Skip to content

Latest commit

 

History

History
42 lines (25 loc) · 2.32 KB

README.md

File metadata and controls

42 lines (25 loc) · 2.32 KB
Raw

Batch_FID_PyTorchPix2Pix

FID is one of the main metrics used to evaluate generative adversarial networks models, so in this repository we propose a calculation method to check the FID score over each epoch using the test data, the method is formated to the PyTorch implementation of Pix2Pix and uses the packages move-my-files and pytorch-fid. The method is composed of a few steps:

  1. test model with 5th_EpochCheckpoint
  2. Separate REAL images and FAKE images of target domain
  3. Calculate FID between REAL_B ←→ FAKE_B
  4. test model with 10th_EpochCheckpoint ..... AND LOOP AGAIN UNTIL 200 / 400 epoch checkpoint

Note that the FID is calculated over the real data and generated data --> INPUT DATA IS SIMPLY IGNORED

Basic Usage

Run test.py of all checkpoints of the trained model using the PyTorch implementation of Pix2Pix

!python test.py  --num_test 160 --epoch 5 --dataroot ./datasets/EXPERIMENT_NAME --name EXPERIMENT_NAME --netG --norm  --phase test --preprocess --model pix2pix
!python test.py  --num_test 160 --epoch 10 --dataroot ./datasets/EXPERIMENT_NAME --name EXPERIMENT_NAME --netG --norm --phase test --preprocess --model pix2pix
!python test.py  --num_test 160 --epoch 15 --dataroot ./datasets/EXPERIMENT_NAME --name EXPERIMENT_NAME --netG --norm --phase test --preprocess --model pix2pix
!python test.py  --num_test 160 --epoch 20 --dataroot ./datasets/EXPERIMENT_NAME --name EXPERIMENT_NAME --netG --norm --phase test --preprocess --model pix2pix
...

Then open Batch_FID_PyTorchPix2Pix.ipynb and set EXPERIMENT_NAME and Epoch_range

Run the cells to generate the lines of code for the calculation method

Copy the output and run on the next cell --> CALCULATE FID FOR EACH EPOCH

TODO

Generate a script to delete generated files after calculation

Append fid_value from pytorch-fid to a list

Create csv file with FID per EPOCH

Plot a line graph to check increase or decrease on the quality of the generated images

Think about other implementation methods like: input domain ←→ target domain or the usability of it with CycleGAN...