In Machine Learning, an important question is "How to fuse two modalities in a same space". For instance, in Visual Question Answering, one must fuse the image and the question embeddings in a same bi-modal space. This multimodal embedding is latter classified to provide the answer.

We introduce a novel module (BLOCK) to fuse two representations together. First, we experimentaly demonstrate that it is better than any available fusion for our tasks. Secondly, we provide a theoritical-grounded analysis around the notion of tensor complexity. For further details, please see our AAAI 2019 paper and poster.

In this repo, we make our BLOCK fusion available via pip install including several powerful fusions from the state-of-the-art (MLB, MUTAN, MCB, MFB, MFH, etc.). Also, we provide pretrained models and all the code needed to reproduce our experiments.

• Installation
  • Python 3 & Anaconda
  • As a standalone project
  • Download datasets
  • As a python library
• Quick start
  • Train a model
  • Evaluate a model
• Reproduce results
  • VRD
  • VQA2
  • TDIUC
• Pretrained models
• Fusions
  • Block
  • LinearSum
  • ConcatMLP
  • MLB
  • Tucker
  • Mutan
  • BlockTucker
  • MFB
  • MFH
  • MCB
• Useful commands
• Citation
• Poster
• Authors
• Acknowledgment

We don't provide support for python 2. We advise you to install python 3 with Anaconda. Then, you can create an environment.

conda create --name block python=3
source activate block
git clone --recursive https://github.com/Cadene/block.bootstrap.pytorch.git
cd block.bootstrap.pytorch
pip install -r requirements.txt

Download annotations, images and features for VRD experiments:

bash block/datasets/scripts/download_vrd.sh

Download annotations, images and features for VQA experiments:

bash block/datasets/scripts/download_vqa2.sh
bash block/datasets/scripts/download_vgenome.sh
bash block/datasets/scripts/download_tdiuc.sh

Note: The features have been extracted from a pretrained Faster-RCNN with caffe. We don't provide the code for pretraining or extracting features for now.

By importing the block python module, you can access every fusions, datasets and models in a simple way:

import torch
from block import fusions
mm = fusions.Block([100,100], 300)
inputs = [torch.randn(10,100), torch.randn(10,100)]
out = mm(inputs) # torch.Size([10,300])
# ...
fusions.LinearSum
fusions.ConcatMLP
fusions.MLB
fusions.Mutan
fusions.Tucker
fusions.BlockTucker
fusions.MFB
fusions.MFH
fusions.MCB
# ...
from block.datasets.vqa2 import VQA2
from block.datasets.tdiuc import TDIUC
from block.datasets.vg import VG
from block.datasets.vrd import VRD
# ...
from block.models.networks.vqa_net import VQANet
from block.models.networks.vrd_net import VRDNet
# ...

To be able to do so, you can use pip:

pip install block.bootstrap.pytorch

Or install from source:

git clone https://github.com/Cadene/block.bootstrap.pytorch.git
python setup.py install

The boostrap/run.py file load the options contained in a yaml file, create the corresponding experiment directory and start the training procedure. For instance, you can train our best model on VRD by running:

python -m bootstrap.run -o block/options/vrd/block.yaml

Then, several files are going to be created in logs/vrd/block:

• options.yaml (copy of options)
• logs.txt (history of print)
• logs.json (batchs and epochs statistics)
• view.html (learning curves)
• ckpt_last_engine.pth.tar (checkpoints of last epoch)
• ckpt_last_model.pth.tar
• ckpt_last_optimizer.pth.tar
• ckpt_best_eval_epoch.predicate.R_50_engine.pth.tar (checkpoints of best epoch)
• ckpt_best_eval_epoch.predicate.R_50_model.pth.tar
• ckpt_best_eval_epoch.predicate.R_50_optimizer.pth.tar

Many options are available in the options directory.

At the end of the training procedure, you can evaluate your model on the testing set. In this example, boostrap/run.py load the options from your experiment directory, resume the best checkpoint on the validation set and start an evaluation on the testing set instead of the validation set while skipping the training set (train_split is empty). Thanks to --misc.logs_name, the logs will be written in the new logs_predicate.txt and logs_predicate.json files, instead of being appended to the logs.txt and logs.json files.

python -m bootstrap.run \
-o logs/vrd/block/options.yaml \
--exp.resume best_eval_epoch.predicate.R_50 \
--dataset.train_split \
--dataset.eval_split test \
--misc.logs_name predicate

1. Train block on trainset with early stopping on valset
2. Evaluate the best checkpoint on testset (Predicate Prediction)
3. Evaluate the best checkpoint on testset (Relationship and Phrase Detection)

python -m bootstrap.run \
-o block/options/vrd/block.yaml \
--exp.dir logs/vrd/block

python -m bootstrap.run \
-o logs/vrd/block/options.yaml \
--dataset.train_split \
--dataset.eval_split test \
--exp.resume best_eval_epoch.predicate.R_50 \
--misc.logs_name predicate

python -m bootstrap.run \
-o logs/vrd/block/options.yaml \
--dataset.train_split \
--dataset.eval_split test \
--dataset.mode rel_phrase \
--model.metric.name vrd_rel_phrase \
--exp.resume best_eval_epoch.predicate.R_50 \
--misc.logs_name rel_phrase

Note: You can copy past the three commands at once in the terminal to run one after each other seamlessly.

Note: Block is not the only option available. You can find several others here.

Note: Learning curves can be viewed in the experiment directy (logs/vrd/block/view.html). An example is available here.

Note: In our article, we report result for a negative sampling ratio of 0.5. Better results in Predicate Prediction can be achieve with a ratio of 0.0. Better results in Phrase Detection and Relationship Detection can be achieve with a ratio of 0.8. You can change the ratio by doing so:

python -m bootstrap.run \
-o block/options/vrd/block.yaml \
--exp.dir logs/vrd/block_ratio,0.0 \
--dataset.neg_ratio 0.0

Finally you can compare experiments on the valset or testset metrics:

python -m block.compare_vrd_val -d \
logs/vrd/block \
logs/vrd/block_tucker \
logs/vrd/mutan \
logs/vrd/mfh \
logs/vrd/mlb

python -m block.compare_vrd_test -d \
logs/vrd/block \
logs/vrd/block_tucker

Example:

## eval_epoch.predicate.R_50
  Place  Method          Score    Epoch
-------  ------------  -------  -------
      1  block         86.3708       13
      2  block_tucker  86.2529        9

## eval_epoch.predicate.R_100
  Place  Method          Score    Epoch
-------  ------------  -------  -------
      1  block         92.4588       13
      2  block_tucker  91.5816        9

## eval_epoch.phrase.R_50
  Place  Method          Score    Epoch
-------  ------------  -------  -------
      1  block         25.4779       13
      2  block_tucker  23.7759        9

## eval_epoch.phrase.R_100
  Place  Method          Score    Epoch
-------  ------------  -------  -------
      1  block         29.7198       13
      2  block_tucker  27.9131        9

## eval_epoch.rel.R_50
  Place  Method          Score    Epoch
-------  ------------  -------  -------
      1  block         18.0806       13
      2  block_tucker  17.0856        9

## eval_epoch.rel.R_100
  Place  Method          Score    Epoch
-------  ------------  -------  -------
      1  block         21.1181       13
      2  block_tucker  19.7565        9

We use this simple setup to tune our hyperparameters on the valset.

python -m bootstrap.run \
-o block/options/vqa2/block.yaml \
--exp.dir logs/vqa2/block

This heavier setup allows us to train a model on 95% of the concatenation of train and val sets, and to evaluate it on the 5% rest. Then we extract the predictions of our best checkpoint on the testset. Finally, we submit a json file on the EvalAI web site.

python -m bootstrap.run \
-o block/options/vqa2/block.yaml \
--exp.dir logs/vqa2/block_trainval \
--dataset.proc_split trainval

python -m bootstrap.run \
-o logs/vqa2/block_trainval/options.yaml \
--exp.resume best_eval_epoch.accuracy_top1 \
--dataset.train_split \
--dataset.eval_split test \
--misc.logs_name test

Same, but we add pairs from the VisualGenome dataset.

python -m bootstrap.run \
-o block/options/vqa2/block.yaml \
--exp.dir logs/vqa2/block_trainval_vg \
--dataset.proc_split trainval \
--dataset.vg True

python -m bootstrap.run \
-o logs/vqa2/block_trainval_vg/options.yaml \
--exp.resume best_eval_epoch.accuracy_top1 \
--dataset.train_split \
--dataset.eval_split test \
--misc.logs_name test

You can compare experiments by displaying their best metrics on the valset.

python -m block.compare_vqa_val -d logs/vqa2/block logs/vqa2/mutan

It is not possible to automaticaly compute the accuracies on the testset. You need to submit a json file on the EvalAI platform. The evaluation step on the testset creates the json file that contains the prediction of your model on the full testset. For instance: logs/vqa2/block_trainval_vg/results/test/epoch,19/OpenEnded_mscoco_test2015_model_results.json. To get the accuracies on testdev or test sets, you must submit this file.

The full training set is split into a trainset and a valset. At the end of the training, we evaluate our best checkpoint on the testset. The TDIUC metrics are computed and displayed at the end of each epoch. They are also stored in logs.json and logs_test.json.

python -m bootstrap.run \
-o block/options/tdiuc/block.yaml \
--exp.dir logs/tdiuc/block

python -m bootstrap.run \
-o logs/tdiuc/block/options.yaml \
--exp.resume best_eval_epoch.accuracy_top1 \
--dataset.train_split \
--dataset.eval_split test \
--misc.logs_name test

You can compare experiments by displaying their best metrics on the valset or testset.

python -m block.compare_tdiuc_val -d logs/tdiuc/block logs/tdiuc/mutan
python -m block.compare_tdiuc_test -d logs/tdiuc/block logs/tdiuc/mutan

Note: These pretrained models have been trained using the Pytorch 1.0 to make sure that our results are reproducible in this version. We also used a more efficient learning rate scheduling strategy which turned out to give slightly better results.

Download Block:

mkdir -p logs/vrd
cd logs/vrd
wget http://data.lip6.fr/cadene/block/vrd/block.tar.gz
tar -xzvf block.tar.gz

Results python -m block.compare_vrd_test -d logs/vrd/block:

• predicate.R_50: 86.3708
• predicate.R_100: 92.4588
• phrase.R_50: 25.4779
• phrase.R_100: 29.7198
• rel.R_50: 18.0806
• rel.R_100: 21.1181

Download Block train/val:

mkdir -p logs/vqa2
cd logs/vqa2
wget http://data.lip6.fr/cadene/block/vqa2/block.tar.gz
tar -xzvf block.tar.gz

Results val (python -m block.compare_vqa2_val -d logs/vqa2/block):

• overall (oe): 63.6
• accuracy_top1: 54.4254

Download Block train+val/val/test:

mkdir -p logs/vqa2
cd logs/vqa2
wget http://data.lip6.fr/cadene/block/vqa2/block_trainval.tar.gz
tar -xzvf block_trainval.tar.gz

Results test-dev (EvalAI):

• overall: 66.74
• yes/no: 83.73
• number: 46.51
• other: 56.84

Download Block train+val+vg/val/test:

mkdir -p logs/vqa2
cd logs/vqa2
wget http://data.lip6.fr/cadene/block/vqa2/block_trainval_vg.tar.gz
tar -xzvf block_trainval_vg.tar.gz

Results test-dev (EvalAI):

• overall: 67.41
• yes/no: 83.89
• number: 46.22
• other: 58.18

Download Block train+val/val/test:

mkdir -p logs/tdiuc
cd logs/tdiuc
wget http://data.lip6.fr/cadene/block/tdiuc/block_trainval.tar.gz
tar -xzvf block_trainval.tar.gz

Results val (python -m block.compare_tdiuc_val -d logs/tdiuc/block):

• accuracy_top1: 88.0195
• acc_mpt_a: 72.2555
• acc_mpt_h: 59.9484
• acc_mpt_a_norm: 60.9635
• acc_mpt_h_norm: 44.7724

Results test (python -m block.compare_tdiuc_test -d logs/tdiuc/block):

• accuracy_top1: 86.3242
• acc_mpt_a: 72.4447
• acc_mpt_h: 66.15
• acc_mpt_a_norm: 58.5728
• acc_mpt_h_norm: 38.8279

fusion = fusions.Block([100,100], 300)

Parameters:

• input_dims: list containing the dimensions of each input vector
• output_dim: desired output dimension
• mm_dim: dimension of the multi-modal space. Here, it is equal the sum of output dimensions of all the D_c tensors (default: 1600)
• chunks: number of blocks in the block-diagonal tensor. Equal to C in the previous equations (default: 20)
• rank: upper-bound of the rank of mode-3 slice matrices of D_c tensors (default: 15)
• shared: boolean that specifies if we want to share the values of input mono-modal projections (default: False)
• dropout_input: dropout rate right after the input projections (default: 0.)
• dropout_pre_lin: dropout rate just before the output linear (default: 0.)
• dropout_output: dropout rate right after the output linear (default: 0.)
• pos_norm: string that specifies if the signed-square root - l2 normalization should be done on every chunk outputs or on the concatenations of every outputs. Accepted values: 'before_cat' and 'after_cat'. (default: 'before_cat')

Reference: BLOCK: Bilinear Superdiagonal Fusion for Visual Question Answering and Visual Relationship Detection, *Hedi Ben-younes, Rémi Cadene, Nicolas Thome, Matthieu Cord *

code

fusion = fusions.LinearSum([100, 100], 300)

Parameters:

• input_dims: list containing the dimensions of each input vector
• output_dim: desired output dimension
• mm_dim: dimension of the multi-modal space (default: 1200)
• activ_input: name of the activation function that follows mono-modal projections, before the sum (default: relu)
• activ_output: name of the activation function that follows output projection (default: relu)
• normalize: boolean that specifies whether or not we want to apply the signed square root - l2 normalization (default: False)
• dropout_input: dropout rate right after the activ_input (default: 0.)
• dropout_pre_lin: dropout rate just before the output linear (default: 0.)
• dropout_output: dropout rate right after the activ_output (default: 0.)

code

fusion = fusions.ConcatMLP([100, 100], 300)

Parameters:

• input_dims: list containing the dimensions of each input vector
• output_dim: desired output dimension
• dimensions: list of hidden dimensions (default: [500,500])
• activation: stringname of the activation function of the network, applied at each layer but the last (default: 'relu')
• dropout: dropout rate, applied at each layer but the last (default: 0.)

code

fusion = fusions.MLB([100,100], 300)

Parameters:

• input_dims: list containing the dimensions of each input vector
• output_dim: desired output dimension
• mm_dim: dimension of the multi-modal space (default: 1200)
• activ_input: name of the activation function that follows mono-modal projections, before the element-wise product (default: 'relu')
• activ_output: name of the activation function that follows output projection (default: 'relu')
• normalize: boolean that specifies whether or not we want to apply the signed square root - l2 normalization (default: False)
• dropout_input: dropout rate right after the activ_input (default: 0.)
• dropout_pre_lin: dropout rate just before the output linear (default: 0.)
• dropout_output: dropout rate right after the activ_output (default: 0.)

Reference: Hadamard Product for Low-rank Bilinear Pooling, Jin-Hwa Kim, Kyoung-Woon On, Woosang Lim, Jeonghee Kim, Jung-Woo Ha, Byoung-Tak Zhang

code

fusion = fusions.Mutan([100, 100], 300)

Parameters:

• input_dims: list containing the dimensions of each input vector
• output_dim: desired output dimension
• mm_dim: dimension of the multi-modal space. Here, it is equal to the output dimensions of the D tensor (default: 1600)
• rank: upper-bound of the rank of mode-3 slice matrices of the D tensor (default: 15)
• shared: boolean that specifies if we want to share the values of input mono-modal projections (default: False)
• normalize: boolean that specifies whether or not we want to apply the signed square root - l2 normalization (default: False)
• dropout_input: dropout rate right after the input projections (default: 0.)
• dropout_pre_lin: dropout rate just before the output linear (default: 0.)
• dropout_output: dropout rate right after the output linear (default: 0.)

Reference: MUTAN: Multimodal Tucker Fusion for Visual Question Answering, Hedi Ben-younes*, Rémi Cadene*, Nicolas Thome, Matthieu Cord

code

This module correponds to Mutan without the low-rank constraint on third-mode slices of the D tensor.

fusion = fusions.Tucker([100, 100], 300)

Parameters:

• input_dims: list containing the dimensions of each input vector
• output_dim: desired output dimension
• mm_dim: dimension of the multi-modal space. Here, it is equal to the output dimensions of the D tensor (default: 1600)
• shared: boolean that specifies if we want to share the values of input mono-modal projections (default: False)
• normalize: boolean that specifies whether or not we want to apply the signed square root - l2 normalization (default: False)
• dropout_input: dropout rate right after the input projections (default: 0.)
• dropout_pre_lin: dropout rate just before the output linear (default: 0.)
• dropout_output: dropout rate right after the output linear (default: 0.)

Reference: MUTAN: Multimodal Tucker Fusion for Visual Question Answering, Hedi Ben-younes*, Rémi Cadene*, Nicolas Thome, Matthieu Cord

code

This module correponds to Block without the low-rank constraint on third-mode slices of D_c tensors

fusion = fusions.BlockTucker([100,100], 300)

Parameters:

• input_dims: list containing the dimensions of each input vector
• output_dim: desired output dimension
• mm_dim: dimension of the multi-modal space. Here, it is equal the sum of output dimensions of all the D_c tensors (default: 1600)
• chunks: number of blocks in the block-diagonal tensor. Equal to C in the previous equations (default: 20)
• shared: boolean that specifies if we want to share the values of input mono-modal projections (default: False)
• dropout_input: dropout rate right after the input projections (default: 0.)
• dropout_pre_lin: dropout rate just before the output linear (default: 0.)
• dropout_output: dropout rate right after the output linear (default: 0.)
• pos_norm: string that specifies if the signed-square root - l2 normalization should be done on every chunk outputs or on the concatenations of every outputs. Accepted values: 'before_cat' and 'after_cat'. (default: 'before_cat')

Reference: BLOCK: Bilinear Superdiagonal Fusion for Visual Question Answering and Visual Relationship Detection, *Hedi Ben-younes, Rémi Cadene, Nicolas Thome, Matthieu Cord *

code

fusion = fusions.MFB([100,100], 300)

Parameters:

• input_dims: list containing the dimensions of each input vector
• output_dim: desired output dimension
• mm_dim: dimension of the multi-modal space. Here, it is equal to the output dimension of the MFB layer (default: 1200)
• factor: MFB factor (default: 2)
• activ_input: name of the activation function that follows mono-modal projections, before the element-wise product (default: 'relu')
• activ_output: name of the activation function that follows output projection (default: 'relu')
• normalize: boolean that specifies whether or not we want to apply the signed square root - l2 normalization (default: False)
• dropout_input: dropout rate right after the activ_input (default: 0.)
• dropout_pre_lin: dropout rate just before the output linear (default: 0.)
• dropout_output: dropout rate right after the activ_output (default: 0.)

Reference: Multi-modal Factorized Bilinear Pooling with Co-Attention Learning for Visual Question Answering, *Zhou Yu, Jun Yu, Jianping Fan, Dacheng Tao *

code

fusion = fusions.MFH([100,100], 300)

Parameters:

• input_dims: list containing the dimensions of each input vector
• output_dim: desired output dimension
• mm_dim: dimension of the multi-modal space. Here, it is equal to the output dimension of the MFH layer (default: 1200)
• factor: MFB factor (default: 2)
• activ_input: name of the activation function that follows mono-modal projections, before the element-wise product (default: 'relu')
• activ_output: name of the activation function that follows output projection (default: 'relu')
• normalize: boolean that specifies whether or not we want to apply the signed square root - l2 normalization (default: False)
• dropout_input: dropout rate right after the activ_input (default: 0.)
• dropout_pre_lin: dropout rate just before the output linear (default: 0.)
• dropout_output: dropout rate right after the activ_output (default: 0.)

Reference: Beyond Bilinear: Generalized Multi-modal Factorized High-order Pooling for Visual Question Answering, Zhou Yu, Jun Yu, Chenchao Xiang, Jianping Fan, Dacheng Tao

code

/!\ Not available in pytorch 1.0 - Avaiable in pytorch 0.3 and 0.4

fusion = fusions.MCB([100,100], 300)

Parameters:

• input_dims: list containing the dimensions of each input vector
• output_dim: desired output dimension
• mm_dim: dimension of the multi-modal space. Here, it is equal to the output dimension of the MCB layer (default: 16000)
• activ_output: name of the activation function that follows output projection (default: 'relu')
• dropout_output: dropout rate right after the activ_output (default: 0.)

Reference: Multimodal Compact Bilinear Pooling for Visual Question Answering and Visual Grounding, Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, Marcus Rohrbach

code

Instead of creating a view.html file, a tensorboard file will be created:

python -m bootstrap.run -o block/options/vqa2/block.yaml \
--view.name tensorboard

tensorboard --logdir=logs/vqa2

You can use plotly and tensorboard at the same time by updating the yaml file like this one.

For a specific experiment:

CUDA_VISIBLE_DEVICES=0 python -m boostrap.run -o block/options/vqa2/block.yaml

For the current terminal session:

export CUDA_VISIBLE_DEVICES=0

The boostrap.pytorch framework makes it easy to overwrite a hyperparameter. In this example, we run an experiment with a non-default learning rate. Thus, I also overwrite the experiment directory path:

python -m bootstrap.run -o block/options/vqa2/block.yaml \
--optimizer.lr 0.0003 \
--exp.dir logs/vqa2/block_lr,0.0003

If a problem occurs, it is easy to resume the last epoch by specifying the options file from the experiment directory while overwritting the exp.resume option (default is None):

python -m bootstrap.run -o logs/vqa2/block/options.yaml \
--exp.resume last

TODO

TODO

@InProceedings{BenYounes_2019_AAAI,
    author = {Ben-Younes, Hedi and Cadene, Remi and Thome, Nicolas and Cord, Matthieu},
    title = {BLOCK: {B}ilinear {S}uperdiagonal {F}usion for {V}isual {Q}uestion {A}nswering and {V}isual {R}elationship {D}etection},
    booktitle = {The Thirty-Third AAAI Conference on Artificial Intelligence},
    year = {2019},
    url = {http://remicadene.com/pdfs/paper_aaai2019.pdf}
}

This code was made available by Hedi Ben-Younes (Sorbonne-Heuritech), Remi Cadene (Sorbonne), Matthieu Cord (Sorbonne) and Nicolas Thome (CNAM).

Special thanks to the authors of VQA2, TDIUC, VisualGenome and VRD, the datasets used in this research project.