This translation example is adapted from a pytorch tutorial.
The s4d.py is adapted from state-spaces/s4.
All jupyter notebooks assume the usable of a GPU to train the model.
Change the load_ckpt = False into load_ckpt = True (in second code cell) and then the checkpoints can be loaded from checkpoint.
The comparison of training loss curve is provided in the /assets.
The following snippet is the code for environment construction and data preparation.
conda create -n SSM_Examples python=3.10
conda activate SSM_Examples
pip install -r requirements.txt
cd SSM_Examples
wget https://download.pytorch.org/tutorial/data.zip
unzip data.zip