Generationg Visual Explanations

This repository contains code for the following paper:

Hendricks, L.A., Akata, Z., Rohrbach, M., Donahue, J., Schiele, B. and Darrell, T., 2016. Generating Visual Explanations. ECCV 2016.

@article{hendricks2016generating,
  title={Generating Visual Explanations},
  author={Hendricks, Lisa Anne and Akata, Zeynep and Rohrbach, Marcus and Donahue, Jeff and Schiele, Bernt and Darrell, Trevor},
  journal={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2016}
}

This code has been edited extensively (you can see old code on deprecated branch). Hopefully it is easier to use, but please bug me if you run into issues.

Getting Started

1. Please clone my git repo. You will need to use my version of caffe, specifically the "bilinear" branch.
2. Download data using the "download_data.sh" script. This will also preprocess the CUB sentences. All my ECCV 2016 models will be put in "gve_models"

Building the models

All the models are generated using NetSpec. Please build them by running "build_nets.sh". "build_nets.sh" will also generate bash scripts you can use to train models.

Training the models

If you would like to retrain my models, please use the following instructions. Note that all my trained models are in "gve_models". All the training scripts will be built using "build_nets.sh"

1. First train the description model ("./train_description.sh"). The learned hidden units of the description model are used to build a representation for the 200 CUB classes.
2. Run "make_class_embedding.sh" to build the class embeddings
3. Train definition and explanation_label models ("./train_definition.sh", "./train_explanation_label.sh").
4. Train sentence classification model ("./train_caption_classifer.sh"). This is needed for the reinforce loss. I found that using an embedding and LSTM hidden dim of 1000 and dropout of 0.75 worked best.
5. Train the explanation_dis and explanation models ("./train_explanation_dis.sh", "./train_explanation.sh"). These models are fine-tuned from the description and explanation_label model respecitvely. The weighting between the relevance and discriminative loss can impact perforance substantially. I found that loss weights of 80/20 on the relevance/discriminative losses worked best for the explanation-dis model and that loss weights of 110/20 on the relevance/discriminative losses worked best for the explanation model.

Evaluate

Please use the bash scripts eval_*.sh to compute image relevance metrics. To compute class relevance metrics, run "analyze_cider_scores.py". This relies on precomputed CIDEr scores between each generated sentence and reference sentences from each class. You can recompute these using "class_meteor_similarity_metric.py" but this will take > 10 hours.

Please note that I retrained the models since the initial arXiv version of the paper was released so the numbers are slightly different, though the main trends remain the same.