Home

Awesome

Explainable Neural Computation via Stack Neural Module Networks

This repository contains the code for the following paper:

@inproceedings{hu2018explainable,
  title={Explainable Neural Computation via Stack Neural Module Networks},
  author={Hu, Ronghang and Andreas, Jacob and Darrell, Trevor and Saenko, Kate},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2018}
}

Project Page: http://ronghanghu.com/snmn

Installation

  1. Install Python 3 (Anaconda recommended: https://www.continuum.io/downloads).
  2. Install TensorFlow (we used TensorFlow 1.5.0 in our experiments):
    pip install tensorflow-gpu (or pip install tensorflow-gpu==1.5.0 to install TensorFlow 1.5.0)
  3. Download this repository or clone with Git, and then enter the root directory of the repository:
    git clone https://github.com/ronghanghu/snmn.git && cd snmn

Train and evaluate on the CLEVR (and CLEVR-Ref) dataset

Note (08/04/2019): there was previously an error in the released code -- the gradient clipping was missing in the released version, causing training to be unstable (especially for VQAv1 and VQAv2). This error has been fixed now.

Download and preprocess the data

  1. Download the CLEVR dataset from http://cs.stanford.edu/people/jcjohns/clevr/, and symbol link it to exp_clevr_snmn/clevr_dataset. After this step, the file structure should look like
exp_clevr_snmn/clevr_dataset/
  images/
    train/
      CLEVR_train_000000.png
      ...
    val/
    test/
  questions/
    CLEVR_train_questions.json
    CLEVR_val_questions.json
    CLEVR_test_questions.json
  ...

(Optional) If you want to run any experiments on the CLEVR-Ref dataset for the referential expression grounding task, you can download it from here, and symbol link it to exp_clevr_snmn/clevr_loc_dataset. After this step, the file structure should look like

exp_clevr_snmn/clevr_loc_dataset/
  images/
    loc_train/
      CLEVR_loc_train_000000.png
      ...
    loc_val/
    loc_test/
  questions/
    CLEVR_loc_train_questions.json
    CLEVR_loc_val_questions.json
    CLEVR_loc_test_questions.json
  ...
  1. Extract visual features from the images and store them on the disk. In our experiments, we extract visual features using ResNet-101 C4 block. Then, construct the "expert layout" from ground-truth functional programs, and build image collections (imdb) for CLEVR (and CLEVR-Ref). These procedures can be down as follows.
./exp_clevr_snmn/tfmodel/resnet/download_resnet_v1_101.sh  # download ResNet-101

cd ./exp_clevr_snmn/data/
python extract_resnet101_c4.py  # feature extraction
python get_ground_truth_layout.py  # construct expert policy
python build_clevr_imdb.py  # build image collections
cd ../../

# (Optional, if you want to run on the CLEVR-Ref dataset)
cd ./exp_clevr_snmn/data/
python extract_resnet101_c4_loc.py  # feature extraction
python get_ground_truth_layout_loc.py  # construct expert policy
python build_clevr_imdb_loc.py  # build image collections
cd ../../

Training

  1. Add the root of this repository to PYTHONPATH: export PYTHONPATH=.:$PYTHONPATH

  2. Train on the CLEVR dataset for VQA:

    • with ground-truth layout
      python exp_clevr_snmn/train_net_vqa.py --cfg exp_clevr_snmn/cfgs/vqa_gt_layout.yaml
    • without ground-truth layout
      python exp_clevr_snmn/train_net_vqa.py --cfg exp_clevr_snmn/cfgs/vqa_scratch.yaml
  3. (Optional) Train on the CLEVR-Ref dataset for the REF task:

    • with ground-truth layout
      python exp_clevr_snmn/train_net_loc.py --cfg exp_clevr_snmn/cfgs/loc_gt_layout.yaml
    • without ground-truth layout
      python exp_clevr_snmn/train_net_loc.py --cfg exp_clevr_snmn/cfgs/loc_scratch.yaml
  4. (Optional) Train jointly on the CLEVR and CLEVR-Ref datasets for VQA and REF tasks:

    • with ground-truth layout
      python exp_clevr_snmn/train_net_joint.py --cfg exp_clevr_snmn/cfgs/joint_gt_layout.yaml
    • without ground-truth layout
      python exp_clevr_snmn/train_net_joint.py --cfg exp_clevr_snmn/cfgs/joint_scratch.yaml

Note:

Test

  1. Add the root of this repository to PYTHONPATH: export PYTHONPATH=.:$PYTHONPATH

  2. Evaluate on the CLEVR dataset for the VQA task:
    python exp_clevr_snmn/test_net_vqa.py --cfg exp_clevr_snmn/cfgs/{exp_name}.yaml TEST.ITER 200000
    where {exp_name} should be one of vqa_gt_layout, vqa_scratch, joint_gt_layout and joint_scratch.
    Expected accuracy: 96.6% for vqa_gt_layout, 93.0% for vqa_scratch, 96.5% for joint_gt_layout, 93.9% for joint_scratch. Note:

    • The above evaluation script will print out the accuracy (only for val split) and also save it under exp_clevr_snmn/results/{exp_name}/. It will also save a prediction output file in this directory.
    • The above evaluation script will generate 100 visualizations by default, and save it under exp_clevr_snmn/results/{exp_name}/. You may change the number of visualizations with TEST.NUM_VIS parameter (e.g. appending TEST.NUM_VIS 400 to the commands above to generate 400 visualizations).
    • By default, the above script evaluates on the validation split of CLEVR. To evaluate on the test split, append TEST.SPLIT_VQA test to the command above. As there is no ground-truth answers for test split in the downloaded CLEVR data, the displayed accuracy will be zero on the test split. You may email the prediction outputs in exp_clevr_snmn/results/{exp_name}/ to the CLEVR dataset authors for the test split accuracy.
    • By default, the above script uses GPU 0. To run on a different GPU, append GPU_ID parameter to the commands above (e.g. appending GPU_ID 2 to use GPU 2).
  3. (Optional) Evaluate on the CLEVR-Ref dataset for the REF task:
    python exp_clevr_snmn/test_net_loc.py --cfg exp_clevr_snmn/cfgs/{exp_name}.yaml TEST.ITER 200000
    where {exp_name} should be one of loc_gt_layout, loc_scratch, joint_gt_layout and joint_scratch.
    Expected accuracy (Precision@1): 96.0% for loc_gt_layout, 93.4% for loc_scratch, 96.2% for joint_gt_layout, 95.4% for joint_scratch. Note:

    • The above evaluation script will print out the accuracy (Precision@1) and also save it under exp_clevr_snmn/results/{exp_name}/.
    • The above evaluation script will generate 100 visualizations by default, and save it under exp_clevr_snmn/results/{exp_name}/. You may change the number of visualizations with TEST.NUM_VIS parameter (e.g. appending TEST.NUM_VIS 400 to the commands above to generate 400 visualizations).
    • By default, the above script evaluates on the validation split of CLEVR-Ref. To evaluate on the test split, append TEST.SPLIT_LOC loc_test to the command above.
    • By default, the above script uses GPU 0. To run on a different GPU, append GPU_ID parameter to the commands above (e.g. appending GPU_ID 2 to use GPU 2).

Train and evaluate on the VQAv1 and VQAv2 datasets

Note (08/04/2019): there was previously an error in the released code -- the gradient clipping was missing in the released version, causing training to be unstable (especially for VQAv1 and VQAv2). This error has been fixed now.

Download and preprocess the data

  1. Download the VQAv1 and VQAv2 dataset annotations from http://www.visualqa.org/download.html, and symbol link them to exp_vqa/vqa_dataset. After this step, the file structure should look like
exp_vqa/vqa_dataset/
  Questions/
    OpenEnded_mscoco_train2014_questions.json
    OpenEnded_mscoco_val2014_questions.json
    OpenEnded_mscoco_test-dev2015_questions.json
    OpenEnded_mscoco_test2015_questions.json
    v2_OpenEnded_mscoco_train2014_questions.json
    v2_OpenEnded_mscoco_val2014_questions.json
    v2_OpenEnded_mscoco_test-dev2015_questions.jso
    v2_OpenEnded_mscoco_test2015_questions.json
  Annotations/
    mscoco_train2014_annotations.json
    mscoco_val2014_annotations.json
    v2_mscoco_train2014_annotations.json
    v2_mscoco_val2014_annotations.json
    v2_mscoco_train2014_complementary_pairs.json
    v2_mscoco_val2014_complementary_pairs.json
  1. Download the COCO images from http://mscoco.org/, and symbol link it to exp_vqa/coco_dataset. After this step, the file structure should look like
exp_vqa/coco_dataset/
  images/
    train2014/
      COCO_train2014_000000000009.jpg
      ...
    val2014/
      COCO_val2014_000000000042.jpg
      ...
    test2015/
      COCO_test2015_000000000001.jpg
      ...
  1. Extract visual features from the images and store them on the disk. In our experiments, we extract visual features using ResNet-152 C5 block. Then, build image collections (imdb) for VQAv1 and VQAv2. These procedures can be done as follows.
./exp_vqa/tfmodel/resnet/download_resnet_v1_152.sh  # Download ResNet-152

cd ./exp_vqa/data/
python extract_resnet152_c5_7x7.py  # feature extraction for all COCO images
python build_vqa_imdb_r152_7x7.py  # build image collections for VQAv1
python build_vqa_imdb_r152_7x7_vqa_v2.py  # build image collections for VQAv2
cd ../../

(Note that this repository already contains the "expert layout" from parsing results using Stanford Parser. They are the same as in N2NMN.)

Pre-trained models

You may skip the training procedure and directly download the pretrained models here for evaluation. The downloaded models should be put under exp_vqa/tfmodel/{exp_name}/.

Training

  1. Add the root of this repository to PYTHONPATH: export PYTHONPATH=.:$PYTHONPATH

  2. Train on the VQAv1 dataset:

    • with ground-truth layout
      python exp_vqa/train_net_vqa.py --cfg exp_vqa/cfgs/vqa_v1_gt_layout.yaml
    • without ground-truth layout
      python exp_vqa/train_net_vqa.py --cfg exp_vqa/cfgs/vqa_v1_scratch.yaml
  3. Train on the VQAv2 dataset:

    • with ground-truth layout
      python exp_vqa/train_net_vqa.py --cfg exp_vqa/cfgs/vqa_v2_gt_layout.yaml
    • without ground-truth layout
      python exp_vqa/train_net_vqa.py --cfg exp_vqa/cfgs/vqa_v2_scratch.yaml

Note:

Test

  1. Add the root of this repository to PYTHONPATH: export PYTHONPATH=.:$PYTHONPATH

  2. Evaluate on the VQAv1 dataset:
    python exp_vqa/test_net_vqa.py --cfg exp_vqa/cfgs/{exp_name}.yaml TEST.ITER 20000
    where {exp_name} should be one of vqa_v1_gt_layout and vqa_v1_scratch. Note:

    • By default, the above script evaluates on the test-dev2015 split of VQAv1. To evaluate on the test2015 split, append TEST.SPLIT_VQA test2015 to the command above.
    • By default, the above script uses GPU 0. To run on a different GPU, append GPU_ID parameter to the commands above (e.g. appending GPU_ID 2 to use GPU 2).
    • The above evaluation script will not print out the accuracy (the displayed accuracy will be zero), but will write the prediction outputs to exp_vqa/eval_outputs/{exp_name}/, which can be uploaded to the evaluation sever (http://www.visualqa.org/roe.html) for evaluation. Expected accuracy: 66.0% for vqa_v1_gt_layout, 65.5% for vqa_v1_scratch.
  3. Evaluate on the VQAv2 dataset:
    python exp_vqa/test_net_vqa.py --cfg exp_vqa/cfgs/{exp_name}.yaml TEST.ITER 40000
    where {exp_name} should be one of vqa_v2_gt_layout and vqa_v2_scratch. Note:

    • By default, the above script uses GPU 0. To run on a different GPU, append GPU_ID parameter to the commands above (e.g. appending GPU_ID 2 to use GPU 2).
    • The above evaluation script will not print out the accuracy (the displayed accuracy will be zero), but will write the prediction outputs to exp_vqa/eval_outputs_vqa_v2/{exp_name}/, which can be uploaded to the evaluation sever (http://www.visualqa.org/roe.html) for evaluation. Expected accuracy: 64.0% for vqa_v2_gt_layout, 64.1% for vqa_v2_scratch.

Acknowledgements

The outline of the configuration code (such as models_clevr_snmn/config.py) is obtained from the Detectron codebase.