This repository contains the code to run the experiments presented in the paper Transductive Visual Verb Sense Disambiguation which will be published on WACV2021 (link still not available).

The project is runs on Python 3.6 and relies on the following packages:

• pytorch 1.0.0
• torchvision 0.2.1
• pandas 0.24.2
• numpy 1.17.3

Required packages can be easily installed using the conda YAML file:

conda env create --file conda_env.yml

The data features and required labels can be downloaded from here. The downloaded folder must be called "data" and must be placed into the project folder.

Running the experiments script without any optional parameter will use default ones, running the transduction on VerSe dataset using as input a data-point and a verb, as described in the paper. The number of labelled data-points per class will change from 1 to 20.

python -m run_experiments

The annotations used are the ones from COCO (GOLD).

Optional parameters can be specified:

usage: run_experiments.py [-h] [-G | -P] [-m MAX_LABELS] [-i ITERATIONS] [-a]

optional arguments:
  -h, --help            show this help message and exit
  -G, --gold_captions   Use GOLD captions from COCO.
  -P, --pred_captions   use PRED captions extracted with NeuralBabyTalk.
  -m MAX_LABELS, --max_labels MAX_LABELS
                        The maximum number of labeled data-points to use for
                        each sense.
  -i ITERATIONS, --iterations ITERATIONS
                        The number of Replicator Dynamics iterations to be
                        run.
  -a, --all_senses      Ignore input verb, run inference on the senses of all
                        verbs for each data point.