github/DeepFaceLab
1
0
Fork 0
mirror of https://github.com/iperov/DeepFaceLab.git synced 2025-07-07 13:32:09 -07:00
Code Issues Projects Releases Packages Wiki Activity Actions

SAE/SAEHD:

Browse source 
removed option 'apply random ct'

added option
   Color transfer mode apply to src faceset. ( none/rct/lct/mkl/idt, ?:help skip: none )
   Change color distribution of src samples close to dst samples. Try all modes to find the best.
before was lct mode, but sometime it does not work properly for some facesets.
This commit is contained in:
Colombo 2019-10-15 11:22:19 +04:00
parent 5a2eefaa5b
commit 73511a072a
3 changed files with 29 additions and 22 deletions
Download patch file Download diff file Expand all files Collapse all files

17
models/Model_SAE/Model.py
View file

@ -82,9 +82,9 @@ class SAEModel(ModelBase):
self.options['bg_style_power'] = np.clip ( io.input_number("Background style power ( 0.0 .. 100.0 ?:help skip:%.2f) : " % (default_bg_style_power), default_bg_style_power, self.options['bg_style_power'] = np.clip ( io.input_number("Background style power ( 0.0 .. 100.0 ?:help skip:%.2f) : " % (default_bg_style_power), default_bg_style_power,
help_message="Learn to transfer image around face. This can make face more like dst. Enabling this option increases the chance of model collapse."), 0.0, 100.0 ) help_message="Learn to transfer image around face. This can make face more like dst. Enabling this option increases the chance of model collapse."), 0.0, 100.0 )
default_apply_random_ct = False if is_first_run else self.options.get('apply_random_ct', False) default_ct_mode = False if is_first_run else self.options.get('ct_mode', 'none')
self.options['apply_random_ct'] = io.input_bool (f"Apply random color transfer to src faceset? (y/n, ?:help skip:{yn_str[default_apply_random_ct]}) : ", default_apply_random_ct, help_message="Increase variativity of src samples by apply LCT color transfer from random dst samples. It is like 'face_style' learning, but more precise color transfer and without risk of model collapse, also it does not require additional GPU resources, but the training time may be longer, due to the src faceset is becoming more diverse.") self.options['ct_mode'] = io.input_str (f"Color transfer mode apply to src faceset. ( none/rct/lct/mkl/idt, ?:help skip:{default_ct_mode}) : ", default_ct_mode, ['none','rct','lct','mkl','idt'], help_message="Change color distribution of src samples close to dst samples. Try all modes to find the best.")
if nnlib.device.backend != 'plaidML': # todo https://github.com/plaidml/plaidml/issues/301 if nnlib.device.backend != 'plaidML': # todo https://github.com/plaidml/plaidml/issues/301
default_clipgrad = False if is_first_run else self.options.get('clipgrad', False) default_clipgrad = False if is_first_run else self.options.get('clipgrad', False)
self.options['clipgrad'] = io.input_bool (f"Enable gradient clipping? (y/n, ?:help skip:{yn_str[default_clipgrad]}) : ", default_clipgrad, help_message="Gradient clipping reduces chance of model collapse, sacrificing speed of training.") self.options['clipgrad'] = io.input_bool (f"Enable gradient clipping? (y/n, ?:help skip:{yn_str[default_clipgrad]}) : ", default_clipgrad, help_message="Gradient clipping reduces chance of model collapse, sacrificing speed of training.")
@ -95,7 +95,7 @@ class SAEModel(ModelBase):
self.options['pixel_loss'] = self.options.get('pixel_loss', False) self.options['pixel_loss'] = self.options.get('pixel_loss', False)
self.options['face_style_power'] = self.options.get('face_style_power', default_face_style_power) self.options['face_style_power'] = self.options.get('face_style_power', default_face_style_power)
self.options['bg_style_power'] = self.options.get('bg_style_power', default_bg_style_power) self.options['bg_style_power'] = self.options.get('bg_style_power', default_bg_style_power)
self.options['apply_random_ct'] = self.options.get('apply_random_ct', False) self.options['ct_mode'] = self.options.get('ct_mode', 'none')
self.options['clipgrad'] = self.options.get('clipgrad', False) self.options['clipgrad'] = self.options.get('clipgrad', False)
if is_first_run: if is_first_run:
@ -121,7 +121,6 @@ class SAEModel(ModelBase):
bgr_shape = (resolution, resolution, 3) bgr_shape = (resolution, resolution, 3)
mask_shape = (resolution, resolution, 1) mask_shape = (resolution, resolution, 1)
apply_random_ct = self.options.get('apply_random_ct', False)
masked_training = True masked_training = True
class SAEDFModel(object): class SAEDFModel(object):
@ -476,11 +475,11 @@ class SAEModel(ModelBase):
self.set_training_data_generators ([ self.set_training_data_generators ([
SampleGeneratorFace(training_data_src_path, sort_by_yaw_target_samples_path=training_data_dst_path if sort_by_yaw else None, SampleGeneratorFace(training_data_src_path, sort_by_yaw_target_samples_path=training_data_dst_path if sort_by_yaw else None,
random_ct_samples_path=training_data_dst_path if apply_random_ct else None, random_ct_samples_path=training_data_dst_path if self.options['ct_mode'] != 'none' else None,
debug=self.is_debug(), batch_size=self.batch_size, debug=self.is_debug(), batch_size=self.batch_size,
sample_process_options=SampleProcessor.Options(random_flip=self.random_flip, scale_range=np.array([-0.05, 0.05])+self.src_scale_mod / 100.0 ), sample_process_options=SampleProcessor.Options(random_flip=self.random_flip, scale_range=np.array([-0.05, 0.05])+self.src_scale_mod / 100.0 ),
output_sample_types = [ {'types' : (t.IMG_WARPED_TRANSFORMED, face_type, t_mode_bgr), 'resolution':resolution, 'apply_ct': apply_random_ct}, output_sample_types = [ {'types' : (t.IMG_WARPED_TRANSFORMED, face_type, t_mode_bgr), 'resolution':resolution, 'ct_mode': self.options['ct_mode'] },
{'types' : (t.IMG_TRANSFORMED, face_type, t_mode_bgr), 'resolution': resolution, 'apply_ct': apply_random_ct }, {'types' : (t.IMG_TRANSFORMED, face_type, t_mode_bgr), 'resolution': resolution, 'ct_mode': self.options['ct_mode'] },
{'types' : (t.IMG_TRANSFORMED, face_type, t.MODE_M), 'resolution': resolution } ] {'types' : (t.IMG_TRANSFORMED, face_type, t.MODE_M), 'resolution': resolution } ]
), ),
@ -565,7 +564,7 @@ class SAEModel(ModelBase):
import converters import converters
return self.predictor_func, (self.options['resolution'], self.options['resolution'], 3), converters.ConverterConfigMasked(face_type=face_type, return self.predictor_func, (self.options['resolution'], self.options['resolution'], 3), converters.ConverterConfigMasked(face_type=face_type,
default_mode = 1 if self.options['apply_random_ct'] or self.options['face_style_power'] or self.options['bg_style_power'] else 4, default_mode = 1 if self.options['ct_mode'] != 'none' or self.options['face_style_power'] or self.options['bg_style_power'] else 4,
clip_hborder_mask_per=0.0625 if (self.options['face_type'] == 'f') else 0, clip_hborder_mask_per=0.0625 if (self.options['face_type'] == 'f') else 0,
) )

15
models/Model_SAEHD/Model.py
View file

@ -76,8 +76,8 @@ class SAEHDModel(ModelBase):
self.options['bg_style_power'] = np.clip ( io.input_number("Background style power ( 0.0 .. 100.0 ?:help skip:%.2f) : " % (default_bg_style_power), default_bg_style_power, self.options['bg_style_power'] = np.clip ( io.input_number("Background style power ( 0.0 .. 100.0 ?:help skip:%.2f) : " % (default_bg_style_power), default_bg_style_power,
help_message="Learn to transfer image around face. This can make face more like dst. Enabling this option increases the chance of model collapse."), 0.0, 100.0 ) help_message="Learn to transfer image around face. This can make face more like dst. Enabling this option increases the chance of model collapse."), 0.0, 100.0 )
default_apply_random_ct = False if is_first_run else self.options.get('apply_random_ct', False) default_ct_mode = False if is_first_run else self.options.get('ct_mode', 'none')
self.options['apply_random_ct'] = io.input_bool (f"Apply random color transfer to src faceset? (y/n, ?:help skip:{yn_str[default_apply_random_ct]}) : ", default_apply_random_ct, help_message="Increase variativity of src samples by apply LCT color transfer from random dst samples. It is like 'face_style' learning, but more precise color transfer and without risk of model collapse, also it does not require additional GPU resources, but the training time may be longer, due to the src faceset is becoming more diverse.") self.options['ct_mode'] = io.input_str (f"Color transfer mode apply to src faceset. ( none/rct/lct/mkl/idt, ?:help skip:{default_ct_mode}) : ", default_ct_mode, ['none','rct','lct','mkl','idt'], help_message="Change color distribution of src samples close to dst samples. Try all modes to find the best.")
if nnlib.device.backend != 'plaidML': # todo https://github.com/plaidml/plaidml/issues/301 if nnlib.device.backend != 'plaidML': # todo https://github.com/plaidml/plaidml/issues/301
default_clipgrad = False if is_first_run else self.options.get('clipgrad', False) default_clipgrad = False if is_first_run else self.options.get('clipgrad', False)
@ -90,7 +90,7 @@ class SAEHDModel(ModelBase):
self.options['true_face_training'] = self.options.get('true_face_training', default_true_face_training) self.options['true_face_training'] = self.options.get('true_face_training', default_true_face_training)
self.options['face_style_power'] = self.options.get('face_style_power', default_face_style_power) self.options['face_style_power'] = self.options.get('face_style_power', default_face_style_power)
self.options['bg_style_power'] = self.options.get('bg_style_power', default_bg_style_power) self.options['bg_style_power'] = self.options.get('bg_style_power', default_bg_style_power)
self.options['apply_random_ct'] = self.options.get('apply_random_ct', False) self.options['ct_mode'] = self.options.get('ct_mode', 'none')
self.options['clipgrad'] = self.options.get('clipgrad', False) self.options['clipgrad'] = self.options.get('clipgrad', False)
if is_first_run: if is_first_run:
@ -115,7 +115,6 @@ class SAEHDModel(ModelBase):
bgr_shape = (resolution, resolution, 3) bgr_shape = (resolution, resolution, 3)
mask_shape = (resolution, resolution, 1) mask_shape = (resolution, resolution, 1)
apply_random_ct = self.options.get('apply_random_ct', False)
self.true_face_training = self.options.get('true_face_training', False) self.true_face_training = self.options.get('true_face_training', False)
masked_training = True masked_training = True
@ -537,11 +536,11 @@ class SAEHDModel(ModelBase):
self.set_training_data_generators ([ self.set_training_data_generators ([
SampleGeneratorFace(training_data_src_path, sort_by_yaw_target_samples_path=training_data_dst_path if sort_by_yaw else None, SampleGeneratorFace(training_data_src_path, sort_by_yaw_target_samples_path=training_data_dst_path if sort_by_yaw else None,
random_ct_samples_path=training_data_dst_path if apply_random_ct else None, random_ct_samples_path=training_data_dst_path if self.options['ct_mode'] != 'none' else None,
debug=self.is_debug(), batch_size=self.batch_size, debug=self.is_debug(), batch_size=self.batch_size,
sample_process_options=SampleProcessor.Options(random_flip=self.random_flip, scale_range=np.array([-0.05, 0.05])+self.src_scale_mod / 100.0 ), sample_process_options=SampleProcessor.Options(random_flip=self.random_flip, scale_range=np.array([-0.05, 0.05])+self.src_scale_mod / 100.0 ),
output_sample_types = [ {'types' : (t_img_warped, face_type, t_mode_bgr), 'resolution':resolution, 'apply_ct': apply_random_ct}, output_sample_types = [ {'types' : (t_img_warped, face_type, t_mode_bgr), 'resolution':resolution, 'ct_mode': self.options['ct_mode'] },
{'types' : (t.IMG_TRANSFORMED, face_type, t_mode_bgr), 'resolution': resolution, 'apply_ct': apply_random_ct }, {'types' : (t.IMG_TRANSFORMED, face_type, t_mode_bgr), 'resolution': resolution, 'ct_mode': self.options['ct_mode'] },
{'types' : (t.IMG_TRANSFORMED, face_type, t.MODE_M), 'resolution': resolution } ] {'types' : (t.IMG_TRANSFORMED, face_type, t.MODE_M), 'resolution': resolution } ]
), ),
@ -640,7 +639,7 @@ class SAEHDModel(ModelBase):
import converters import converters
return self.predictor_func, (self.options['resolution'], self.options['resolution'], 3), converters.ConverterConfigMasked(face_type=face_type, return self.predictor_func, (self.options['resolution'], self.options['resolution'], 3), converters.ConverterConfigMasked(face_type=face_type,
default_mode = 1 if self.options['apply_random_ct'] or self.options['face_style_power'] or self.options['bg_style_power'] else 4, default_mode = 1 if self.options['ct_mode'] != 'none' or self.options['face_style_power'] or self.options['bg_style_power'] else 4,
clip_hborder_mask_per=0.0625 if (face_type != FaceType.HALF) else 0, clip_hborder_mask_per=0.0625 if (face_type != FaceType.HALF) else 0,
) )

19
samplelib/SampleProcessor.py
View file

@ -37,7 +37,7 @@ opts:
'resolution' : N 'resolution' : N
'motion_blur' : (chance_int, range) - chance 0..100 to apply to face (not mask), and max_size of motion blur 'motion_blur' : (chance_int, range) - chance 0..100 to apply to face (not mask), and max_size of motion blur
'apply_ct' : bool 'ct_mode' :
'normalize_tanh' : bool 'normalize_tanh' : bool
""" """
@ -125,7 +125,7 @@ class SampleProcessor(object):
gaussian_blur = opts.get('gaussian_blur', None) gaussian_blur = opts.get('gaussian_blur', None)
random_hsv_shift = opts.get('random_hsv_shift', None) random_hsv_shift = opts.get('random_hsv_shift', None)
apply_ct = opts.get('apply_ct', False) ct_mode = opts.get('ct_mode', 'None')
normalize_tanh = opts.get('normalize_tanh', False) normalize_tanh = opts.get('normalize_tanh', False)
img_type = SPTF.NONE img_type = SPTF.NONE
@ -244,14 +244,23 @@ class SampleProcessor(object):
img_bgr = img[...,0:3] img_bgr = img[...,0:3]
img_mask = img[...,3:4] img_mask = img[...,3:4]
if apply_ct and ct_sample is not None: if ct_mode is not None and ct_sample is not None:
if ct_sample_bgr is None: if ct_sample_bgr is None:
ct_sample_bgr = ct_sample.load_bgr() ct_sample_bgr = ct_sample.load_bgr()
ct_sample_bgr_resized = cv2.resize( ct_sample_bgr, (resolution,resolution), cv2.INTER_LINEAR ) ct_sample_bgr_resized = cv2.resize( ct_sample_bgr, (resolution,resolution), cv2.INTER_LINEAR )
img_bgr = imagelib.linear_color_transfer (img_bgr, ct_sample_bgr_resized) if ct_mode == 'lct':
img_bgr = np.clip( img_bgr, 0.0, 1.0) img_bgr = imagelib.linear_color_transfer (img_bgr, ct_sample_bgr_resized)
img_bgr = np.clip( img_bgr, 0.0, 1.0)
elif ct_mode == 'rct':
img_bgr = imagelib.reinhard_color_transfer ( np.clip( (img_bgr*255).astype(np.uint8), 0, 255),
np.clip( (ct_sample_bgr_resized*255).astype(np.uint8), 0, 255) )
img_bgr = np.clip( img_bgr.astype(np.float32) / 255.0, 0.0, 1.0)
elif ct_mode == 'mkl':
img_bgr = imagelib.color_transfer_mkl (img_bgr, ct_sample_bgr_resized)
elif ct_mode == 'idt':
img_bgr = imagelib.color_transfer_idt (img_bgr, ct_sample_bgr_resized)
if random_hsv_shift: if random_hsv_shift:
rnd_state = np.random.RandomState (sample_rnd_seed) rnd_state = np.random.RandomState (sample_rnd_seed)