Implementation of ulmfit using basenet wrappers.
Code in ulmfit.py is adapted directly from fastai code.
conda create -n ulm_env python=3.6 pip -y
source activate ulm_env
# pytorch
conda install pytorch pytorch=0.3.1 cuda90 -c pytorch -y
# spacy (for tokenization)
conda install -c conda-forge spacy -y
python -m spacy download en
# additional requirements
pip install -r requirements.txt
See ./run.sh for usage.
- Update to pytorch==0.4
- !! Looked into this on 2018-12-02 -- pytorch==0.4 has a bug that doesn't allow freezing RNNs in the way we need to. May need to wait for pytorch==1.0 for a fix, so I'm waiting for now.