Is index reproduction text generation image is score quantitative experiment whole process reproduction inception score quantitative evaluation experiment step on the pit and avoid the pit process

One 、IS Introduction to scores

The evaluation of text generated images is also a very challenging work ,IS Score is a general index to measure image quality .
IS Yes KL Mathematical knowledge of divergence and entropy , Its main principle is to calculate p(y|x) and p（y） The divergence between ,IS The higher the score, the better .

Two 、IS fraction CUB Quantitative experiment steps

With DF-GAN As an example ：

First step ：B_VALIDATION Change it to True

according to DF-GAN Official documents of , First of all, will bird.yml Medium B_VALIDATION Change it to True：

The second step ： Configure the trained generator

Configure trained model, And put it in code/bird in , Rename it to netG.pth,DF-GAN The official also gave a well-trained model ： The official link , Those who cannot open the official link can click CSDN link , After downloading, put it into the specified folder .

The third step ： Sampling generates images

function ：python main.py --cfg cfg/bird.yml

Possible problems 1：RuntimeError: Attempting to deserialize object on CUDA device
1 but torch.cuda.device_count() is 1
Question why ： The model uses several blocks during training GPU, But this machine only 1 block GPU
Solution ： Designated for “cuda:0”, Find two torch.load(）
The first place is main.py Of the 57 That's ok ：netG.load_state_dict(torch.load('models/%s/netG.pth' % (cfg.CONFIG_NAME)) Change it to ：netG.load_state_dict(torch.load('models/%s/netG.pth' % (cfg.CONFIG_NAME), map_location='cuda:0'))
The second place is the second 254 Change it to ：state_dict = torch.load(cfg.TEXT.DAMSM_NAME, map_location='cuda:0')

Possible problems 2：
RuntimeError: Error(s) in loading state_dict for NetG:
size mismatch for fc.weight: copying a param with shape torch.Size([4096, 100]) from checkpoint, the shape in current model is torch.Size([1024, 100]).
size mismatch for fc.bias: copying a param with shape torch.Size([4096]) from checkpoint, the shape in current model is torch.Size([1024]).
size mismatch for block0.c1.weight: copying a param with shape torch.Size([256, 256, 3, 3]) from checkpoint, the shape in current model is torch.Size([64, 64, 3, 3])
Question why ： When loading the model , The parameter does not correspond to , Size mismatch
Solution ： Check the trained model model.py And do quantitative at this time model.py Whether or not the same , see yml Whether the size of the file is set the same , especially NF Parameter settings . If used DF-GAN Officially trained models , To put NF Set to 32, Others should also be consistent ：

After successful operation , It will automatically sample and generate several pictures , The successful

Step four ： download IS Code and configure

Official IS fraction （StackGAN Used ） Code ：github link , Put it in the folder in the following location ：

Step five ： Download pre trained inception mode And configuration

Download pre trained inception model： The official link 、 If you can't open it, you can click ：CSDN link , Put it in the specified location of the folder , The location is as follows ：

Step six ： To configure inception_score.py, Change the read path

open inception_score.py Code for , Locate the 41 Row to 49 That's ok ：

You can see that there are two file addresses , The first address indicates where to read the pre trained inception model checkpoints, Change it to your own path ：

tf.app.flags.DEFINE_string('checkpoint_dir',
                           '../birds_valid299/model.ckpt',
                           """Path where to read model checkpoints.""")

The second path refers to where to read the sampled pictures , Also change it to your own path ：

tf.app.flags.DEFINE_string('image_folder',
                           '../test/valid/single',
                           """Path where to load the images """)

Step seven ： Start quantitative evaluation

Input... At the terminal ：python inception_score.py --image_folder ../test/valid/single
It is shown as follows ：

It's done , It can be seen that the result of this time is 4.61, Error is ±0.13！

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.

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