Feature Distillation Improves Zero-Shot Transfer from Synthetic Images

This is the code release for the TMLR publication "Feature Distillation Improves Zero-Shot Transfer from Synthetic Images" by Niclas Popp, Jan Hendrik Metzen and Matthias Hein. The following figure summarizes the experimental setup:

The codebase consists of three components:

1. Domain-agnostic Distillation
2. Synthetic Data Generation
3. Domain-specific Distillation

The required packages are listed in the requirements.txt file. The code was tested on NVIDIA v100 and h100 gpus.

Step 1: Domain-agnostic Distillation

For an example file to run domain-agnostic distillation together with the available hyperparameters see: example_domain_agnostic.sh The code is built to use the webdataset dataloader together with .tar files. For details in how to setup the data for this kind of dataloader see here

Step 2: Synthetic data Generation

The synthetic data generation process can be started as shown in the example_data_generation.sh file. For different domains, select the corresponding dataset option.

Step 3: Domain-specific Distillation

The final step of our framework is domain-specific distillation. An example together with the available options is given in the file example_domain_specific.sh. This step requires the final model checkpoint from step 1 and the synthetic data from step 2.