Emergent Communication of Generalizations

https://arxiv.org/abs/2106.02668

NeurIPS 2021

Setup

• Download and process birds (CUB) data here, then unzip into /data/cub directory (i.e. the filepath should be data/cub/CUB_200_2011/*), then run python save_cub_np.py in the /data directory to save cub images to easily accessible npz format.
• Download and process ShapeWorld data: use data/download_shapeworld.sh. There are 3 datasets:
  • shapeworld: 20k games over the 312 conjunctive concepts
  • shapeworld_ref: 20k games over the 30 conjunctive concepts possible for reference games only
  • shapeworld_all: 20k games over the 312 conjunctive concepts, but no compositional split.

Note that running concept/setref game agents loads the shapeworld ref dataset (and vice versa) because we do zero shot eval, so download everything.

Code used for generating the ShapeWorld data is located here.

Running experiments

Quickstart

./run_cub.sh and ./run_sw.sh contain ready-to-go commands for running reference, set reference, and concept agents for birds and ShapeWorld, respectively. See those scripts for more details. If you need more info, read on for the basic workflow.

1. Train model

Run python code/train.py --cuda --name NAME --dataset DATASET where NAME is an experiment name (results saved in exp/NAME) and DATASET is the dataset (data/cub, data/shapeworld, or data/shapeworld_ref). To specify the type of game use the following additional flags:

• --percent_novel 1.0: runs concept game (i.e. percent novel indicates what % of images are novel to the student; note you can try values in between if you'd like)
• --percent_novel 0.0: runs setref game
• --percent_novel 0.0 --reference_game: runs reference game. For ShapeWorld, be sure to pass in the 30-concept reference game dataset shapeworld_ref, not the standard 312-concept dataset shapeworld!

Additional flags relevant for experiments:

• --max_lang_length: maximum message length. This includes sos/eos token, so the true length is this value minus 2
• --vocab_size: vocab size of the agents.
• --n_examples: number of examples given to agents
• --uniform_weight: add uniform noise to gumbel softmax exploration policy
• --wandb: activate wandb logging (run wandb init yourself)

There are other options documented in code/io_util.py.

Metrics

Metrics are logged into exp/NAME/metrics.csv and logged to wandb (if --wandb is enabled). The relevant ones are:

• train_acc:
• test. For shapeworld, this metric is split into {test,val}_acc and {test,val}_same_acc to denote unseen and seen splits, respectively, where {test,val}_avg_acc averages the two.
• {train,test}_langts/{train,test}_ts: edit-distance based topographic similarity. For Birds (cub), the metric is ts; for ShapeWorld the metric is langts.
• {train,test}_hausdorff: Hausdorff distance based topographic similarity.

There are many other metrics, most of which should be reasonably intuitive, though contact authors for clarifications.

There are also all of the above metrics split by game type (we eval ref agents on setref, concept, etc).

Mutual information and entropy are measured later (see below).

2. Sample language from model

The above command produces a metrics.csv with most metrics, but I measure entropy and AMI at the end by sampling a bunch of language from the model and analyzing that corpus. To do so, run

# (no --cuda flag needed; will use whatever flag was set at train time)
python code/sample.py exp/NAME

which by default samples 200k messages from a trained model into exp/NAME/sampled_lang.csv and some summary statistics into exp/NAME/sampled_stats.json.

Now, if you just want the information theoretic systematicity metrics, for both Birds and ShapeWorld run python code/acre.py exp/NAME/sampled_lang.csv --dataset DATASET --cuda --stats_only which does not run ACRe, but rather just dumps some summary statistics:

• exp/NAME/sampled_lang_overall_stats.json: this contains entropy, unnormalized mutual information, and adjusted mutual information
• exp/NAME/sampled_lang_stats.csv: this is a list of utterances generated for each concept, with their counts. Also entropy information. This can be used to plot the sunburst (i.e. nested pie) plots in the paper. See "4. Visualizing Model Outputs"

Again, we haven't actually run ACRe. If you want to run ACRe, read on:

3. Train ACRe

If you actually want to train an ACRe model you should train your model with the shapeworld_all dataset, which doesn't involve the compositional split (though you can still do ACRe analysis on models trained normally).

Run ACRe without the --stats_only flag. Rather, run

python code/acre.py exp/NAME/sampled_lang.csv --dataset DATASET --cuda

which trains an ACRe model to reconstruct the agent language according to the concepts of DATASET. This prints out some top1 acc/loss metrics and the following files:

• exp/NAME/sampled_lang_{train,test}_acre_metrics.csv: overall loss/top1 acc for ACRe reconstruction compared to the ground truth language only (i.e. not evaluating a listener model yet), as well as these metrics broken down by concept
• exp/NAME/sampled_lang_{train,test}_sampled_lang.pt: Contains ground truth model language for both train/test ACRe splits, as well as ACRe reconstructions. This gets used to evaluate a listener in the next section.
• exp/NAME/acre_split.json: The split of train/test concepts used for ACRe.

4. Evaluate ACRe on Listener

Run

python code/eval_zero_shot.py exp/NAME --cuda.

which evaluates across --epochs epochs (default 5), categorizing concepts by whether they belong to the ACRe train or test split, and evaluates several types of language on the listener:

• ground_truth_1: the model lang located in exp/NAME/sampled_lang_{train,test}_sampled_lang.pt.
• same_concept: language sampled from other model utterances from the same concept
• acre: ACRe reconstructed language.
• random: random language uniformly sampled from the possible set of utterances (not reported in paper; worse than any_concept)
• any_concept: random language sampled from utterances from any concept (the random baseline in the paper)
• closest_concept: language sampled from utterances for the "closest" concept as measured by edit distance
• ground_truth_2: (sanity check) re-sample language from the teacher; should be close to ground_truth_1 performance.

These results are saved into

• exp/NAME/zero_shot_{train,test}.json: BLEU-1 and listener acc aggregated across all concepts, and for each concept individually
• exp/NAME/zero_shot_lang_type_stats.csv: a lang stats file similar to exp/NAME/sampled_lang_stats.csv described above, which can be used to visualize outputs for the various language distributions as described in the next section.

5. Visualizing model outputs

This requires R and the sunburstR package, as well as a generated sampled_lang_stats.csv which is produced by acre.py (just the --stats_only flag will do). Then an example usage is located in lines 425--441 of analysis/analysis.Rmd.

6. Evaluating across different games

Accuracy and topographic similarity metrics are evaluated zero-shot across different games in the main train script, though entropy/AMI metrics aren't collected. To obtain those, and to get all the results in one place, sample language while using a --force_* flag to force the game to be ref, setref, or concept. This adds a _force_{ref,setref,concept} prefix to every file outputted by sample.py, e.g. sampled_lang_force_ref.csv. For example:

python code/sample.py exp/NAME --force_reference_game
python code/acre.py exp/NAME/sampled_lang_force_ref.csv --dataset data/shapeworld_ref

which now produces exp/NAME/sampled_stats_force_ref.json, exp/NAME/sampled_lang_force_ref_overall_stats.json, exp/NAME/sampled_lang_force_ref_stats.csv, etc.

If you're analyzing ShapeWorld, remember to specify the right dataset - either ref, or setref/concept - when printing summary statistics via acre.py.

Dependencies

This code was tested with python 3.8 and torch==1.8.1. A specific environments file is located in requirements.txt, but other common package versions are likely to be compatible as well.