animeganv2_tensorflow2's Introduction

AnimeGANv2_Tensorflow2 中文

Reference source AnimeGANv2 project, rewritten with TensorFlow2 to implement

Installation and testing environment

GPU：3060 batch_size=10 Training time is 11min/epoch

tensorflow==2.8.0
tensorflow-addons==0.16.1
wandb
tqdm==4.63.1
PyYAML
opencv-python==4.5.5

Usage

Training

python train.py --config_path config/config-defaults.yaml --dataset Hayao --hyperparameters False

--config_path the configuration file path is under config/config-defaults.yaml by default, which is the hyperparameter configuration of the project
--dataset the name of the dataset
--hyperparameters whether to enable hyperparameter search for wandb
--pre_train_weight Pre-training weights, which can be loaded with previously trained model weights for fine-tuning and then trained into a new model

Testing

Training process

loss variation

Discriminator related losses

Generator related losses

Relative change in loss of generators and discriminators

As can be seen from the loss, the generator and discriminator produced an obvious confrontation effect, the generator loss into an upward trend, discriminator loss into a downward trend, due to the training of the relevant loss weight is in accordance with the way recommended by the original author, and the original author training effect has a certain difference, you need to adjust again

Image verification results

The overall animation style of the picture effect is relatively strong compared to the original author, but the picture details are missing more and the picture color is stronger, which needs to be further improved by adjusting the loss weight

License

This version is for academic research and non-commercial use only, if used for commercial purposes, please contact me for licensing approval

animeganv2_tensorflow2's People

Contributors

Stargazers

Watchers

animeganv2_tensorflow2's Issues

如果我想动漫风格化程度更高, 应该调整哪些参数

为啥我cuda\cudnn\tensorflow版本都对应, 还是不能使用gpu训练. 换其他代码就行, 大哥救命

cuda 11.2
cudnn 8.2.1
tensorflow 2.8

你好，我看这个工程里面数据来自于wandb，数据集可以共享一下吗

How many epoches you have trained to achieve your training results?

Thank you very much for your work, I did some attempts to convert from animeganv2_tf1 to tf2 before but failed.
But there is a question I want to ask. How many epochs you have trained to achieve your training results?
My results after 400 epochs are still poor, I don't know if it's a configuration issue or something else.

About style loss

According to the generator's style loss provided by https://github.com/TachibanaYoshino/AnimeGANv2/blob/617ab10ce21877df90c9cf2b705712bc81b1e2e6/tools/ops.py
Maybe function gram() should be expressed like:

feat = tf.linalg.einsum('bijc,bijd->bcd', x, x)
num_locations = tf.math.reduce_prod(x.shape[1:])
num_locations = tf.cast(num_locations, dtype=tf.float32)

feat = feat/num_locations

Otherwise, it may cause generator gradient explosion, and difficult to achieve Antagonism effect

Recommend Projects