training * Add error handling to threads. Add non-mp queues to queue_handler * Improved Model Building and NNMeta * refactor lib/models * training refactor. DFL H128 model Implementation * Dfaker - use hashes * Move timelapse. Remove perceptual loss arg * Update INSTALL.md. Add logger formatting. Update Dfaker training * DFL h128 partially ported * Add mask to dfaker (#573) * Remove old models. Add mask to dfaker * dfl mask. Make masks selectable in config (#575) * DFL H128 Mask. Mask type selectable in config. * remove gan_v2_2 * Creating Input Size config for models Creating Input Size config for models Will be used downstream in converters. Also name change of image_shape to input_shape to clarify ( for future models with potentially different output_shapes) * Add mask loss options to config * MTCNN options to config.ini. Remove GAN config. Update USAGE.md * Add sliders for numerical values in GUI * Add config plugins menu to gui. Validate config * Only backup model if loss has dropped. Get training working again * bugfixes * Standardise loss printing * GUI idle cpu fixes. Graph loss fix. * mutli-gpu logging bugfix * Merge branch 'staging' into train_refactor * backup state file * Crash protection: Only backup if both total losses have dropped * Port OriginalHiRes_RC4 to train_refactor (OriginalHiRes) * Load and save model structure with weights * Slight code update * Improve config loader. Add subpixel opt to all models. Config to state * Show samples... wrong input * Remove AE topology. Add input/output shapes to State * Port original_villain (birb/VillainGuy) model to faceswap * Add plugin info to GUI config pages * Load input shape from state. IAE Config options. * Fix transform_kwargs. Coverage to ratio. Bugfix mask detection * Suppress keras userwarnings. Automate zoom. Coverage_ratio to model def. * Consolidation of converters & refactor (#574) * Consolidation of converters & refactor Initial Upload of alpha Items - consolidate convert_mased & convert_adjust into one converter -add average color adjust to convert_masked -allow mask transition blur size to be a fixed integer of pixels and a fraction of the facial mask size -allow erosion/dilation size to be a fixed integer of pixels and a fraction of the facial mask size -eliminate redundant type conversions to avoid multiple round-off errors -refactor loops for vectorization/speed -reorganize for clarity & style changes TODO - bug/issues with warping the new face onto a transparent old image...use a cleanup mask for now - issues with mask border giving black ring at zero erosion .. investigate - remove GAN ?? - test enlargment factors of umeyama standard face .. match to coverage factor - make enlargment factor a model parameter - remove convert_adjusted and referencing code when finished * Update Convert_Masked.py default blur size of 2 to match original... description of enlargement tests breakout matrxi scaling into def * Enlargment scale as a cli parameter * Update cli.py * dynamic interpolation algorithm Compute x & y scale factors from the affine matrix on the fly by QR decomp. Choose interpolation alogrithm for the affine warp based on an upsample or downsample for each image * input size input size from config * fix issues with <1.0 erosion * Update convert.py * Update Convert_Adjust.py more work on the way to merginf * Clean up help note on sharpen * cleanup seamless * Delete Convert_Adjust.py * Update umeyama.py * Update training_data.py * swapping * segmentation stub * changes to convert.str * Update masked.py * Backwards compatibility fix for models Get converter running * Convert: Move masks to class. bugfix blur_size some linting * mask fix * convert fixes - missing facehull_rect re-added - coverage to % - corrected coverage logic - cleanup of gui option ordering * Update cli.py * default for blur * Update masked.py * added preliminary low_mem version of OriginalHighRes model plugin * Code cleanup, minor fixes * Update masked.py * Update masked.py * Add dfl mask to convert * histogram fix & seamless location * update * revert * bugfix: Load actual configuration in gui * Standardize nn_blocks * Update cli.py * Minor code amends * Fix Original HiRes model * Add masks to preview output for mask trainers refactor trainer.__base.py * Masked trainers converter support * convert bugfix * Bugfix: Converter for masked (dfl/dfaker) trainers * Additional Losses (#592) * initial upload * Delete blur.py * default initializer = He instead of Glorot (#588) * Allow kernel_initializer to be overridable * Add ICNR Initializer option for upscale on all models. * Hopefully fixes RSoDs with original-highres model plugin * remove debug line * Original-HighRes model plugin Red Screen of Death fix, take #2 * Move global options to _base. Rename Villain model * clipnorm and res block biases * scale the end of res block * res block * dfaker pre-activation res * OHRES pre-activation * villain pre-activation * tabs/space in nn_blocks * fix for histogram with mask all set to zero * fix to prevent two networks with same name * GUI: Wider tooltips. Improve TQDM capture * Fix regex bug * Convert padding=48 to ratio of image size * Add size option to alignments tool extract * Pass through training image size to convert from model * Convert: Pull training coverage from model * convert: coverage, blur and erode to percent * simplify matrix scaling * ordering of sliders in train * Add matrix scaling to utils. Use interpolation in lib.aligner transform * masked.py Import get_matrix_scaling from utils * fix circular import * Update masked.py * quick fix for matrix scaling * testing thus for now * tqdm regex capture bugfix * Minor ammends * blur size cleanup * Remove coverage option from convert (Now cascades from model) * Implement convert for all model types * Add mask option and coverage option to all existing models * bugfix for model loading on convert * debug print removal * Bugfix for masks in dfl_h128 and iae * Update preview display. Add preview scaling to cli * mask notes * Delete training_data_v2.py errant file * training data variables * Fix timelapse function * Add new config items to state file for legacy purposes * Slight GUI tweak * Raise exception if problem with loaded model * Add Tensorboard support (Logs stored in model directory) * ICNR fix * loss bugfix * convert bugfix * Move ini files to config folder. Make TensorBoard optional * Fix training data for unbalanced inputs/outputs * Fix config "none" test * Keep helptext in .ini files when saving config from GUI * Remove frame_dims from alignments * Add no-flip and warp-to-landmarks cli options * Revert OHR to RC4_fix version * Fix lowmem mode on OHR model * padding to variable * Save models in parallel threads * Speed-up of res_block stability * Automated Reflection Padding * Reflect Padding as a training option Includes auto-calculation of proper padding shapes, input_shapes, output_shapes Flag included in config now * rest of reflect padding * Move TB logging to cli. Session info to state file * Add session iterations to state file * Add recent files to menu. GUI code tidy up * [GUI] Fix recent file list update issue * Add correct loss names to TensorBoard logs * Update live graph to use TensorBoard and remove animation * Fix analysis tab. GUI optimizations * Analysis Graph popup to Tensorboard Logs * [GUI] Bug fix for graphing for models with hypens in name * [GUI] Correctly split loss to tabs during training * [GUI] Add loss type selection to analysis graph * Fix store command name in recent files. Switch to correct tab on open * [GUI] Disable training graph when 'no-logs' is selected * Fix graphing race condition * rename original_hires model to unbalanced
7.1 KiB
Executable File
Before attempting any of this, please make sure you have read, understood and completed the installation instructions. If you are experiencing issues, please raise them in the faceswap-playground repository instead of the main repo.
Workflow
So, you want to swap faces in pictures and videos? Well hold up, because first you gotta understand what this collection of scripts will do, how it does it and what it can't currently do.
The basic operation of this script is simple. It trains a machine learning model to recognize and transform two faces based on pictures. The machine learning model is our little "bot" that we're teaching to do the actual swapping and the pictures are the "training data" that we use to train it. Note that the bot is primarily processing faces. Other objects might not work.
So here's our plan. We want to create a reality where Donald Trump lost the presidency to Nic Cage; we have his inauguration video; let's replace Trump with Cage.
Gathering raw data
In order to accomplish this, the bot needs to learn to recognize both face A (Trump) and face B (Nic Cage). By default, the bot doesn't know what a Trump or a Nic Cage looks like. So we need to show it some pictures and let it guess which is which. So we need pictures of both of these faces first.
A possible source is Google, DuckDuckGo or Bing image search. There are scripts to download large amounts of images. Alternatively, if you have a video of the person you're looking for (from interviews, public speeches, or movies), you can convert this video to still images and use those. see Extracting video frames for more information.
Feel free to list your image sets in the faceswap-playground, or add more methods to this file.
So now we have a folder full of pictures of Trump and a separate folder of Nic Cage. Let's save them in our directory where we put the faceswap project. Example: ~/faceswap/photo/trump and ~/faceswap/photo/cage
EXTRACT
So here's a problem. We have a ton of pictures of both our subjects, but they're just pictures of them doing stuff or in an environment with other people. Their bodies are on there, they're on there with other people... It's a mess. We can only train our bot if the data we have is consistent and focusses on the subject we want to swap. This is where faceswap first comes in.
# To convert trump:
python faceswap.py extract -i ~/faceswap/photo/trump -o ~/faceswap/data/trump
# To convert cage:
python faceswap.py extract -i ~/faceswap/photo/cage -o ~/faceswap/data/cage
We specify our photo input directory and the output folder where our training data will be saved. The script will then try its best to recognize face landmarks, crop the image to that size, and save it to the output folder. Note: this script will make grabbing test data much easier, but it is not perfect. It will (incorrectly) detect multiple faces in some photos and does not recognize if the face is the person who we want to swap. Therefore: Always check your training data before you start training. The training data will influence how good your model will be at swapping.
You can see the full list of arguments for extracting via help flag. i.e.
python faceswap.py extract -h
Some of the plugins have configurable options. You can find the config options in: <faceswap_folder>\plugins\extract\config.ini. Extract needs to have been run at least once to generate the config file
TRAIN
The training process will take the longest, especially on CPU. We specify the folders where the two faces are, and where we will save our training model. It will start hammering the training data once you run the command. I personally really like to go by the preview and quit the processing once I'm happy with the results.
python faceswap.py train -A ~/faceswap/data/trump -B ~/faceswap/data/cage -m ~/faceswap/models/
# or -p to show a preview
python faceswap.py train -A ~/faceswap/data/trump -B ~/faceswap/data/cage -m ~/faceswap/models/ -p
If you use the preview feature, select the preview window and press ENTER to save your processed data and quit gracefully. Without the preview enabled, you might have to forcefully quit by hitting Ctrl+C to cancel the command. Note that it will save the model once it's gone through about 100 iterations, which can take quite a while. So make sure you save before stopping the process.
You can see the full list of arguments for training via help flag. i.e.
python faceswap.py train -h
Some of the plugins have configurable options. You can find the config options in: <faceswap_folder>\plugins\traom\config.ini. Train needs to have been run at least once to generate the config file
CONVERT
Now that we're happy with our trained model, we can convert our video. How does it work? Similarly to the extraction script, actually! The conversion script basically detects a face in a picture using the same algorithm, quickly crops the image to the right size, runs our bot on this cropped image of the face it has found, and then (crudely) pastes the processed face back into the picture.
Remember those initial pictures we had of Trump? Let's try swapping a face there. We will use that directory as our input directory, create a new folder where the output will be saved, and tell them which model to use.
python faceswap.py convert -i ~/faceswap/photo/trump/ -o ~/faceswap/output/ -m ~/faceswap/models/
It should now start swapping faces of all these pictures.
You can see the full list of arguments available for converting via help flag. i.e.
python faceswap.py convert -h
GUI
All of the above commands and options can be run from the GUI. This is launched with:
python faceswap.py gui
Video's
A video is just a series of pictures in the form of frames. Therefore you can gather the raw images from them for your dataset or combine your results into a video.
EFFMPEG
You can perform various video processes with the built in effmpeg tool. You can see the full list of arguments available by running:
python tools.py effmpeg -h
Extracting video frames with FFMPEG
Alternatively you can split a video into separate frames using ffmpeg for instance. Below is an example command to process a video to separate frames.
ffmpeg -i /path/to/my/video.mp4 /path/to/output/video-frame-%d.png
Generating a video
If you split a video, using ffmpeg for example, and used them as a target for swapping faces onto you can combine these frames again. The command below stitches the png frames back into a single video again.
ffmpeg -i video-frame-%0d.png -c:v libx264 -vf "fps=25,format=yuv420p" out.mp4
Notes
This guide is far from complete. Functionality may change over time, and new dependencies are added and removed as time goes on.
If you are experiencing issues, please raise them in the faceswap-playground repository instead of the main repo.