1
0
Fork 0
mirror of https://github.com/deepfakes/faceswap synced 2025-06-07 10:43:27 -04:00
faceswap/lib/alignments.py
torzdf cd00859c40
model_refactor (#571) (#572)
* 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
2019-02-09 18:35:12 +00:00

361 lines
15 KiB
Python

#!/usr/bin/env python3
""" Alignments file functions for reading, writing and manipulating
a serialized alignments file """
import logging
import os
from datetime import datetime
import cv2
from lib import Serializer
from lib.utils import rotate_landmarks
logger = logging.getLogger(__name__) # pylint: disable=invalid-name
class Alignments():
""" Holds processes pertaining to the alignments file.
folder: folder alignments file is stored in
filename: Filename of alignments file excluding extension. If a
valid extension is provided, then it will be used to
decide the serializer, and the serializer argument will
be ignored.
serializer: If provided, this will be the format that the data is
saved in (if data is to be saved). Can be 'json', 'pickle'
or 'yaml'
"""
# pylint: disable=too-many-public-methods
def __init__(self, folder, filename="alignments", serializer="json"):
logger.debug("Initializing %s: (folder: '%s', filename: '%s', serializer: '%s')",
self.__class__.__name__, folder, filename, serializer)
self.serializer = self.get_serializer(filename, serializer)
self.file = self.get_location(folder, filename)
self.data = self.load()
logger.debug("Initialized %s", self.__class__.__name__)
# << PROPERTIES >> #
@property
def frames_count(self):
""" Return current frames count """
retval = len(self.data)
logger.trace(retval)
return retval
@property
def faces_count(self):
""" Return current faces count """
retval = sum(len(faces) for faces in self.data.values())
logger.trace(retval)
return retval
@property
def have_alignments_file(self):
""" Return whether an alignments file exists """
retval = os.path.exists(self.file)
logger.trace(retval)
return retval
@property
def hashes_to_frame(self):
""" Return a dict of each face_hash with their parent
frame name(s) and their index in the frame
"""
hash_faces = dict()
for frame_name, faces in self.data.items():
for idx, face in enumerate(faces):
hash_faces.setdefault(face["hash"], dict())[frame_name] = idx
return hash_faces
# << INIT FUNCTIONS >> #
@staticmethod
def get_serializer(filename, serializer):
""" Set the serializer to be used for loading and
saving alignments
If a filename with a valid extension is passed in
this will be used as the serializer, otherwise the
specified serializer will be used """
logger.debug("Getting serializer: (filename: '%s', serializer: '%s')",
filename, serializer)
extension = os.path.splitext(filename)[1]
if extension in (".json", ".p", ".yaml", ".yml"):
logger.debug("Serializer set from file extension: '%s'", extension)
retval = Serializer.get_serializer_from_ext(extension)
elif serializer not in ("json", "pickle", "yaml"):
raise ValueError("Error: {} is not a valid serializer. Use "
"'json', 'pickle' or 'yaml'")
else:
logger.debug("Serializer set from argument: '%s'", serializer)
retval = Serializer.get_serializer(serializer)
logger.verbose("Using '%s' serializer for alignments", retval.ext)
return retval
def get_location(self, folder, filename):
""" Return the path to alignments file """
logger.debug("Getting location: (folder: '%s', filename: '%s')", folder, filename)
extension = os.path.splitext(filename)[1]
if extension in (".json", ".p", ".yaml", ".yml"):
logger.debug("File extension set from filename: '%s'", extension)
location = os.path.join(str(folder), filename)
else:
location = os.path.join(str(folder),
"{}.{}".format(filename,
self.serializer.ext))
logger.debug("File extension set from serializer: '%s'", self.serializer.ext)
logger.verbose("Alignments filepath: '%s'", location)
return location
# << I/O >> #
def load(self):
""" Load the alignments data
Override for custom loading logic """
logger.debug("Loading alignments")
if not self.have_alignments_file:
raise ValueError("Error: Alignments file not found at "
"{}".format(self.file))
try:
logger.info("Reading alignments from: '%s'", self.file)
with open(self.file, self.serializer.roptions) as align:
data = self.serializer.unmarshal(align.read())
except IOError as err:
logger.error("'%s' not read: %s", self.file, err.strerror)
exit(1)
logger.debug("Loaded alignments")
return data
def reload(self):
""" Read the alignments data from the correct format """
logger.debug("Re-loading alignments")
self.data = self.load()
logger.debug("Re-loaded alignments")
def save(self):
""" Write the serialized alignments file """
logger.debug("Saving alignments")
try:
logger.info("Writing alignments to: '%s'", self.file)
with open(self.file, self.serializer.woptions) as align:
align.write(self.serializer.marshal(self.data))
logger.debug("Saved alignments")
except IOError as err:
logger.error("'%s' not written: %s", self.file, err.strerror)
def backup(self):
""" Backup copy of old alignments """
logger.debug("Backing up alignments")
if not os.path.isfile(self.file):
logger.debug("No alignments to back up")
return
now = datetime.now().strftime("%Y%m%d_%H%M%S")
src = self.file
split = os.path.splitext(src)
dst = split[0] + "_" + now + split[1]
logger.info("Backing up original alignments to '%s'", dst)
os.rename(src, dst)
logger.debug("Backed up alignments")
# << VALIDATION >> #
def frame_exists(self, frame):
""" return path of images that have faces """
retval = frame in self.data.keys()
logger.trace("'%s': %s", frame, retval)
return retval
def frame_has_faces(self, frame):
""" Return true if frame exists and has faces """
retval = bool(self.data.get(frame, list()))
logger.trace("'%s': %s", frame, retval)
return retval
def frame_has_multiple_faces(self, frame):
""" Return true if frame exists and has faces """
if not frame:
retval = False
else:
retval = bool(len(self.data.get(frame, list())) > 1)
logger.trace("'%s': %s", frame, retval)
return retval
# << DATA >> #
def get_faces_in_frame(self, frame):
""" Return the alignments for the selected frame """
logger.trace("Getting faces for frame: '%s'", frame)
return self.data.get(frame, list())
def get_full_frame_name(self, frame):
""" Return a frame with extension for when the extension is
not known """
retval = next(key for key in self.data.keys()
if key.startswith(frame))
logger.trace("Requested: '%s', Returning: '%s'", frame, retval)
return retval
def count_faces_in_frame(self, frame):
""" Return number of alignments within frame """
retval = len(self.data.get(frame, list()))
logger.trace(retval)
return retval
# << MANIPULATION >> #
def delete_face_at_index(self, frame, idx):
""" Delete the face alignment for given frame at given index """
logger.debug("Deleting face %s for frame '%s'", idx, frame)
idx = int(idx)
if idx + 1 > self.count_faces_in_frame(frame):
logger.debug("No face to delete: (frame: '%s', idx %s)", frame, idx)
return False
del self.data[frame][idx]
logger.debug("Deleted face: (frame: '%s', idx %s)", frame, idx)
return True
def add_face(self, frame, alignment):
""" Add a new face for a frame and return it's index """
logger.debug("Adding face to frame: '%s'", frame)
self.data[frame].append(alignment)
retval = self.count_faces_in_frame(frame) - 1
logger.debug("Returning new face index: %s", retval)
return retval
def update_face(self, frame, idx, alignment):
""" Replace a face for given frame and index """
logger.debug("Updating face %s for frame '%s'", idx, frame)
self.data[frame][idx] = alignment
def filter_hashes(self, hashlist, filter_out=False):
""" Filter in or out faces that match the hashlist
filter_out=True: Remove faces that match in the hashlist
filter_out=False: Remove faces that are not in the hashlist
"""
hashset = set(hashlist)
for filename, frame in self.data.items():
for idx, face in reversed(list(enumerate(frame))):
if ((filter_out and face.get("hash", None) in hashset) or
(not filter_out and face.get("hash", None) not in hashset)):
logger.verbose("Filtering out face: (filename: %s, index: %s)", filename, idx)
del frame[idx]
else:
logger.trace("Not filtering out face: (filename: %s, index: %s)",
filename, idx)
# << GENERATORS >> #
def yield_faces(self):
""" Yield face alignments for one image """
for frame_fullname, alignments in self.data.items():
frame_name = os.path.splitext(frame_fullname)[0]
face_count = len(alignments)
logger.trace("Yielding: (frame: '%s', faces: %s, frame_fullname: '%s')",
frame_name, face_count, frame_fullname)
yield frame_name, alignments, face_count, frame_fullname
@staticmethod
def yield_original_index_reverse(image_alignments, number_alignments):
""" Return the correct original index for
alignment in reverse order """
for idx, _ in enumerate(reversed(image_alignments)):
original_idx = number_alignments - 1 - idx
logger.trace("Yielding: face index %s", original_idx)
yield original_idx
# << LEGACY FUNCTIONS >> #
# < Rotation > #
# The old rotation method would rotate the image to find a face, then
# store the rotated landmarks along with a rotation value to tell the
# convert process that it had to rotate the frame to find the landmarks.
# This is problematic for numerous reasons. The process now rotates the
# landmarks to correctly correspond with the original frame. The below are
# functions to convert legacy alignments to the currently supported
# infrastructure.
# This can eventually be removed
def get_legacy_rotation(self):
""" Return a list of frames with legacy rotations
Looks for an 'r' value in the alignments file that
is not zero """
logger.debug("Getting alignments containing legacy rotations")
keys = list()
for key, val in self.data.items():
if any(alignment.get("r", None) for alignment in val):
keys.append(key)
logger.debug("Got alignments containing legacy rotations: %s", len(keys))
return keys
def rotate_existing_landmarks(self, frame_name, frame):
""" Backwards compatability fix. Rotates the landmarks to
their correct position and deletes r
NB: The original frame must be passed in otherwise
the transformation cannot be performed """
logger.trace("Rotating existing landmarks for frame: '%s'", frame_name)
dims = frame.shape[:2]
for face in self.get_faces_in_frame(frame_name):
angle = face.get("r", 0)
if not angle:
logger.trace("Landmarks do not require rotation: '%s'", frame_name)
return
logger.trace("Rotating landmarks: (frame: '%s', angle: %s)", frame_name, angle)
r_mat = self.get_original_rotation_matrix(dims, angle)
rotate_landmarks(face, r_mat)
del face["r"]
logger.trace("Rotatated existing landmarks for frame: '%s'", frame_name)
@staticmethod
def get_original_rotation_matrix(dimensions, angle):
""" Calculate original rotation matrix and invert """
logger.trace("Getting original rotation matrix: (dimensions: %s, angle: %s)",
dimensions, angle)
height, width = dimensions
center = (width/2, height/2)
r_mat = cv2.getRotationMatrix2D( # pylint: disable=no-member
center, -1.0 * angle, 1.)
abs_cos = abs(r_mat[0, 0])
abs_sin = abs(r_mat[0, 1])
rotated_width = int(height*abs_sin + width*abs_cos)
rotated_height = int(height*abs_cos + width*abs_sin)
r_mat[0, 2] += rotated_width/2 - center[0]
r_mat[1, 2] += rotated_height/2 - center[1]
logger.trace("Returning rotation matrix: %s", r_mat)
return r_mat
# <Face Hashes> #
# The old index based method of face matching is problematic.
# The SHA1 Hash of the extracted face is now stored in the alignments file.
# This has it's own issues, but they are far reduced from the index/filename method
# This can eventually be removed
def get_legacy_no_hashes(self):
""" Get alignments without face hashes """
logger.debug("Getting alignments without face hashes")
keys = list()
for key, val in self.data.items():
for alignment in val:
if "hash" not in alignment.keys():
keys.append(key)
break
logger.debug("Got alignments without face hashes: %s", len(keys))
return keys
def add_face_hashes(self, frame_name, hashes):
""" Backward compatability fix. Add face hash to alignments """
logger.trace("Adding face hash: (frame: '%s', hashes: %s)", frame_name, hashes)
faces = self.get_faces_in_frame(frame_name)
count_match = len(faces) - len(hashes)
if count_match != 0:
msg = "more" if count_match > 0 else "fewer"
logger.warning("There are %s %s face(s) in the alignments file than exist in the "
"faces folder. Check your sources for frame '%s'.",
abs(count_match), msg, frame_name)
for idx, i_hash in hashes.items():
faces[idx]["hash"] = i_hash