Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/lunarwhite/gen-anime-face
Train GAN model ref NTU MLDS. 李宏毅老师的二次元人脸/头像数据集
https://github.com/lunarwhite/gen-anime-face
acg animation deep-learning keras python tensorflow
Last synced: about 5 hours ago
JSON representation
Train GAN model ref NTU MLDS. 李宏毅老师的二次元人脸/头像数据集
- Host: GitHub
- URL: https://github.com/lunarwhite/gen-anime-face
- Owner: lunarwhite
- License: mit
- Created: 2021-06-03T01:20:00.000Z (over 3 years ago)
- Default Branch: main
- Last Pushed: 2024-07-31T04:23:37.000Z (3 months ago)
- Last Synced: 2024-07-31T06:23:04.039Z (3 months ago)
- Topics: acg, animation, deep-learning, keras, python, tensorflow
- Language: Python
- Homepage:
- Size: 150 KB
- Stars: 8
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# gen-anime-face
GAN-based image generation of animated/ACGN faces. 基于李宏毅老师的二次元人脸/头像数据集，应用训练GAN，实现二次元人脸生成
```
.
├── LICENSE
├── main.py
├── README.md
├── requirements.txt
├── res
│ └── image # 数据集
└── tmp # 保存 model 图片
```
## 1 Overview
- 数据初始来源：[MakeGirlsMoe](https://make.girls.moe/#/)
- 数据集下载：[Kaggle](https://www.kaggle.com/lunarwhite/anime-face-dataset-ntumlds)
- 选择 GAN 搭建模型，借助 Keras 搭建训练
- 主要工具包版本为 TensorFlow `2.2.0`、和 Python `3.7.10`
## 2 Setup
- clone repo：`git clone https://github.com/lunarwhite/gen-anime-face.git`
- 更新 pip：`pip3 install --upgrade pip`
- 为项目创建虚拟环境：`conda create --name python=3.7`
- 激活 env：`conda activate `
- 安装 Python 库依赖：`pip3 install -r requirements.txt`
## 3 Train
- 运行：`python main.py`
- 调参：在 `main.py` 文件中修改常用参数
```python
def main():
gan = GAN(
optimizer = Adam(0.0002, 0.5), # 优化方法，初始学习率
loss="binary_crossentropy" # 损失函数
)
gan.train(
epochs=300, # 训练轮数
batch_size=32, # 一次训练所抓取的数据样本数量
sample_interval=50 # 多少个 epoch之 后 save 一次实验结果
)
```
## 4 Workflow
- 观察数据
- 数据集大小
- 数据集样本
- 图像分辨率
- 数据预处理
- 归一化
- 搭建模型
- Update discriminator # do not update the parameters of generator
- real_images = `sample_batch_data(training_data, batch_size)`
- noise = `sample_batch_noise(batch_size, noise_dim)`
- fake_images = `generator(noise)`
- real_predicts = `discriminator(real_images)`
- fake_predicts = `discriminator(fake_images)`
- d_loss = `loss_d_fn(real_predicts, real_labels, fake_predicts, fake_labels)`
- d_grad = `gradients(d_loss, d_params)`
- d_params = `updates(d_params, d_grad)`
- Update Generator # do not update the parameters of discriminator
- noise = `sample_batch_noise(noise_dim, batch_size)`
- fake_images = `generator(noise)`
- fake_predicts = `discriminator(fake_images)`
- g_loss = `loss_g_fn( fake_predicts, real_labels)`
- g_grad = `gradients(g_loss, g_params)`
- g_params = `updates(g_params, g_grad)`
- 可视化分析
- 模型可视化
- 改进模型
- 1: Normalize the inputs
- 2: A modified loss function
- 4: BatchNorm
- 5: Avoid Sparse Gradients: ReLU, MaxPool
- 6: Use Soft and Noisy Labels
- 9: Use the ADAM Optimizer
- 10: Track failures early
- 13: Add noise to inputs, decay over time
- 17: Use Dropouts in G in both train and test phase
## 5 Reference
- [How to Train a GAN? Tips and tricks to make GANs work](https://github.com/soumith/ganhacks)