The Penn Treebank dataset can be found at:

https://github.com/wojzaremba/lstm/tree/master/data

After downloading the dataset, please run the following script to preprocess the files {train,valid,test}.txt into a pkl bundle:

python -m enas_lm.src.process

To search for a cell using default parameters:

python -m enas_lm.src.search \
  --output_dir="$(pwd)/output" \
  --data_path="$(pwd)/ptb/ptb.pkl"

To evaluate the found architecture with default parameters on TPUs, please first follow this guide to set up a Cloud TPU account. A 1x1 donut of TPUv2 should provide enough memory for our run. The running command is:

python -m enas_lm.src.tpu.fixed \
  --fixed_arc='0 2 1 0 2 1 2 2 4 0 5 0 3 2 6 2' \
  --output_dir=${STORAGE_BUCKET}/output \
  --data_path="${STORAGE_BUCKET}/data/ptb.pkl" \
  --use_tpu \
  --tpu_job_name=${TPU_NAME}

Our 10 runs produce test perplexity: 56.1 ± 0.1.

We also provide a GPU implementation of the method, which is much slower than the TPU version above. To evaluate the found architecture with default parameters on GPU, you can directly use the command:

python -m enas_lm.src.fixed \
  --fixed_arc='0 2 1 0 2 1 2 2 4 0 5 0 3 2 6 2' \
  --output_dir="$(pwd)/output" \
  --data_path="$(pwd)/ptb/ptb.pkl"

Our 10 runs produce test perplexity: 57.3 ± 0.1.

To evaluate the architecture found by DARTS:

python -m enas_lm.src.fixed \
  --fixed_arc='0 2 1 1 1 1 1 3 2 0 5 2 3 0 5 1' \
  --output_dir="$(pwd)/output" \
  --data_path="$(pwd)/ptb/ptb.pkl"

Expected test perplexity: 58.1

To run the AWD-LSTM baseline:

python -m enas_lm.src.lstm \
  --output_dir="$(pwd)/output" \
  --data_path="$(pwd)/ptb/ptb.pkl"

Expected test perplexity: 58.21