# 论文复现：Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation

---

English|简体中文

• paddle-PSP
  • 1 Introduction
  • 2 Result
  • 3 Datasets
  • 4 Environment
  • 5 Pretrained models
  • 6 Quick start
    • train
    • inferernce
  • 7 Code structure
    • structure
    • Parameter description
  • 8 Model information

1 Introduction

---

This project is based on the pixel2style2pixel (pSp). pSp framework generates a series of style vectors based on a novel encoder network, which is fed into a pre-trained style generator to form an extended W + potential space. The encoder can directly reconstruct real input images.

Paper

• [1] Richardson E , Alaluf Y , Patashnik O , et al. Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation[J]. 2020.

Reference project

• https://github.com/eladrich/pixel2style2pixel

Project on Ai Studio

• notebook：https://aistudio.baidu.com/aistudio/projectdetail/2331440

2 Result

Results（Test on CelebA-HQ）

Model	LPIPS	Similarity	MSE
Paper	0.17	0.56	0.03
Pytorch	0.15	0.57	0.03
Paddle	0.17	0.57	0.03

Visual comparison

Pytorch	Paddle

3 Datasets

• Training： FFHQ-1024. saved in FFHQ/.

• Testing：CelebA-HQ.saved inCelebA_test/.

4 Environment

Hardware：GPU、CPU

Framework：PaddlePaddle >=2.0.0

5 Pretrained models

Pretrained models saved inpretrained_models/.

Pretrained models	Description
FFHQ StyleGAN(stylegan2-ffhq-config-f.pdparams)	StyleGAN trained with the FFHQ dataset fromrosinality ，output size:1024x1024
IR-SE50 Model(model_ir_se50.pdparams)	IR_SE model (TreB1eN)trained for computering ID loss.
CurricularFace Backbone(CurricularFace_Backbone.paparams)	Pretrained CurricularFace model(HuangYG123)evaled Similarity
AlexNet(alexnet.pdparams和lin_alex.pdparams)	computered lpips loss
StyleGAN Inversion(psp_ffhq_inverse.pdparams)	pSp trained with the FFHQ dataset for StyleGAN inversion.

Baidu driver：https://pan.baidu.com/s/1G-Ffs8-y93R0ZlD9mEU6Eg password：m3nb

Pretrained pSp encoder：

模型	Description
StyleGAN Inversion(psp_ffhq_inverse.pdparams)	pSp trained with the FFHQ dataset for StyleGAN inversion.

6 Quick start

Compile operation

python scripts/compile_ranger.py

Train

python scripts/train.py \
--dataset_type=ffhq_encode \
--exp_dir=exp/test \
--workers=0 \
--batch_size=8 \
--test_batch_size=8 \
--test_workers=0 \
--val_interval=2500 \
--save_interval=5000 \
--encoder_type=GradualStyleEncoder \
--start_from_latent_avg \
--lpips_lambda=0.8 \
--l2_lambda=1 \
--id_lambda=0.1 \
--optim_name=ranger

inferernce

python scripts/inference.py \
--exp_dir=inference \
--checkpoint_path=pretrained_models/psp_ffhq_inverse.pdparams \
--data_path=CelebA_test \
--test_batch_size=8 \
--test_workers=4

Others

• LPIPS

python scripts/calc_losses_on_images.py \
--mode lpips \
--data_path=inference/inference_results \
--gt_path=CelebA_test

• MSE

python scripts/calc_losses_on_images.py \
--mode l2 \
--data_path=inference/inference_results \
--gt_path=CelebA_test

• Similarity

python scripts/calc_id_loss_parallel.py \
--data_path=inference/inference_results \
--gt_path=CelebA_test

7 Code structure

Structure

├─config          # 配置
    ├─data            #数据集加载
       ├─CelebA_test  # 测试数据图像
    ├─logs            #日志
       ├─train        # 训练日志
       ├─test         # 测试日志
    ├─models          # 模型
        ├─encoders    # 编码器
        ├─loss        # 损失函数
        ├─mtcnn       #     
        ├─stylegan2   #       
        ├─utils       # 编译算子
    ├─scripts         #算法执行
        trian         #训练
        inference     #测试
    ├─utils           # 工具代码
    │  README.md      #英文readme
    │  README_cn.md   #中文readme

Parameter description

Parameter	Default
Config	None

8 Model information

The overall information of the model is as follows:

Information	Descriptions
Version	Paddle 2.1.2
Application	Image Generation
Hardware	GPU / CPU

License

# encoding=utf8
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.

# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at

#     http://www.apache.org/licenses/LICENSE-2.0

# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.