We conducted most of our experiments using the following versions of packages:
yaml~=0.2.5
pyyaml~=5.4.1
tensorflow~=2.4.2
tensorflow-datasets~=4.3.0
numpy~=1.19.5
tqdm~=4.61.2
However, different versions of TensorFlow (2.3, 2.5) have been shown to work as well.
We expect readers who want to replicate our experiments to be using a machine with at least one GPU.
Recreating them using CPU will be very time-consuming and might require some changes in code.
Ready-to-use experiment definitions are available in the experiments directory.
That can be easily modified to get different experiments.
We tried to select parameter names to be self-explanatory.
You can run each experiment using the run.py script with --exp flag.
For example:
python run.py --exp=experiments/resnet-20-iterative.yaml
If you have multiple GPUs available, you can use --gpu flag to choose which one should be used.
Otherwise, the default is the first listed GPU (usually /device:GPU:0).
GPU indexing starts at 0.
To use the second GPU, you can run:
python run.py --exp=experiments/resnet-20-one-shot.yaml --gpu=1
Model checkpoints will be saved under the path specified in .yaml file, for example:
steps_per_epoch: 2000
save_model:
4: data/resnet-20-iterative/baseline-ep4.h5
36: data/resnet-20-iterative/baseline-ep36.h5
In this example, there will be two checkpoints: one after 8000 (2000 * 4) and the second after 72000 (2000 * 36) iterations. Those can be used as starting points for other experiments.
Simple logs will be available under the path specified in .yaml file:
logs: data/resnet-20-iterative/info.yaml
More detailed logs in the form of TensorBoard logs will be available under:
tensorboard: data/resnet-20-one-shot/tb
To open those, please refer to TensorBoard.
We can upload some checkpoints via Zenodo platform. Please contact us directly.