Generate images

Select a checkpoint using run logs

The validation step that occurs regularly during training generates a set of examples and metrics. We can use these to choose the best model checkpoint.

You can choose your checkpoint based on the following metrics, which you will find in the wandb logs:

  • The translation rate shows how many of the conversion network’s output are classified by the pre-trained classifier as the target class.

  • The conversion rate is similar, except it gives the model several tries to correctly convert an image. This is possible because the conversion network includes some randomness.

Note

Generally, we choose the checkpoint with the highest average conversion rate.

Make sure that the generated images, which are also in the logs as ema_fake_x_reference and ema_fake_x_latent , look realistic at that point.

Run image generation

The generate_images.py script will use the model that you have chosen to try to convert your images from one class to another.

It requires an additional set of parameters:

  • dataset: Which of the datasets to run the translation on. By default this will be the “test” dataset, if that does not exist it will revert to the “validation” dataset.

  • source_class: The name of the class to take images from.

  • target_class: The name of the class to convert images to.

  • checkpoint_iter: The number of the checkpoint you chose above.

For example, to convert the validation dataset from class_A to class_B using checkpoint 50000 you would run:

python generate_images.py --dataset validation --source_class class_A --target_class class_B --checkpoint_iter 50000

If you want to have a look at the other modifiable options, run the following:

python generate_images.py --help

Output

Once you’re done with this step, your experiment directory (solver.root_directory, in your YAML file) should look something like this:

<solver.root_directory>/
├── checkpoints/
└── generated_images/
    └── latent/
        ├── class_A/
        │   └── class_B/
        │       ├── xxx.png
        │       ├── xxy.png
        │       ├── abc.png
        │       └── ...
        ├── class_A/
        │   └── class_C/
        │       ├── xxx.png
        │       ├── xxy.png
        │       ├── abc.png
        │       └── ...
        ├── ...
        └── class_C/
            └── class_B/
                ├── asd.png
                ├── fgh.png
                ├── hjk.png
                └── ...