mirror of
https://github.com/iperov/DeepFaceLab.git
synced 2025-07-06 21:12:07 -07:00
4d37fd62cd
SAE: added "rare sample booster" SAE: pixel loss replaced to smooth transition from DSSIM to PixelLoss in 15k epochs by default
203 lines
No EOL
8.9 KiB
Python
203 lines
No EOL
8.9 KiB
Python
import numpy as np
|
|
|
|
from nnlib import nnlib
|
|
from models import ModelBase
|
|
from facelib import FaceType
|
|
from samples import *
|
|
from utils.console_utils import *
|
|
|
|
class Model(ModelBase):
|
|
|
|
encoderH5 = 'encoder.h5'
|
|
decoder_srcH5 = 'decoder_src.h5'
|
|
decoder_dstH5 = 'decoder_dst.h5'
|
|
|
|
#override
|
|
def onInitializeOptions(self, is_first_run, ask_override):
|
|
if is_first_run:
|
|
self.options['lighter_ae'] = input_bool ("Use lightweight autoencoder? (y/n, ?:help skip:n) : ", False, help_message="Lightweight autoencoder is faster, requires less VRAM, sacrificing overall quality. If your GPU VRAM <= 4, you should to choose this option.")
|
|
else:
|
|
default_lighter_ae = self.options.get('created_vram_gb', 99) <= 4 #temporally support old models, deprecate in future
|
|
if 'created_vram_gb' in self.options.keys():
|
|
self.options.pop ('created_vram_gb')
|
|
self.options['lighter_ae'] = self.options.get('lighter_ae', default_lighter_ae)
|
|
|
|
#override
|
|
def onInitialize(self, **in_options):
|
|
exec(nnlib.import_all(), locals(), globals())
|
|
self.set_vram_batch_requirements( {1.5:2,2:2,3:8,4:16,5:24,6:32,7:40,8:48} )
|
|
|
|
|
|
bgr_shape, mask_shape, self.encoder, self.decoder_src, self.decoder_dst = self.Build(self.options['lighter_ae'])
|
|
|
|
if not self.is_first_run():
|
|
self.encoder.load_weights (self.get_strpath_storage_for_file(self.encoderH5))
|
|
self.decoder_src.load_weights (self.get_strpath_storage_for_file(self.decoder_srcH5))
|
|
self.decoder_dst.load_weights (self.get_strpath_storage_for_file(self.decoder_dstH5))
|
|
|
|
input_src_bgr = Input(bgr_shape)
|
|
input_src_mask = Input(mask_shape)
|
|
input_dst_bgr = Input(bgr_shape)
|
|
input_dst_mask = Input(mask_shape)
|
|
|
|
rec_src_bgr, rec_src_mask = self.decoder_src( self.encoder(input_src_bgr) )
|
|
rec_dst_bgr, rec_dst_mask = self.decoder_dst( self.encoder(input_dst_bgr) )
|
|
|
|
self.ae = Model([input_src_bgr,input_src_mask,input_dst_bgr,input_dst_mask], [rec_src_bgr, rec_src_mask, rec_dst_bgr, rec_dst_mask] )
|
|
|
|
self.ae.compile(optimizer=Adam(lr=5e-5, beta_1=0.5, beta_2=0.999),
|
|
loss=[ DSSIMMaskLoss([input_src_mask]), 'mae', DSSIMMaskLoss([input_dst_mask]), 'mae' ] )
|
|
|
|
self.src_view = K.function([input_src_bgr],[rec_src_bgr, rec_src_mask])
|
|
self.dst_view = K.function([input_dst_bgr],[rec_dst_bgr, rec_dst_mask])
|
|
|
|
if self.is_training_mode:
|
|
f = SampleProcessor.TypeFlags
|
|
self.set_training_data_generators ([
|
|
SampleGeneratorFace(self.training_data_src_path, sort_by_yaw_target_samples_path=self.training_data_dst_path if self.sort_by_yaw else None,
|
|
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 ),
|
|
output_sample_types=[ [f.WARPED_TRANSFORMED | f.FACE_ALIGN_HALF | f.MODE_BGR, 64],
|
|
[f.TRANSFORMED | f.FACE_ALIGN_HALF | f.MODE_BGR, 64],
|
|
[f.TRANSFORMED | f.FACE_ALIGN_HALF | f.MODE_M | f.FACE_MASK_FULL, 64] ] ),
|
|
|
|
SampleGeneratorFace(self.training_data_dst_path, debug=self.is_debug(), batch_size=self.batch_size,
|
|
sample_process_options=SampleProcessor.Options(random_flip=self.random_flip),
|
|
output_sample_types=[ [f.WARPED_TRANSFORMED | f.FACE_ALIGN_HALF | f.MODE_BGR, 64],
|
|
[f.TRANSFORMED | f.FACE_ALIGN_HALF | f.MODE_BGR, 64],
|
|
[f.TRANSFORMED | f.FACE_ALIGN_HALF | f.MODE_M | f.FACE_MASK_FULL, 64] ] )
|
|
])
|
|
|
|
#override
|
|
def onSave(self):
|
|
self.save_weights_safe( [[self.encoder, self.get_strpath_storage_for_file(self.encoderH5)],
|
|
[self.decoder_src, self.get_strpath_storage_for_file(self.decoder_srcH5)],
|
|
[self.decoder_dst, self.get_strpath_storage_for_file(self.decoder_dstH5)]] )
|
|
|
|
#override
|
|
def onTrainOneEpoch(self, sample, generators_list):
|
|
warped_src, target_src, target_src_full_mask = sample[0]
|
|
warped_dst, target_dst, target_dst_full_mask = sample[1]
|
|
|
|
total, loss_src_bgr, loss_src_mask, loss_dst_bgr, loss_dst_mask = self.ae.train_on_batch( [warped_src, target_src_full_mask, warped_dst, target_dst_full_mask], [target_src, target_src_full_mask, target_dst, target_dst_full_mask] )
|
|
|
|
return ( ('loss_src', loss_src_bgr), ('loss_dst', loss_dst_bgr) )
|
|
|
|
#override
|
|
def onGetPreview(self, sample):
|
|
test_A = sample[0][1][0:4] #first 4 samples
|
|
test_A_m = sample[0][2][0:4]
|
|
test_B = sample[1][1][0:4]
|
|
test_B_m = sample[1][2][0:4]
|
|
|
|
AA, mAA = self.src_view([test_A])
|
|
AB, mAB = self.src_view([test_B])
|
|
BB, mBB = self.dst_view([test_B])
|
|
|
|
mAA = np.repeat ( mAA, (3,), -1)
|
|
mAB = np.repeat ( mAB, (3,), -1)
|
|
mBB = np.repeat ( mBB, (3,), -1)
|
|
|
|
st = []
|
|
for i in range(0, len(test_A)):
|
|
st.append ( np.concatenate ( (
|
|
test_A[i,:,:,0:3],
|
|
AA[i],
|
|
#mAA[i],
|
|
test_B[i,:,:,0:3],
|
|
BB[i],
|
|
#mBB[i],
|
|
AB[i],
|
|
#mAB[i]
|
|
), axis=1) )
|
|
|
|
return [ ('H64', np.concatenate ( st, axis=0 ) ) ]
|
|
|
|
def predictor_func (self, face):
|
|
|
|
face_64_bgr = face[...,0:3]
|
|
face_64_mask = np.expand_dims(face[...,3],-1)
|
|
|
|
x, mx = self.src_view ( [ np.expand_dims(face_64_bgr,0) ] )
|
|
x, mx = x[0], mx[0]
|
|
|
|
return np.concatenate ( (x,mx), -1 )
|
|
|
|
#override
|
|
def get_converter(self, **in_options):
|
|
from models import ConverterMasked
|
|
return ConverterMasked(self.predictor_func,
|
|
predictor_input_size=64,
|
|
output_size=64,
|
|
face_type=FaceType.HALF,
|
|
base_erode_mask_modifier=100,
|
|
base_blur_mask_modifier=100,
|
|
**in_options)
|
|
|
|
def Build(self, lighter_ae):
|
|
exec(nnlib.code_import_all, locals(), globals())
|
|
|
|
bgr_shape = (64, 64, 3)
|
|
mask_shape = (64, 64, 1)
|
|
|
|
def downscale (dim):
|
|
def func(x):
|
|
return LeakyReLU(0.1)(Conv2D(dim, 5, strides=2, padding='same')(x))
|
|
return func
|
|
|
|
def upscale (dim):
|
|
def func(x):
|
|
return PixelShuffler()(LeakyReLU(0.1)(Conv2D(dim * 4, 3, strides=1, padding='same')(x)))
|
|
return func
|
|
|
|
def Encoder(input_shape):
|
|
input_layer = Input(input_shape)
|
|
x = input_layer
|
|
if not lighter_ae:
|
|
x = downscale(128)(x)
|
|
x = downscale(256)(x)
|
|
x = downscale(512)(x)
|
|
x = downscale(1024)(x)
|
|
x = Dense(1024)(Flatten()(x))
|
|
x = Dense(4 * 4 * 1024)(x)
|
|
x = Reshape((4, 4, 1024))(x)
|
|
x = upscale(512)(x)
|
|
else:
|
|
x = downscale(128)(x)
|
|
x = downscale(256)(x)
|
|
x = downscale(512)(x)
|
|
x = downscale(768)(x)
|
|
x = Dense(512)(Flatten()(x))
|
|
x = Dense(4 * 4 * 512)(x)
|
|
x = Reshape((4, 4, 512))(x)
|
|
x = upscale(256)(x)
|
|
return Model(input_layer, x)
|
|
|
|
def Decoder():
|
|
if not lighter_ae:
|
|
input_ = Input(shape=(8, 8, 512))
|
|
x = input_
|
|
|
|
x = upscale(512)(x)
|
|
x = upscale(256)(x)
|
|
x = upscale(128)(x)
|
|
|
|
else:
|
|
input_ = Input(shape=(8, 8, 256))
|
|
|
|
x = input_
|
|
x = upscale(256)(x)
|
|
x = upscale(128)(x)
|
|
x = upscale(64)(x)
|
|
|
|
y = input_ #mask decoder
|
|
y = upscale(256)(y)
|
|
y = upscale(128)(y)
|
|
y = upscale(64)(y)
|
|
|
|
x = Conv2D(3, kernel_size=5, padding='same', activation='sigmoid')(x)
|
|
y = Conv2D(1, kernel_size=5, padding='same', activation='sigmoid')(y)
|
|
|
|
return Model(input_, [x,y])
|
|
|
|
return bgr_shape, mask_shape, Encoder(bgr_shape), Decoder(), Decoder() |