FID index reproduction step on the pit to avoid the pit text generation image FID quantitative experiment whole process reproduction (FR é Chet inception distance) quantitative evaluation experiment s

One 、FID Introduction to scores

FID Its full name is ：Fréchet Inception Distance.

FID Scores are used to extract features from the pre training network , Measure the distance between the real image distribution and the generated image distribution . Real images obey a distribution in space （ Suppose it's a normal distribution ）, and GAN The generated feature is also a distribution ,GAN The thing to do is to keep training so that the two distributions are as same as possible .FID Is to calculate the direct distance between these two distributions , The distance algorithm used is called Frechet distance.

FID Calculate the distance between two distributions , The smaller the distance, the closer the generated distribution is to the real distribution , so FID The smaller the better. .

Two 、FID fraction CUB Quantitative experiment steps

2.1、 download FID Computational code

github download ：https://github.com/MinfengZhu/DM-GAN/tree/master/eval/FID

Put it in the code Directory

2.2、 download FID Pre trained models

Google cloud disk link ：https://drive.google.com/file/d/1747il5vnY2zNkmQ1x_8hySx537ZAJEtj
CSDN link ：FID Trained models in the light of CUB-birds Of FID Pre training model

After downloading is a npz file , Put it in the specified folder location

2.3、 Enter terminal command

Open the terminal , The input command is ：python fid_score.py --gpu 0 --batch-size 24 --path1 eval/FID/bird_val.npz --path2 ../test/valid/single
The latter two parameters path1 On behalf of you FID The location of the pre trained model ,path2 Represents the position where you put the generated image .

Display after running ：

2.4、 matters needing attention

1、 Calculation FID Generally need 30000 Zhang generates images , If only 3000 Zhang is unpredictable ;
2、 Calculation FID The standard method of is not unified at present , No use imagenet The parameters of pre training are measured FID Are not the same as , This article is just a practice , At present, it may not be rigorous enough , For reference only .

Last

Personal profile ： Graduate students in the field of artificial intelligence , At present, I mainly focus on text generation and image generation （text to image） Direction

Personal home page ： Medium coke with more ice

Time limited free subscribe ： Text generated images T2I special column

Stand by me ： give the thumbs-up + Collection ️+ Leaving a message.

in addition , We have established wechat T2I Learning exchange group , If you are T2I Fans or researchers in this field can send me a private message to join .

If this article helps you a lot , I hope you can give me a coke ！ Add more ice ！