Integrating a discriminator to guide the model toward generating more realistic facial details. This did introduce some texture to the faces.
Browse files- Experimenting with Adversarial Loss/Discriminatorv3_3.py +196 -0
- Experimenting with Adversarial Loss/discriminator-16796-16328-37280.pth +3 -0
- Experimenting with Adversarial Loss/discriminator-580-596-640.pth +3 -0
- Experimenting with Adversarial Loss/reswapper-1679500.pth +3 -0
- Experimenting with Adversarial Loss/reswapper-1683150.pth +3 -0
- Experimenting with Adversarial Loss/train_dis.3_3_1_Good_1.1.1.py +435 -0
Experimenting with Adversarial Loss/Discriminatorv3_3.py
ADDED
@@ -0,0 +1,196 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# a modified version of https://github.com/deepinsight/insightface/blob/master/recognition/arcface_torch/backbones/iresnet.py
|
2 |
+
|
3 |
+
import torch
|
4 |
+
from torch import nn
|
5 |
+
from torch.utils.checkpoint import checkpoint
|
6 |
+
|
7 |
+
__all__ = ['iresnet18', 'iresnet34', 'iresnet50', 'iresnet100', 'iresnet200']
|
8 |
+
using_ckpt = False
|
9 |
+
|
10 |
+
def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
|
11 |
+
"""3x3 convolution with padding"""
|
12 |
+
return nn.Conv2d(in_planes,
|
13 |
+
out_planes,
|
14 |
+
kernel_size=3,
|
15 |
+
stride=stride,
|
16 |
+
padding=dilation,
|
17 |
+
groups=groups,
|
18 |
+
bias=True,
|
19 |
+
dilation=dilation)
|
20 |
+
|
21 |
+
|
22 |
+
def conv1x1(in_planes, out_planes, stride=1):
|
23 |
+
"""1x1 convolution"""
|
24 |
+
return nn.Conv2d(in_planes,
|
25 |
+
out_planes,
|
26 |
+
kernel_size=1,
|
27 |
+
stride=stride,
|
28 |
+
bias=True)
|
29 |
+
|
30 |
+
|
31 |
+
class IBasicBlock(nn.Module):
|
32 |
+
expansion = 1
|
33 |
+
def __init__(self, inplanes, planes, stride=1, downsample=None,
|
34 |
+
groups=1, base_width=64, dilation=1):
|
35 |
+
super(IBasicBlock, self).__init__()
|
36 |
+
if groups != 1 or base_width != 64:
|
37 |
+
raise ValueError('BasicBlock only supports groups=1 and base_width=64')
|
38 |
+
if dilation > 1:
|
39 |
+
raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
|
40 |
+
self.bn1 = nn.BatchNorm2d(inplanes, eps=1e-05,)
|
41 |
+
self.conv1 = conv3x3(inplanes, planes)
|
42 |
+
self.bn2 = nn.BatchNorm2d(planes, eps=1e-05,)
|
43 |
+
self.prelu = nn.PReLU(planes)
|
44 |
+
self.conv2 = conv3x3(planes, planes, stride)
|
45 |
+
self.bn3 = nn.BatchNorm2d(planes, eps=1e-05,)
|
46 |
+
self.downsample = downsample
|
47 |
+
self.stride = stride
|
48 |
+
|
49 |
+
def forward_impl(self, x):
|
50 |
+
identity = x
|
51 |
+
out = self.bn1(x)
|
52 |
+
out = self.conv1(out)
|
53 |
+
out = self.bn2(out)
|
54 |
+
out = self.prelu(out)
|
55 |
+
out = self.conv2(out)
|
56 |
+
out = self.bn3(out)
|
57 |
+
if self.downsample is not None:
|
58 |
+
identity = self.downsample(x)
|
59 |
+
out += identity
|
60 |
+
return out
|
61 |
+
|
62 |
+
def forward(self, x):
|
63 |
+
if self.training and using_ckpt:
|
64 |
+
return checkpoint(self.forward_impl, x)
|
65 |
+
else:
|
66 |
+
return self.forward_impl(x)
|
67 |
+
|
68 |
+
|
69 |
+
class IResNet(nn.Module):
|
70 |
+
fc_scale = 14 * 14
|
71 |
+
def __init__(self,
|
72 |
+
block, layers, dropout=0, num_features=512, zero_init_residual=False,
|
73 |
+
groups=1, width_per_group=64, replace_stride_with_dilation=None, fp16=False):
|
74 |
+
super(IResNet, self).__init__()
|
75 |
+
self.extra_gflops = 0.0
|
76 |
+
self.fp16 = fp16
|
77 |
+
self.inplanes = 64
|
78 |
+
self.dilation = 1
|
79 |
+
if replace_stride_with_dilation is None:
|
80 |
+
replace_stride_with_dilation = [False, False, False]
|
81 |
+
if len(replace_stride_with_dilation) != 3:
|
82 |
+
raise ValueError("replace_stride_with_dilation should be None "
|
83 |
+
"or a 3-element tuple, got {}".format(replace_stride_with_dilation))
|
84 |
+
self.groups = groups
|
85 |
+
self.base_width = width_per_group
|
86 |
+
self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=True)
|
87 |
+
self.bn1 = nn.BatchNorm2d(self.inplanes, eps=1e-05)
|
88 |
+
self.prelu = nn.PReLU(self.inplanes)
|
89 |
+
self.layer1 = self._make_layer(block, 64, layers[0], stride=2)
|
90 |
+
self.layer2 = self._make_layer(block,
|
91 |
+
128,
|
92 |
+
layers[1],
|
93 |
+
stride=2,
|
94 |
+
dilate=replace_stride_with_dilation[0])
|
95 |
+
self.layer3 = self._make_layer(block,
|
96 |
+
256,
|
97 |
+
layers[2],
|
98 |
+
stride=2,
|
99 |
+
dilate=replace_stride_with_dilation[1])
|
100 |
+
self.layer4 = self._make_layer(block,
|
101 |
+
512,
|
102 |
+
layers[3],
|
103 |
+
stride=2,
|
104 |
+
dilate=replace_stride_with_dilation[2])
|
105 |
+
self.bn2 = nn.BatchNorm2d(512 * block.expansion, eps=1e-05,)
|
106 |
+
self.dropout = nn.Dropout(p=dropout, inplace=True)
|
107 |
+
self.fc = nn.Linear(512 * block.expansion * self.fc_scale, num_features)
|
108 |
+
self.features = nn.BatchNorm1d(num_features, eps=1e-05)
|
109 |
+
nn.init.constant_(self.features.weight, 1.0)
|
110 |
+
self.features.weight.requires_grad = False
|
111 |
+
|
112 |
+
# for m in self.modules():
|
113 |
+
# if isinstance(m, nn.Conv2d):
|
114 |
+
# nn.init.normal_(m.weight, 0, 0.1)
|
115 |
+
# elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
|
116 |
+
# nn.init.constant_(m.weight, 1)
|
117 |
+
# nn.init.constant_(m.bias, 0)
|
118 |
+
|
119 |
+
# if zero_init_residual:
|
120 |
+
# for m in self.modules():
|
121 |
+
# if isinstance(m, IBasicBlock):
|
122 |
+
# nn.init.constant_(m.bn2.weight, 0)
|
123 |
+
|
124 |
+
def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
|
125 |
+
downsample = None
|
126 |
+
previous_dilation = self.dilation
|
127 |
+
if dilate:
|
128 |
+
self.dilation *= stride
|
129 |
+
stride = 1
|
130 |
+
if stride != 1 or self.inplanes != planes * block.expansion:
|
131 |
+
downsample = nn.Sequential(
|
132 |
+
conv1x1(self.inplanes, planes * block.expansion, stride),
|
133 |
+
nn.BatchNorm2d(planes * block.expansion, eps=1e-05, ),
|
134 |
+
)
|
135 |
+
layers = []
|
136 |
+
layers.append(
|
137 |
+
block(self.inplanes, planes, stride, downsample, self.groups,
|
138 |
+
self.base_width, previous_dilation))
|
139 |
+
self.inplanes = planes * block.expansion
|
140 |
+
for _ in range(1, blocks):
|
141 |
+
layers.append(
|
142 |
+
block(self.inplanes,
|
143 |
+
planes,
|
144 |
+
groups=self.groups,
|
145 |
+
base_width=self.base_width,
|
146 |
+
dilation=self.dilation))
|
147 |
+
|
148 |
+
return nn.Sequential(*layers)
|
149 |
+
|
150 |
+
def forward(self, x):
|
151 |
+
with torch.cuda.amp.autocast(self.fp16):
|
152 |
+
x = self.conv1(x)
|
153 |
+
x = self.bn1(x)
|
154 |
+
x = self.prelu(x)
|
155 |
+
x = self.layer1(x)
|
156 |
+
x = self.layer2(x)
|
157 |
+
x = self.layer3(x)
|
158 |
+
x = self.layer4(x)
|
159 |
+
x = self.bn2(x)
|
160 |
+
x = torch.flatten(x, 1)
|
161 |
+
x = self.dropout(x)
|
162 |
+
x = self.fc(x.float() if self.fp16 else x)
|
163 |
+
# x = self.features(x)
|
164 |
+
return x
|
165 |
+
|
166 |
+
|
167 |
+
def _iresnet(arch, block, layers, pretrained, progress, **kwargs):
|
168 |
+
model = IResNet(block, layers, **kwargs)
|
169 |
+
if pretrained:
|
170 |
+
raise ValueError()
|
171 |
+
return model
|
172 |
+
|
173 |
+
|
174 |
+
def iresnet18(pretrained=False, progress=True, **kwargs):
|
175 |
+
return _iresnet('iresnet18', IBasicBlock, [2, 2, 2, 2], pretrained,
|
176 |
+
progress, **kwargs)
|
177 |
+
|
178 |
+
|
179 |
+
def iresnet34(pretrained=False, progress=True, **kwargs):
|
180 |
+
return _iresnet('iresnet34', IBasicBlock, [3, 4, 6, 3], pretrained,
|
181 |
+
progress, **kwargs)
|
182 |
+
|
183 |
+
|
184 |
+
def iresnet50(pretrained=False, progress=True, **kwargs):
|
185 |
+
return _iresnet('iresnet50', IBasicBlock, [3, 4, 14, 3], pretrained,
|
186 |
+
progress, **kwargs)
|
187 |
+
|
188 |
+
|
189 |
+
def iresnet100(pretrained=False, progress=True, **kwargs):
|
190 |
+
return _iresnet('iresnet100', IBasicBlock, [3, 13, 30, 3], pretrained,
|
191 |
+
progress, **kwargs)
|
192 |
+
|
193 |
+
|
194 |
+
def iresnet200(pretrained=False, progress=True, **kwargs):
|
195 |
+
return _iresnet('iresnet200', IBasicBlock, [6, 26, 60, 6], pretrained,
|
196 |
+
progress, **kwargs)
|
Experimenting with Adversarial Loss/discriminator-16796-16328-37280.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:bd44167f0badfb6adfa6975b76c29ff360fb9e4eaa80b248768e15cb0145bec1
|
3 |
+
size 328890738
|
Experimenting with Adversarial Loss/discriminator-580-596-640.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:6f0cfcd96fcefa9b06d182d5a471ddf94db35c9143d0c3afb5b073502ec1cc07
|
3 |
+
size 328887546
|
Experimenting with Adversarial Loss/reswapper-1679500.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:0653401aad18b8c82a565ea4f954e044b2c2d72b5dde965b4c06e52abddac2cf
|
3 |
+
size 553194302
|
Experimenting with Adversarial Loss/reswapper-1683150.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:c13e960d555a7075fb38bfaa2b0fa59a6c84ca470fd85b3b0c54f526ecb32e8f
|
3 |
+
size 553194302
|
Experimenting with Adversarial Loss/train_dis.3_3_1_Good_1.1.1.py
ADDED
@@ -0,0 +1,435 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from datetime import datetime
|
2 |
+
import os
|
3 |
+
import random
|
4 |
+
import torch
|
5 |
+
import torch.optim as optim
|
6 |
+
import torch.nn.functional as F
|
7 |
+
|
8 |
+
from Discriminatorv3_3 import iresnet50
|
9 |
+
import Image
|
10 |
+
import ModelFormat
|
11 |
+
from StyleTransferLoss import StyleTransferLoss
|
12 |
+
import onnxruntime as rt
|
13 |
+
|
14 |
+
import cv2
|
15 |
+
from insightface.data import get_image as ins_get_image
|
16 |
+
from insightface.app import FaceAnalysis
|
17 |
+
import face_align
|
18 |
+
|
19 |
+
from StyleTransferModel_128 import StyleTransferModel
|
20 |
+
from torch.utils.tensorboard import SummaryWriter
|
21 |
+
|
22 |
+
inswapper_128_path = 'inswapper_128.onnx'
|
23 |
+
img_size = 128
|
24 |
+
|
25 |
+
providers = ['CUDAExecutionProvider', 'CPUExecutionProvider']
|
26 |
+
|
27 |
+
inswapperInferenceSession = rt.InferenceSession(inswapper_128_path, providers=providers)
|
28 |
+
|
29 |
+
faceAnalysis = FaceAnalysis(name='buffalo_l')
|
30 |
+
faceAnalysis.prepare(ctx_id=0, det_size=(512, 512))
|
31 |
+
|
32 |
+
class FocalLoss(torch.nn.Module):
|
33 |
+
def __init__(self, gamma=0, eps=1e-7):
|
34 |
+
super(FocalLoss, self).__init__()
|
35 |
+
self.gamma = gamma
|
36 |
+
self.eps = eps
|
37 |
+
self.ce = torch.nn.CrossEntropyLoss()
|
38 |
+
|
39 |
+
def forward(self, input, target):
|
40 |
+
logp = self.ce(input, target)
|
41 |
+
p = torch.exp(-logp)
|
42 |
+
loss = (1 - p) ** self.gamma * logp
|
43 |
+
return loss.mean()
|
44 |
+
|
45 |
+
def get_device():
|
46 |
+
return torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
47 |
+
style_loss_fn = StyleTransferLoss().to(get_device())
|
48 |
+
|
49 |
+
def patchgan_prediction(pred, threshold=0.5):
|
50 |
+
"""Process PatchGAN output to image-level decision"""
|
51 |
+
# pred shape: (batch_size, 1, 8, 8)
|
52 |
+
probabilities = torch.sigmoid(pred)
|
53 |
+
|
54 |
+
# Two aggregation strategies
|
55 |
+
patch_confidence = probabilities.mean(dim=[1,2,3]) # Average all patches
|
56 |
+
any_patch_positive = (probabilities > threshold).any(dim=[1,2,3]).float() # Any patch thinks it's real
|
57 |
+
|
58 |
+
return patch_confidence, any_patch_positive
|
59 |
+
|
60 |
+
def compute_gradient_penalty(D, real, fake):
|
61 |
+
alpha = torch.rand(real.size(0), 1, 1, 1).to(real.device)
|
62 |
+
interpolates = (alpha * real + (1 - alpha) * fake).requires_grad_(True)
|
63 |
+
d_interpolates = D(interpolates)
|
64 |
+
|
65 |
+
gradients = torch.autograd.grad(
|
66 |
+
outputs=d_interpolates,
|
67 |
+
inputs=interpolates,
|
68 |
+
grad_outputs=torch.ones_like(d_interpolates),
|
69 |
+
create_graph=True,
|
70 |
+
retain_graph=True
|
71 |
+
)[0]
|
72 |
+
|
73 |
+
gradient_penalty = ((gradients.norm(2, dim=1) - 1) ** 2).mean()
|
74 |
+
return gradient_penalty
|
75 |
+
|
76 |
+
def cosin_metric(x1,x2):
|
77 |
+
return torch.sum(x1 * x2, dim=1) / (torch.norm(x1, dim=1) * torch.norm(x2, dim=1))
|
78 |
+
|
79 |
+
def createFakeImage(datasetDir, image, enableDataAugmentation, steps, resolution, device):
|
80 |
+
targetFaceIndex = random.randint(0, len(image)-1)
|
81 |
+
sourceFaceIndex = random.randint(0, len(image)-1)
|
82 |
+
|
83 |
+
target_img=cv2.imread(f"{datasetDir}/{image[targetFaceIndex]}")
|
84 |
+
if enableDataAugmentation and steps % 2 == 0:
|
85 |
+
target_img = cv2.cvtColor(target_img, cv2.COLOR_BGR2GRAY)
|
86 |
+
target_img = cv2.cvtColor(target_img, cv2.COLOR_GRAY2BGR)
|
87 |
+
faces = faceAnalysis.get(target_img)
|
88 |
+
|
89 |
+
if targetFaceIndex != sourceFaceIndex:
|
90 |
+
source_img = cv2.imread(f"{datasetDir}/{image[sourceFaceIndex]}")
|
91 |
+
faces2 = faceAnalysis.get(source_img)
|
92 |
+
else:
|
93 |
+
faces2 = faces
|
94 |
+
|
95 |
+
if len(faces) > 0 and len(faces2) > 0:
|
96 |
+
new_aligned_face, _ = face_align.norm_crop2(target_img, faces[0].kps, img_size)
|
97 |
+
blob = Image.getBlob(new_aligned_face)
|
98 |
+
latent = Image.getLatent(faces2[0])
|
99 |
+
else:
|
100 |
+
return createFakeImage(datasetDir, image, enableDataAugmentation, steps, resolution, device)
|
101 |
+
|
102 |
+
if targetFaceIndex != sourceFaceIndex:
|
103 |
+
input = {inswapperInferenceSession.get_inputs()[0].name: blob,
|
104 |
+
inswapperInferenceSession.get_inputs()[1].name: latent}
|
105 |
+
|
106 |
+
expected_output = inswapperInferenceSession.run([inswapperInferenceSession.get_outputs()[0].name], input)[0]
|
107 |
+
else:
|
108 |
+
expected_output = blob
|
109 |
+
|
110 |
+
expected_output_tensor = torch.from_numpy(expected_output).to(device)
|
111 |
+
|
112 |
+
if resolution != 128:
|
113 |
+
new_aligned_face, _ = face_align.norm_crop2(target_img, faces[0].kps, resolution)
|
114 |
+
blob = Image.getBlob(new_aligned_face, (resolution, resolution))
|
115 |
+
|
116 |
+
latent_tensor = torch.from_numpy(latent).to(device)
|
117 |
+
target_input_tensor = torch.from_numpy(blob).to(device)
|
118 |
+
|
119 |
+
return target_input_tensor, latent_tensor, expected_output_tensor
|
120 |
+
|
121 |
+
def train(datasetDir, learning_rate=0.0001, model_path=None, outputModelFolder='', saveModelEachSteps = 1, stopAtSteps=None, logDir=None, previewDir=None, saveAs_onnx = False, resolutions = [128], enableDataAugmentation = False):
|
122 |
+
device = get_device()
|
123 |
+
print(f"Using device: {device}")
|
124 |
+
train_g = True #True
|
125 |
+
train_d = True #False
|
126 |
+
|
127 |
+
model = StyleTransferModel().to(device)
|
128 |
+
discriminator = iresnet50().to(device) # Add discriminator
|
129 |
+
# discriminator.features.weight.requires_grad = True
|
130 |
+
optimizer_D = optim.Adam(discriminator.parameters(), lr=0.00005) # S
|
131 |
+
fake_correct_count = 0
|
132 |
+
real_correct_count = 0
|
133 |
+
d_steps = 0
|
134 |
+
|
135 |
+
if model_path is not None:
|
136 |
+
model.load_state_dict(torch.load(model_path, map_location=device), strict=False)
|
137 |
+
lastSteps = 0
|
138 |
+
# lastSteps = 200
|
139 |
+
d_steps = 640
|
140 |
+
fake_correct_count=580
|
141 |
+
real_correct_count=596
|
142 |
+
|
143 |
+
discriminator.load_state_dict(torch.load(f"D:\\ReSwapper\\model\\discriminatorV4\\discriminator-{fake_correct_count}-{real_correct_count}-{d_steps}.pth", map_location=device), strict=False)
|
144 |
+
print(f"Loaded model from {model_path}")
|
145 |
+
if train_g:
|
146 |
+
lastSteps = int(model_path.split('-')[-1].split('.')[0])
|
147 |
+
print(f"Resuming training from step {lastSteps}")
|
148 |
+
d_steps *= 2
|
149 |
+
else:
|
150 |
+
lastSteps = 0
|
151 |
+
|
152 |
+
model.train()
|
153 |
+
model = model.to(device)
|
154 |
+
# criterion = FocalLoss(gamma=2).to(device)
|
155 |
+
# # criterion = torch.nn.CrossEntropyLoss().to(device)
|
156 |
+
# criterion = torch.nn.BCELoss().to(device)
|
157 |
+
criterion = torch.nn.BCELoss().to(device)
|
158 |
+
|
159 |
+
# Initialize optimizer
|
160 |
+
optimizer = optim.Adam(model.parameters(), lr=learning_rate)
|
161 |
+
# torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=50, eta_min=1e-6)
|
162 |
+
|
163 |
+
# Initialize TensorBoard writer
|
164 |
+
if logDir is not None:
|
165 |
+
train_writer = SummaryWriter(os.path.join(logDir, "training"))
|
166 |
+
val_writer = SummaryWriter(os.path.join(logDir, "validation"))
|
167 |
+
|
168 |
+
steps = 0
|
169 |
+
|
170 |
+
image = os.listdir(datasetDir)
|
171 |
+
|
172 |
+
resolutionIndex = 0
|
173 |
+
|
174 |
+
|
175 |
+
batch_size = 5
|
176 |
+
# Training loop
|
177 |
+
while True:
|
178 |
+
start_time = datetime.now()
|
179 |
+
|
180 |
+
resolution = resolutions[resolutionIndex%len(resolutions)]
|
181 |
+
optimizer.zero_grad()
|
182 |
+
|
183 |
+
# if steps % 100 == 0 or True:
|
184 |
+
real_images_list = []
|
185 |
+
|
186 |
+
fake_images_list = []
|
187 |
+
while len(real_images_list)!=batch_size:
|
188 |
+
realFaceIndex = random.randint(0, len(image)-1)
|
189 |
+
real_img = cv2.imread(f"{datasetDir}/{image[realFaceIndex]}")
|
190 |
+
faces3 = faceAnalysis.get(real_img)
|
191 |
+
if len(faces3) == 0 : continue
|
192 |
+
|
193 |
+
aligned_real_face, _ = face_align.norm_crop2(real_img, faces3[0].kps, resolution)
|
194 |
+
real_images = torch.from_numpy(Image.getBlob(aligned_real_face, (resolution, resolution))).to(device)
|
195 |
+
|
196 |
+
real_images = F.interpolate(real_images, size=(224, 224), mode='bilinear', align_corners=False)
|
197 |
+
real_images_list.append(real_images)
|
198 |
+
|
199 |
+
while len(fake_images_list)!=batch_size:
|
200 |
+
target_input_tensor, latent_tensor, expected_output_tensor = createFakeImage(datasetDir, image, enableDataAugmentation, steps, resolution, device)
|
201 |
+
|
202 |
+
with torch.no_grad():
|
203 |
+
output = model(target_input_tensor, latent_tensor)
|
204 |
+
|
205 |
+
fake_images = output.detach() # Detach to avoid backprop through generator
|
206 |
+
fake_images = F.interpolate(fake_images, size=(224, 224), mode='bilinear', align_corners=False)
|
207 |
+
|
208 |
+
fake_images_list.append(fake_images)
|
209 |
+
|
210 |
+
if train_d and resolution == 256:
|
211 |
+
# ---------------------
|
212 |
+
# Train Discriminator
|
213 |
+
# ---------------------
|
214 |
+
optimizer_D.zero_grad()
|
215 |
+
|
216 |
+
# Use ground truth as real samples
|
217 |
+
fake_images_list = torch.stack(fake_images_list, 1).to(device)
|
218 |
+
real_images_list = torch.stack(real_images_list, 1).to(device)
|
219 |
+
|
220 |
+
real_pred = discriminator(real_images_list[0])
|
221 |
+
# real_label = torch.from_numpy([1]) * 1
|
222 |
+
# real_label = real_label.float().to(device)
|
223 |
+
|
224 |
+
d_loss_real = F.binary_cross_entropy_with_logits(real_pred, torch.ones_like(real_pred))
|
225 |
+
# d_loss_real= 1- F.cosine_similarity(real_pred, torch.ones_like(real_pred))
|
226 |
+
|
227 |
+
# old
|
228 |
+
# if real_pred.mean() > 0.5:
|
229 |
+
# real_correct_count += 1
|
230 |
+
#new
|
231 |
+
mean_per_real_sample = real_pred.mean(dim=1)
|
232 |
+
|
233 |
+
# Create a boolean mask where mean > 0.5
|
234 |
+
real_mask = mean_per_real_sample > 0.5
|
235 |
+
|
236 |
+
# Sum the True values to get the count
|
237 |
+
real_correct_count += real_mask.sum().item()
|
238 |
+
#new end
|
239 |
+
|
240 |
+
# Use generator output as fake samples
|
241 |
+
# fake_images = output.detach() # Detach to avoid backprop through generator
|
242 |
+
# fake_images = F.interpolate(fake_images, size=(224, 224), mode='bilinear', align_corners=False)
|
243 |
+
fake_pred = discriminator(fake_images_list[0])
|
244 |
+
|
245 |
+
# fake_label = [0] * 1
|
246 |
+
# fake_label = fake_label.float().to(device)
|
247 |
+
|
248 |
+
# if fake_pred.mean() < 0.5:
|
249 |
+
# fake_correct_count += 1
|
250 |
+
mean_per_fake_sample = fake_pred.mean(dim=1)
|
251 |
+
|
252 |
+
# Create a boolean mask where mean > 0.5
|
253 |
+
fake_mask = mean_per_fake_sample < 0.5
|
254 |
+
|
255 |
+
# Sum the True values to get the count
|
256 |
+
fake_correct_count += fake_mask.sum().item()
|
257 |
+
|
258 |
+
d_loss_fake = F.binary_cross_entropy_with_logits(fake_pred, torch.zeros_like(real_pred) * -1)
|
259 |
+
# d_loss_fake= 1- F.cosine_similarity(fake_pred, torch.zeros_like(real_pred))
|
260 |
+
# d_loss_fake_v2 = 1 - cosin_metric(fake_pred[0], torch.zeros_like(real_pred)[0])
|
261 |
+
d_loss = d_loss_real + d_loss_fake
|
262 |
+
d_loss.backward()
|
263 |
+
optimizer_D.step()
|
264 |
+
d_steps += batch_size * 2
|
265 |
+
|
266 |
+
# real, p = patchgan_prediction(real_pred)
|
267 |
+
# fake, p2 = patchgan_prediction(fake_pred)
|
268 |
+
|
269 |
+
#Train Gen
|
270 |
+
if train_g:
|
271 |
+
|
272 |
+
target_input_tensor, latent_tensor, expected_output_tensor = createFakeImage(datasetDir, image, enableDataAugmentation, steps, resolution, device)
|
273 |
+
|
274 |
+
output = model(target_input_tensor, latent_tensor)
|
275 |
+
|
276 |
+
if (resolution != 128):
|
277 |
+
output_128 = F.interpolate(output, size=(128, 128), mode='bilinear', align_corners=False)
|
278 |
+
|
279 |
+
content_loss, identity_loss = style_loss_fn(output_128, expected_output_tensor)
|
280 |
+
# Adversarial loss
|
281 |
+
output_224 = F.interpolate(output, size=(224, 224), mode='bilinear', align_corners=False)
|
282 |
+
|
283 |
+
fake_pred = discriminator(output_224)
|
284 |
+
adversarial_loss = F.binary_cross_entropy_with_logits(fake_pred, torch.ones_like(fake_pred))
|
285 |
+
|
286 |
+
loss = content_loss + adversarial_loss
|
287 |
+
|
288 |
+
if identity_loss is not None:
|
289 |
+
loss +=identity_loss
|
290 |
+
|
291 |
+
|
292 |
+
loss.backward()
|
293 |
+
|
294 |
+
optimizer.step()
|
295 |
+
|
296 |
+
steps += 1
|
297 |
+
totalSteps = steps + lastSteps
|
298 |
+
|
299 |
+
acc = (fake_correct_count+real_correct_count)/ d_steps
|
300 |
+
|
301 |
+
if logDir is not None:
|
302 |
+
if train_g:
|
303 |
+
train_writer.add_scalar("Loss/total", loss.item(), totalSteps)
|
304 |
+
train_writer.add_scalar("Loss/content_loss", content_loss.item(), totalSteps)
|
305 |
+
train_writer.add_scalar("Loss/adversarial_loss", adversarial_loss.item(), totalSteps)
|
306 |
+
|
307 |
+
if identity_loss is not None:
|
308 |
+
train_writer.add_scalar("Loss/identity_loss", identity_loss.item(), totalSteps)
|
309 |
+
|
310 |
+
if train_d:
|
311 |
+
train_writer.add_scalar("Loss/d_acc", acc, totalSteps)
|
312 |
+
|
313 |
+
train_writer.add_scalar("Loss/d_loss", d_loss.item(), totalSteps)
|
314 |
+
train_writer.add_scalar("Loss/d_loss_fake", d_loss_fake.item(), totalSteps)
|
315 |
+
train_writer.add_scalar("Loss/d_loss_real", d_loss_real.item(), totalSteps)
|
316 |
+
|
317 |
+
elapsed_time = datetime.now() - start_time
|
318 |
+
|
319 |
+
if train_d:
|
320 |
+
print(f"Total Steps: {totalSteps}, Step: {steps}, D_Loss: {d_loss.item():.4f}, d_loss_real: {d_loss_real.item():.4f}, d_loss_fake: {d_loss_fake.item():.4f}, acc: {(acc):.4f}, Elapsed time: {elapsed_time}")
|
321 |
+
if train_g:
|
322 |
+
print(f"Total Steps: {totalSteps}, Step: {steps}, G_Loss: {loss.item():.4f}, Elapsed time: {elapsed_time}")
|
323 |
+
|
324 |
+
if steps % saveModelEachSteps == 0:
|
325 |
+
if train_g:
|
326 |
+
outputModelPath = f"reswapper-{totalSteps}.pth"
|
327 |
+
if outputModelFolder != '':
|
328 |
+
outputModelPath = f"{outputModelFolder}/{outputModelPath}"
|
329 |
+
saveModel(model, outputModelPath)
|
330 |
+
if train_d:
|
331 |
+
discriminatorModelPath = f"discriminator-{fake_correct_count}-{real_correct_count}-{d_steps}.pth"
|
332 |
+
if outputModelFolder != '':
|
333 |
+
discriminatorModelPath = f"{outputModelFolder}/{discriminatorModelPath}"
|
334 |
+
saveModel(discriminator, discriminatorModelPath)
|
335 |
+
|
336 |
+
if train_g:
|
337 |
+
validation_total_loss, validation_content_loss, validation_identity_loss, swapped_face, swapped_face_256 = validate(outputModelPath)
|
338 |
+
if previewDir is not None:
|
339 |
+
cv2.imwrite(f"{previewDir}/{totalSteps}.jpg", swapped_face)
|
340 |
+
cv2.imwrite(f"{previewDir}/{totalSteps}_256.jpg", swapped_face_256)
|
341 |
+
|
342 |
+
if logDir is not None:
|
343 |
+
val_writer.add_scalar("Loss/total", validation_total_loss.item(), totalSteps)
|
344 |
+
val_writer.add_scalar("Loss/content_loss", validation_content_loss.item(), totalSteps)
|
345 |
+
if validation_identity_loss is not None:
|
346 |
+
val_writer.add_scalar("Loss/identity_loss", validation_identity_loss.item(), totalSteps)
|
347 |
+
|
348 |
+
if saveAs_onnx :
|
349 |
+
ModelFormat.save_as_onnx_model(outputModelPath)
|
350 |
+
|
351 |
+
if stopAtSteps is not None and steps == stopAtSteps:
|
352 |
+
exit()
|
353 |
+
|
354 |
+
resolutionIndex += 1
|
355 |
+
|
356 |
+
def saveModel(model, outputModelPath):
|
357 |
+
torch.save(model.state_dict(), outputModelPath)
|
358 |
+
|
359 |
+
def load_model(model_path):
|
360 |
+
device = get_device()
|
361 |
+
model = StyleTransferModel().to(device)
|
362 |
+
model.load_state_dict(torch.load(model_path, map_location=device), strict=False)
|
363 |
+
|
364 |
+
model.eval()
|
365 |
+
return model
|
366 |
+
|
367 |
+
def swap_face(model, target_face, source_face_latent):
|
368 |
+
device = get_device()
|
369 |
+
|
370 |
+
target_tensor = torch.from_numpy(target_face).to(device)
|
371 |
+
source_tensor = torch.from_numpy(source_face_latent).to(device)
|
372 |
+
|
373 |
+
with torch.no_grad():
|
374 |
+
swapped_tensor = model(target_tensor, source_tensor)
|
375 |
+
|
376 |
+
swapped_face = Image.postprocess_face(swapped_tensor)
|
377 |
+
|
378 |
+
return swapped_face, swapped_tensor
|
379 |
+
|
380 |
+
# test image
|
381 |
+
test_img = ins_get_image('t1')
|
382 |
+
|
383 |
+
test_faces = faceAnalysis.get(test_img)
|
384 |
+
test_faces = sorted(test_faces, key = lambda x : x.bbox[0])
|
385 |
+
test_target_face, _ = face_align.norm_crop2(test_img, test_faces[0].kps, img_size)
|
386 |
+
test_target_face = Image.getBlob(test_target_face)
|
387 |
+
test_l = Image.getLatent(test_faces[2])
|
388 |
+
|
389 |
+
test_target_face_256, _ = face_align.norm_crop2(test_img, test_faces[0].kps, 256)
|
390 |
+
test_target_face_256 = Image.getBlob(test_target_face_256, (256, 256))
|
391 |
+
|
392 |
+
test_input = {inswapperInferenceSession.get_inputs()[0].name: test_target_face,
|
393 |
+
inswapperInferenceSession.get_inputs()[1].name: test_l}
|
394 |
+
|
395 |
+
test_inswapperOutput = inswapperInferenceSession.run([inswapperInferenceSession.get_outputs()[0].name], test_input)[0]
|
396 |
+
|
397 |
+
def validate(modelPath):
|
398 |
+
model = load_model(modelPath)
|
399 |
+
swapped_face, swapped_tensor= swap_face(model, test_target_face, test_l)
|
400 |
+
swapped_face_256, _= swap_face(model, test_target_face_256, test_l)
|
401 |
+
|
402 |
+
validation_content_loss, validation_identity_loss = style_loss_fn(swapped_tensor, torch.from_numpy(test_inswapperOutput).to(get_device()))
|
403 |
+
|
404 |
+
validation_total_loss = validation_content_loss
|
405 |
+
if validation_identity_loss is not None:
|
406 |
+
validation_total_loss += validation_identity_loss
|
407 |
+
|
408 |
+
return validation_total_loss, validation_content_loss, validation_identity_loss, swapped_face, swapped_face_256
|
409 |
+
|
410 |
+
def main():
|
411 |
+
outputModelFolder = "model/discriminatorV4"
|
412 |
+
modelPath = None
|
413 |
+
modelPath = f"model/discriminatorV4/reswapper-1679500.pth"
|
414 |
+
|
415 |
+
logDir = "training/log/moreRes"
|
416 |
+
previewDir = "training/preview/moreRes"
|
417 |
+
datasetDir = "FFHQ"
|
418 |
+
|
419 |
+
os.makedirs(outputModelFolder, exist_ok=True)
|
420 |
+
os.makedirs(previewDir, exist_ok=True)
|
421 |
+
|
422 |
+
train(
|
423 |
+
datasetDir=datasetDir,
|
424 |
+
model_path=modelPath,
|
425 |
+
learning_rate=0.000001,
|
426 |
+
resolutions = [256],
|
427 |
+
enableDataAugmentation=True,
|
428 |
+
outputModelFolder=outputModelFolder,
|
429 |
+
saveModelEachSteps = 100,
|
430 |
+
stopAtSteps = 70000,
|
431 |
+
logDir=f"{logDir}/{datetime.now().strftime('%Y%m%d %H%M%S')}",
|
432 |
+
previewDir=previewDir)
|
433 |
+
|
434 |
+
if __name__ == "__main__":
|
435 |
+
main()
|