lucaslingle/e-lra

Streamlined variant of Long-Range Arena with pinned dependencies, automated data downloads, and deterministic shuffling.

E/LRA

A streamlined variant of LRA with

• pinned dependencies, avoiding installation difficulties;
• helpful examples, improving productivity;
• automatic dataset setup, avoiding manual wrangling;
• complete factory function, supporting all the models implemented by LRA;
• well-defined variables, avoiding crashing scripts;
• deterministic shuffling, aiding reproducibility.

Installation

It is recommended to install the dependencies using a virtual environment such as venv or miniconda. After the virtual environment is activated, run:

pip3 install --upgrade pip;
git clone https://github.com/lucaslingle/e-lra.git;
cd e-lra;

##### CPU-Only #####
pip3 install -e '.[cpu]' \
    -f https://storage.googleapis.com/jax-releases/jax_releases.html;

##### Nvidia GPU, CUDA 11 #####
pip3 install -e '.[cuda11]' \
    -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html;

##### Cloud TPU VM #####
pip3 install -e '.[tpu]' \
    -f https://storage.googleapis.com/jax-releases/jax_releases.html \
    -f https://storage.googleapis.com/jax-releases/libtpu_releases.html;

Prepare Data

To prepare the data, run:

source ./prep_data.sh;

Examples

To benchmark a vanilla transformer on all tasks, run the following.
To benchmark a different xformer, change the config option.

Test results are printed and written to a file results.json in the specified model_dir.

ListOps

python3 lra_benchmarks/listops/train.py \
      --task_name=basic \
      --config=lra_benchmarks/listops/configs/transformer_base.py \
      --config.eval_frequency=1000 \
      --data_dir=lra_data/listops/ \
      --model_dir=/tmp/listops/;

Text Classification

Sweep over MAX_LENGTH=1000,2000,3000,4000, and report the best result.

python3 lra_benchmarks/text_classification/train.py \
      --task_name=imdb_reviews \
      --config=lra_benchmarks/text_classification/configs/transformer_base.py \
      --config.eval_frequency=1000 \
      --config.max_length=$MAX_LENGTH \
      --data_dir=lra_data/text_classification/ \
      --model_dir=/tmp/text_classification/;

# Clean up model_dir after viewing test metrics,
# since we need to run from scratch for each MAX_LENGTH setting!
rm -rf /tmp/text_classification/;

Retrieval

python3 lra_benchmarks/retrieval/train.py \
      --task_name=basic \
      --config=lra_benchmarks/retrieval/configs/transformer_base.py \
      --config.eval_frequency=1000 \
      --data_dir=lra_data/retrieval/ \
      --model_dir=/tmp/retrieval/;

Image Classification

python3 lra_benchmarks/image/train.py \
      --task_name=cifar10 \
      --config=lra_benchmarks/image/configs/cifar10/transformer_base.py \
      --config.eval_frequency=1000 \
      --model_dir=/tmp/image/;

Pathfinder

python3 lra_benchmarks/image/train.py \
      --task_name=pathfinder32_easy \
      --config=lra_benchmarks/image/configs/pathfinder32/transformer_base.py \
      --config.eval_frequency=1000 \
      --model_dir=/tmp/pathfinder/;

Acknowledgements

Experiments supported by Cloud TPUs from Google's TPU Research Cloud (TRC).