Colors in Context

Code and supplementary material for

• Colors in Context: A Pragmatic Neural Model for Grounded Language Understanding (TACL 2017)
• Generating Bilingual Pragmatic Color References (NAACL 2018)

See http://cocolab.stanford.edu/datasets/colors.html for quick links to the datasets released with these papers.

Dependencies

You'll first need Python 2.7. Creating and activating a new virtualenv or Anaconda environment is recommended. Then run this script to download data and Python package dependencies:

./dependencies

The dependencies script is reasonably simple, so if this fails, it should be possible to look at the script and manually perform the actions it specifies.

This code is written to be run on a Linux system; we've also tested it on Mac OS X (but see "Troubleshooting": missing g++ will cause the program to run impossibly slowly). The code is unlikely to run on Windows, but you're welcome to try.

Running experiments

Bilingual Color References

To re-run pre-trained monolingual and bilingual agents from Table 1, you can use the following commands:

python run_experiment.py --config models/en_l1.config.json        # (en: en)
python run_experiment.py --config models/biling_l1_en.config.json # (en: en+zh)
python run_experiment.py --config models/zh_l1.config.json        # (zh: zh)
python run_experiment.py --config models/biling_l1_zh.config.json # (zh: en+zh)

Look for eval.accuracy.mean in the output to compare with Table 2. This command uses dev set results by default; add --data_source filtered_test to reproduce the test set results instead.

Colors in Context

To re-run the base listener L0 from the paper (Table 3) with pre-trained models, use:

python run_experiment.py --config models/l0.config.json

Re-running the other experiments requires first constructing the "quickload" file for the base speaker:

python quickpickle.py --config models/s0.config.json

Then run the Lb model (again, use --data_source filtered_test to run on the test set):

python run_experiment.py --config models/lb.config.json

This is required before you can run all of the other models. It is also the slowest part, and can take several days on a CPU; using a properly-configured GPU usually takes about 19 hours and can be used by passing --device gpu0 to both the quickpickle.py and run_experiment.py commands above. See

http://deeplearning.net/software/theano/tutorial/using_gpu.html

for necessary configuration.

The output of this file is a 230MB file, runs/lb/grids.0.jsons.gz, which contains half of the information needed to run all of the other experiments without rerunning the model. The other half is scores of the sampled utterances using the S0 model. To generate these (~6 hours on CPU, usually less than 1 hour on GPU):

python s0_score.py --config models/s0.config.json \
                   --grids_file runs/lb/grids.0.jsons.gz

Once this is done, the remaining experiments should be very fast to run. They use a separate script that loads the grids files from the other runs:

python blending.py --config models/l1.config.json
python blending.py --config models/l2.config.json
python blending.py --config models/la.config.json
python blending.py --config models/le.config.json

The results of the experiment, including predictions and log-likelihood scores, will all be logged to the directory

runs/lb

To retrain any models for either paper from scratch, add --load '' to the run_experiment.py command. Like the Lb step, GPU is recommended here; for training, the difference in speed can be dramatic (2 hours vs. days).

Troubleshooting

• Error messages of the form

  error: argument --...: invalid int value: '<pyhocon.config_tree.NoneValue object at ...>'

  should be solved by making sure you're using pyhocon version 0.3.18; if this doesn't work, supplying a number for the argument should fix it. We've seen this with the arguments --train_size, --test_size, and --direct_min_score; to fix these, add:

  --train_size 10000000 --test_size 10000000 --direct_min_score 9999

• A warning message of the form

  WARNING (theano.configdefaults): g++ not detected ! Theano will be unable to execute optimized C-implementations (for both CPU and GPU) and will default to Python implementations. Performance will be severely degraded. To remove this warning, set Theano flags cxx to an empty string.

  should be heeded. Otherwise even just running prediction will take a very long time (days). Check whether you can run g++ from a terminal, or try changing the Theano cxx flag (in ~/.theanorc) to point to an alternative C++ compiler on the system.

• If retrying a run after a previous error, you'll need to add the option --overwrite (or specify a different output directory with --run_dir DIR). The program will remind you of this if you forget.

• Very low accuracies (dev.accuracy.mean < 0.5) could indicate incompatible changes in the version of Lasagne or Theano (we've seen this with Lasagne 0.1). We've reproduced our main results using the development versions of Theano and Lasagne as of June 2, 2016:

  • https://github.com/Theano/Theano/tree/0693ce052725a15b502068a1490b0637216feb00
  • https://github.com/Lasagne/Lasagne/tree/8fe645d28b66f991d547e9b6a314251b8e84446a