Preliminaries

Install Anaconda! Update: conda update -n base -c defaults conda

Pull the submodules:

git submodule init
git submodule update

Chemprop

Publication: Analyzing Learned Molecular Representations for Property Prediction

Instructions

Install requirements:

cd chemprop/
conda env create -f environment.yml
conda activate chemprop
conda install cudatoolkit=10.1 -c pytorch

Decompress data:

tar xvf data.tar.gz

Run microbenchmark:

python train.py --data_path data/tox21.csv --dataset_type classification --save_dir tox21_checkpoints

Results

• GPU activity: ~30%
• CPU activity: ~100%

n2nmn

Publication: Learning to Reason: End-to-End Module Networks for Visual Question Answering

Instructions

Install requirements:

conda create --name n2nmn python=3.5
conda activate n2nmn
python -m pip install tensorflow-gpu==1.0.0
python -m pip install https://storage.googleapis.com/tensorflow_fold/tensorflow_fold-0.0.1-py3-none-linux_x86_64.whl

Get and process the data, per the README:

# Warning: the data is big (~18 GB)!
# Place it somewhere appropriate and change the symlink target path.
wget https://dl.fbaipublicfiles.com/clevr/CLEVR_v1.0.zip
unzip CLEVR_v1.0

cd n2nmn
ln -f -s CLEVR_v1.0 exp_clevr/clevr-dataset

# Preprocessing.
# This part will also consume some 30 GB of disk space.
./exp_clevr/tfmodel/vgg_net/download_vgg_net.sh  # VGG-16 converted to TF

cd ./exp_clevr/data/
python extract_visual_features_vgg_pool5.py  # feature extraction
python get_ground_truth_layout.py  # construct expert policy
python build_clevr_imdb.py  # build image collections
cd ../../

Run microbenchmark:

python exp_clevr/train_clevr_scratch.py

Results

• GPU activity: ~50%

It prints messages about slow I/O, which it does when its prefetch queue is empty.

PINN

An implementation of RobustFill.

Instructions

Install requirements:

conda create --name pinn
conda activate pinn
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

Run microbenchmark:

cd pinn
python test_robustfill.py

Results

• GPU activity: ~70%
• CPU activity: ~100%

Attention Is All You Need

Publication: Attention Is All You Need

Instructions

Install requirements:

# Haven't totally worked out dependencies yet.
# Possibly needs additional CUDA libraries in LD_LIBRARY_PATH.
conda create --name t2t
conda activate t2t
conda install pip
python -m pip install tensorflow-gpu==1.15.0

cd tensor2tensor
python setup.py install

Generate data:

mkdir -p t2t_data t2t_datagen t2t_train
t2t-datagen \
  --data_dir=t2t_data \
  --tmp_dir=t2t_datagen \
  --problem=translate_ende_wmt32k

Run microbenchmark:

t2t-trainer \
  --data_dir=t2t_data \
  --problem=translate_ende_wmt32k \
  --model=transformer \
  --hparams_set=transformer_base_single_gpu \
  --output_dir=t2t_train

Results

• GPU activity: ~90%
• CPU activity: ~110%

meta_nn_scan:

Instructions

Install requirements:

conda create --name meta_nn_scan
conda activate meta_nn_scan
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

Run microbenchmark:

cd meta_nn_scan/meta_net
python pretrain.py --type 'miniscanRBbase' --fn_out_model 'ICML_miniscan_final' \
  --batchsize 128 --episode_type 'rules_gen' --num_pretrain_episodes 100000 \
  --parallel 2

Results

• GPU activity: ~5%, with spikes to ~20%.