texrel

Code for paper "TexRel: a Green Family of Datasets for Emergent Communications on Relations", Perkins, 2021, https://arxiv.org/abs/2105.12804

Pre-requisites

• have python 3.7
• have run the following:

pip install -e .
pip install -r requirements.txt

Download the data

• Download and untar https://public-dataset-model-store.awsdev.asapp.com/hp/papers/texrel/texrel.tar

To create the datasets:

If running on a single machine:

python texrel/create_collection.py --ref macdebug --num-train 32 --num-val-same 32 --num-val-new 32 --num-test-same 32 --num-test-new 32
python texrel/create_collection.py --ref ords064sc9 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 0,2 --num-colors 9 --num-shapes 9
python texrel/create_collection.py --ref ords055sc4 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 4 --num-shapes 4
python texrel/create_collection.py --ref ords056sc3 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 3 --num-shapes 3
python texrel/create_collection.py --ref ords057sc5 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 5 --num-shapes 5
python texrel/create_collection.py --ref ords058sc6 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 6 --num-shapes 6
python texrel/create_collection.py --ref ords059sc7 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 7 --num-shapes 7
python texrel/create_collection.py --ref ords060sc8 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 8 --num-shapes 8
python texrel/create_collection.py --ref ords061sc9 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 9 --num-shapes 9
python texrel/create_collection.py --ref ords062 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --hypothesis-generators Relations --available-preps LeftOf

If you have access to a job submission system, that can distribute jobs over multiple machines, then, given you provide a script ulfs_submit.py, in the PATH, which can be called like -r [reference] [... other args you can ignore] -- [script path] [script args], and which will run in each job [script path] --ds-ref [reference] [script args] then you can run instead:

python texrel/create_collection.py --ref macdebug --num-train 32 --num-val-same 32 --num-val-new 32 --num-test-same 32 --num-test-new 32
python texrel/create_collection_submit.py --ref ords064sc9 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 0,2 --num-colors 9 --num-shapes 9
python texrel/create_collection_submit.py --ref ords055sc4 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 4 --num-shapes 4
python texrel/create_collection_submit.py --ref ords056sc3 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 3 --num-shapes 3
python texrel/create_collection_submit.py --ref ords057sc5 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 5 --num-shapes 5
python texrel/create_collection_submit.py --ref ords058sc6 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 6 --num-shapes 6
python texrel/create_collection_submit.py --ref ords059sc7 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 7 --num-shapes 7
python texrel/create_collection_submit.py --ref ords060sc8 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 8 --num-shapes 8
python texrel/create_collection_submit.py --ref ords061sc9 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --num-colors 9 --num-shapes 9
python texrel/create_collection_submit.py --ref ords062 --inner-train-pos 64 --inner-train-neg 64 --inner-test-pos 64 --inner-test-neg 64 --num-distractors 2 --hypothesis-generators Relations --available-preps LeftOf

Each call to create_collection_submit.py will launch one job for each of the 10 task types, i.e. Colors1, Colors2, Shapes1, etc. Note that for bests results you should provide also a script ulfs_logs.sh in the PATH which can be called like ulfs_logs.sh [reference], and which will output the stdout of the launched job corresponding to reference [reference]. However, if you do not supply this, this will not prevent the jobs being launched.

Dataset md5sums

md5sums for each datafile can be found at md5sums.txt

To use the datasets in your code

• you will need to import texrel.dataset_runtime, and create an instance of texrel.dataset_runtime.TexRelDataset

To run the experiments from the paper

Comparison with Shapeworld

Pre-requisites:

• clone the emergent-communications branch of the LSL repo fork at https://github.com/asappresearch/lsl/tree/emergent-communications, from the parent folder of the texrel repo. That is after the fork, should have a folder containing both this texrel repo, and the forked lsl repo. Then from within the texrel folder do:

python ref_task/runners/run_shapeworld_texrel_comp.py --ref trbase043 --early-stop-metric val_same_acc --batch-size 32 --seed-base 123
python ref_task/runners/run_shapeworld_texrel_comp.py --ref trbase044 --early-stop-metric val_same_acc --batch-size 32 --seed-base 124
python ref_task/runners/run_shapeworld_texrel_comp.py --ref trbase045 --early-stop-metric val_same_acc --batch-size 32 --seed-base 125
python ref_task/runners/run_shapeworld_texrel_comp.py --ref trbase046 --early-stop-metric val_same_acc --batch-size 32 --seed-base 126
python ref_task/runners/run_shapeworld_texrel_comp.py --ref trbase047 --early-stop-metric val_same_acc --batch-size 32 --seed-base 127

Comparison of sender architectures

python ref_task/runners/run_arch_send_comparison.py --ref trs009 --max-mins 5 --batch-size 32 --ds-collection ords064sc9 --seed 123
python ref_task/runners/run_arch_send_comparison.py --ref trs010 --max-mins 5 --batch-size 32 --ds-collection ords064sc9 --seed 124
python ref_task/runners/run_arch_send_comparison.py --ref trs011 --max-mins 5 --batch-size 32 --ds-collection ords064sc9 --seed 125
python ref_task/runners/run_arch_send_comparison.py --ref trs012 --max-mins 5 --batch-size 32 --ds-collection ords064sc9 --seed 126
python ref_task/runners/run_arch_send_comparison.py --ref trs013 --max-mins 5 --batch-size 32 --ds-collection ords064sc9 --seed 127

Comparison of receiver architectures

python ref_task/runners/run_arch_recv_comparison.py --ref trr010 --max-mins 5 --batch-size 32 --ds-collection ords064sc9 --seed 123
python ref_task/runners/run_arch_recv_comparison.py --ref trr011 --max-mins 5 --batch-size 32 --ds-collection ords064sc9 --seed 124
python ref_task/runners/run_arch_recv_comparison.py --ref trr012 --max-mins 5 --batch-size 32 --ds-collection ords064sc9 --seed 125
python ref_task/runners/run_arch_recv_comparison.py --ref trr013 --max-mins 5 --batch-size 32 --ds-collection ords064sc9 --seed 126
python ref_task/runners/run_arch_recv_comparison.py --ref trr014 --max-mins 5 --batch-size 32 --ds-collection ords064sc9 --seed 127

Comparison of end-to-end architectures

python ref_task/runners/run_e2e_comparison.py --ref tree043 --early-stop-metric val_same_acc --batch-size 32 --sampler-model Gumbel --max-mins -1
python ref_task/runners/run_e2e_comparison.py --ref tree044 --early-stop-metric val_same_acc --batch-size 32 --sampler-model Gumbel --max-mins -1 --seed-base 124
python ref_task/runners/run_e2e_comparison.py --ref tree045 --early-stop-metric val_same_acc --batch-size 32 --sampler-model Gumbel --max-mins -1 --seed-base 125
python ref_task/runners/run_e2e_comparison.py --ref tree046 --early-stop-metric val_same_acc --batch-size 32 --sampler-model Gumbel --max-mins -1 --seed-base 126
python ref_task/runners/run_e2e_comparison.py --ref tree047 --early-stop-metric val_same_acc --batch-size 32 --sampler-model Gumbel --max-mins -1 --seed-base 127

Effect of multi-task training

python ref_task/runners/run_multitask_learning.py --ref trmt018 --batch-size 32 --sampler-model Gumbel --early-stop-metric val_same_acc --render-every-seconds -1 --render-every-steps 300 --seed-base 123
python ref_task/runners/run_multitask_learning.py --ref trmt019 --batch-size 32 --sampler-model Gumbel --early-stop-metric val_same_acc --render-every-seconds -1 --render-every-steps 300 --seed-base 124
python ref_task/runners/run_multitask_learning.py --ref trmt020 --batch-size 32 --sampler-model Gumbel --early-stop-metric val_same_acc --render-every-seconds -1 --render-every-steps 300 --seed-base 125
python ref_task/runners/run_multitask_learning.py --ref trmt021 --batch-size 32 --sampler-model Gumbel --early-stop-metric val_same_acc --render-every-seconds -1 --render-every-steps 300 --seed-base 126
python ref_task/runners/run_multitask_learning.py --ref trmt022 --batch-size 32 --sampler-model Gumbel --early-stop-metric val_same_acc --render-every-seconds -1 --render-every-steps 300 --seed-base 127

Effect of number of attributes, and number of attribute values

python ref_task/runners/run_numatts_numvalues.py --ref traa007 --sampler-model Gumbel --batch-size 32 --max-steps 5000 --seed 123
python ref_task/runners/run_numatts_numvalues.py --ref traa008 --sampler-model Gumbel --batch-size 32 --max-steps 5000 --seed 124
python ref_task/runners/run_numatts_numvalues.py --ref traa009 --sampler-model Gumbel --batch-size 32 --max-steps 5000 --seed 125
python ref_task/runners/run_numatts_numvalues.py --ref traa010 --sampler-model Gumbel --batch-size 32 --max-steps 5000 --seed 126
python ref_task/runners/run_numatts_numvalues.py --ref traa011 --sampler-model Gumbel --batch-size 32 --max-steps 5000 --seed 127

Reproducing section 7 of TRE paper

python ref_task/runners/run_measuring_comp_section7.py --ref trmc015 --max-steps 3000 --ds-collection 128leftonly --utt-len 5 --vocab-size 26 --batch-size 32 --sampler-model Gumbel

To reduce the results over the seeds, generate results tables, etc

python ref_task/analysis/texrel/reduce_vs_shapeworld.py --out-ref trred013 --in-refs trbase038 trbase039 trbase040 trbase041 trbase042
python ref_task/analysis/texrel/reduce_send_recv.py --out-ref trsr005 --in-refs trs009 trs010 trs011 trs012 --direction send
python ref_task/analysis/texrel/reduce_send_recv.py --direction recv --in-refs trr010 trr011 trr012 trr013 trr014 --out-ref trs006
python ref_task/analysis/texrel/reduce_e2e.py --out-ref trred015 --in-refs tree043 tree044 tree045 tree046 tree047
python ref_task/analysis/texrel/reduce_multitask.py --out-ref trred014 --in-refs trmt018 trmt019 trmt020 trmt021 trmt021
python ref_task/analysis/texrel/reduce_varatts.py --in-csvs traa00[7-9].csv traa01[0-1].csv --out-csv-long varatts_long.csv --out-csv-wide varatts_wide.csv
# following is not actually a reduction, just adds `gen_err` and `ptre` columns:
python ref_task/analysis/texrel/section7_comp_add_gen_err.py --in-csv trmc015.csv --out-csv trmc015-for-latex.csv  

To run unit-tests

pytest -v .

LSL repo fork

The code changes to LSL, in order to enable use the study of emergent communications using it, will be made available at https://github.com/asappresearch/lsl/tree/emergent-communications. To run this code, please see the section above. Or else, to run standalone, you can do e.g.:

git clone git@github.com:asappresearch/lsl.git -b emergent-communications
cd lsl/shapeworld
python lsl/train.py --cuda --infer_hyp --hypo_lambda 1.0 --batch_size 100 --seed 123 --e2e-emergent-communications --data_dir /gfs/nlp/hp/data exp/l3 --backbone conv4 --epochs 50

Citing

If you find the TexRel code useful, please cite:

@misc{perkins2021texrel,
      title={TexRel: a Green Family of Datasets for Emergent Communications on Relations}, 
      author={Hugh Perkins},
      year={2021},
      eprint={2105.12804},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}