github/DeepFaceLab
1
0
Fork 0
mirror of https://github.com/iperov/DeepFaceLab.git synced 2025-08-20 13:33:24 -07:00
Code Issues Projects Releases Packages Wiki Activity Actions

Upgraded to TF version 1.13.2

Browse source 
Removed the wait at first launch for most graphics cards.

Increased speed of training by 10-20%, but you have to retrain all models from scratch.

SAEHD:

added option 'use float16'
	Experimental option. Reduces the model size by half.
	Increases the speed of training.
	Decreases the accuracy of the model.
	The model may collapse or not train.
	Model may not learn the mask in large resolutions.

true_face_training option is replaced by
"True face power". 0.0000 .. 1.0
Experimental option. Discriminates the result face to be more like the src face. Higher value - stronger discrimination.
Comparison - https://i.imgur.com/czScS9q.png
This commit is contained in:
Colombo 2020-01-25 21:58:19 +04:00
commit 76ca79216e
49 changed files with 1320 additions and 1297 deletions
Download patch file Download diff file Expand all files Collapse all files

2
core/imagelib/IEPolys.py
View file

@ -54,7 +54,7 @@ class IEPolys:
self.n = max(0, self.n-1)
self.dirty = True
return self.n
def n_inc(self):
self.n = min(len(self.list), self.n+1)
self.dirty = True

94
core/imagelib/color_transfer.py
View file

@ -9,7 +9,7 @@ from scipy.sparse.linalg import spsolve
def color_transfer_sot(src,trg, steps=10, batch_size=5, reg_sigmaXY=16.0, reg_sigmaV=5.0):
"""
Color Transform via Sliced Optimal Transfer
ported by @iperov from https://github.com/dcoeurjo/OTColorTransfer
ported by @iperov from https://github.com/dcoeurjo/OTColorTransfer
src - any float range any channel image
dst - any float range any channel image, same shape as src
@ -17,7 +17,7 @@ def color_transfer_sot(src,trg, steps=10, batch_size=5, reg_sigmaXY=16.0, reg_si
batch_size - solver batch size
reg_sigmaXY - apply regularization and sigmaXY of filter, otherwise set to 0.0
reg_sigmaV - sigmaV of filter
return value - clip it manually
"""
if not np.issubdtype(src.dtype, np.floating):
@ -27,11 +27,11 @@ def color_transfer_sot(src,trg, steps=10, batch_size=5, reg_sigmaXY=16.0, reg_si
if len(src.shape) != 3:
raise ValueError("src shape must have rank 3 (h,w,c)")
if src.shape != trg.shape:
raise ValueError("src and trg shapes must be equal")
src_dtype = src.dtype
if src.shape != trg.shape:
raise ValueError("src and trg shapes must be equal")
src_dtype = src.dtype
h,w,c = src.shape
new_src = src.copy()
@ -59,63 +59,63 @@ def color_transfer_sot(src,trg, steps=10, batch_size=5, reg_sigmaXY=16.0, reg_si
src_diff_filt = src_diff_filt[...,None]
new_src = src + src_diff_filt
return new_src
def color_transfer_mkl(x0, x1):
eps = np.finfo(float).eps
h,w,c = x0.shape
h1,w1,c1 = x1.shape
x0 = x0.reshape ( (h*w,c) )
x1 = x1.reshape ( (h1*w1,c1) )
a = np.cov(x0.T)
b = np.cov(x1.T)
Da2, Ua = np.linalg.eig(a)
Da = np.diag(np.sqrt(Da2.clip(eps, None)))
Da = np.diag(np.sqrt(Da2.clip(eps, None)))
C = np.dot(np.dot(np.dot(np.dot(Da, Ua.T), b), Ua), Da)
Dc2, Uc = np.linalg.eig(C)
Dc = np.diag(np.sqrt(Dc2.clip(eps, None)))
Dc = np.diag(np.sqrt(Dc2.clip(eps, None)))
Da_inv = np.diag(1./(np.diag(Da)))
t = np.dot(np.dot(np.dot(np.dot(np.dot(np.dot(Ua, Da_inv), Uc), Dc), Uc.T), Da_inv), Ua.T)
t = np.dot(np.dot(np.dot(np.dot(np.dot(np.dot(Ua, Da_inv), Uc), Dc), Uc.T), Da_inv), Ua.T)
mx0 = np.mean(x0, axis=0)
mx1 = np.mean(x1, axis=0)
result = np.dot(x0-mx0, t) + mx1
return np.clip ( result.reshape ( (h,w,c) ).astype(x0.dtype), 0, 1)
def color_transfer_idt(i0, i1, bins=256, n_rot=20):
relaxation = 1 / n_rot
h,w,c = i0.shape
h1,w1,c1 = i1.shape
i0 = i0.reshape ( (h*w,c) )
i1 = i1.reshape ( (h1*w1,c1) )
n_dims = c
d0 = i0.T
d1 = i1.T
for i in range(n_rot):
r = sp.stats.special_ortho_group.rvs(n_dims).astype(np.float32)
d0r = np.dot(r, d0)
d1r = np.dot(r, d1)
d_r = np.empty_like(d0)
for j in range(n_dims):
lo = min(d0r[j].min(), d1r[j].min())
hi = max(d0r[j].max(), d1r[j].max())
p0r, edges = np.histogram(d0r[j], bins=bins, range=[lo, hi])
p1r, _ = np.histogram(d1r[j], bins=bins, range=[lo, hi])
@ -124,11 +124,11 @@ def color_transfer_idt(i0, i1, bins=256, n_rot=20):
cp1r = p1r.cumsum().astype(np.float32)
cp1r /= cp1r[-1]
f = np.interp(cp0r, cp1r, edges[1:])
d_r[j] = np.interp(d0r[j], edges[1:], f, left=0, right=bins)
d0 = relaxation * np.linalg.solve(r, (d_r - d0r)) + d0
return np.clip ( d0.T.reshape ( (h,w,c) ).astype(i0.dtype) , 0, 1)
@ -137,16 +137,16 @@ def laplacian_matrix(n, m):
mat_D = scipy.sparse.lil_matrix((m, m))
mat_D.setdiag(-1, -1)
mat_D.setdiag(4)
mat_D.setdiag(-1, 1)
mat_A = scipy.sparse.block_diag([mat_D] * n).tolil()
mat_D.setdiag(-1, 1)
mat_A = scipy.sparse.block_diag([mat_D] * n).tolil()
mat_A.setdiag(-1, 1*m)
mat_A.setdiag(-1, -1*m)
mat_A.setdiag(-1, -1*m)
return mat_A
def seamless_clone(source, target, mask):
h, w,c = target.shape
result = []
mat_A = laplacian_matrix(h, w)
laplacian = mat_A.tocsc()
@ -155,7 +155,7 @@ def seamless_clone(source, target, mask):
mask[:,0] = 1
mask[:,-1] = 1
q = np.argwhere(mask==0)
k = q[:,1]+q[:,0]*w
mat_A[k, k] = 1
mat_A[k, k + 1] = 0
@ -163,22 +163,22 @@ def seamless_clone(source, target, mask):
mat_A[k, k + w] = 0
mat_A[k, k - w] = 0
mat_A = mat_A.tocsc()
mat_A = mat_A.tocsc()
mask_flat = mask.flatten()
for channel in range(c):
source_flat = source[:, :, channel].flatten()
target_flat = target[:, :, channel].flatten()
target_flat = target[:, :, channel].flatten()
mat_b = laplacian.dot(source_flat)*0.75
mat_b[mask_flat==0] = target_flat[mask_flat==0]
x = spsolve(mat_A, mat_b).reshape((h, w))
result.append (x)
return np.clip( np.dstack(result), 0, 1 )
def reinhard_color_transfer(target, source, clip=False, preserve_paper=False, source_mask=None, target_mask=None):
"""
Transfers the color distribution from the source to the target
@ -368,26 +368,26 @@ def color_hist_match(src_im, tar_im, hist_match_threshold=255):
def color_transfer_mix(img_src,img_trg):
img_src = (img_src*255.0).astype(np.uint8)
img_trg = (img_trg*255.0).astype(np.uint8)
img_src_lab = cv2.cvtColor(img_src, cv2.COLOR_BGR2LAB)
img_trg_lab = cv2.cvtColor(img_trg, cv2.COLOR_BGR2LAB)
rct_light = np.clip ( linear_color_transfer(img_src_lab[...,0:1].astype(np.float32)/255.0,
rct_light = np.clip ( linear_color_transfer(img_src_lab[...,0:1].astype(np.float32)/255.0,
img_trg_lab[...,0:1].astype(np.float32)/255.0 )[...,0]*255.0,
0, 255).astype(np.uint8)
0, 255).astype(np.uint8)
img_src_lab[...,0] = (np.ones_like (rct_light)*100).astype(np.uint8)
img_src_lab = cv2.cvtColor(img_src_lab, cv2.COLOR_LAB2BGR)
img_src_lab = cv2.cvtColor(img_src_lab, cv2.COLOR_LAB2BGR)
img_trg_lab[...,0] = (np.ones_like (rct_light)*100).astype(np.uint8)
img_trg_lab = cv2.cvtColor(img_trg_lab, cv2.COLOR_LAB2BGR)
img_rct = color_transfer_sot( img_src_lab.astype(np.float32), img_trg_lab.astype(np.float32) )
img_rct = np.clip(img_rct, 0, 255).astype(np.uint8)
img_rct = cv2.cvtColor(img_rct, cv2.COLOR_BGR2LAB)
img_rct = cv2.cvtColor(img_rct, cv2.COLOR_BGR2LAB)
img_rct[...,0] = rct_light
img_rct = cv2.cvtColor(img_rct, cv2.COLOR_LAB2BGR)
return (img_rct / 255.0).astype(np.float32)

10
core/imagelib/common.py
View file

@ -13,24 +13,24 @@ def normalize_channels(img, target_channels):
if c == 0 and target_channels > 0:
img = img[...,np.newaxis]
c = 1
if c == 1 and target_channels > 1:
img = np.repeat (img, target_channels, -1)
c = target_channels
if c > target_channels:
img = img[...,0:target_channels]
c = target_channels
return img
def cut_odd_image(img):
h, w, c = img.shape
wm, hm = w % 2, h % 2
if wm + hm != 0:
if wm + hm != 0:
img = img[0:h-hm,0:w-wm,:]
return img
def overlay_alpha_image(img_target, img_source, xy_offset=(0,0) ):
(h,w,c) = img_source.shape
if c != 4:

4
core/imagelib/text.py
View file

@ -16,7 +16,7 @@ def _get_pil_font (font, size):
def get_text_image( shape, text, color=(1,1,1), border=0.2, font=None):
h,w,c = shape
try:
try:
pil_font = _get_pil_font( localization.get_default_ttf_font_name() , h-2)
canvas = Image.new('RGB', (w,h) , (0,0,0) )
@ -25,7 +25,7 @@ def get_text_image( shape, text, color=(1,1,1), border=0.2, font=None):
draw.text(offset, text, font=pil_font, fill=tuple((np.array(color)*255).astype(np.int)) )
result = np.asarray(canvas) / 255
if c > 3:
result = np.concatenate ( (result, np.ones ((h,w,c-3)) ), axis=-1 )
elif c < 3:

6
core/imagelib/warp.py
View file

@ -6,7 +6,7 @@ def gen_warp_params (source, flip, rotation_range=[-10,10], scale_range=[-0.5, 0
h,w,c = source.shape
if (h != w):
raise ValueError ('gen_warp_params accepts only square images.')
if rnd_seed != None:
rnd_state = np.random.RandomState (rnd_seed)
else:
@ -15,9 +15,9 @@ def gen_warp_params (source, flip, rotation_range=[-10,10], scale_range=[-0.5, 0
rotation = rnd_state.uniform( rotation_range[0], rotation_range[1] )
scale = rnd_state.uniform(1 +scale_range[0], 1 +scale_range[1])
tx = rnd_state.uniform( tx_range[0], tx_range[1] )
ty = rnd_state.uniform( ty_range[0], ty_range[1] )
ty = rnd_state.uniform( ty_range[0], ty_range[1] )
p_flip = flip and rnd_state.randint(10) < 4
#random warp by grid
cell_size = [ w // (2**i) for i in range(1,4) ] [ rnd_state.randint(3) ]
cell_count = w // cell_size + 1