ara_seminar_research

In this repo we will document the reproducibility issues of the paper Aligning Actions Across Recipe Graphs.

Starting Point

To reproduce the paper's claimed results, please start with creating the exactly same environment the paper used:

conda create -n ada python==3.7
conda activate ada
pip install -r requirements.txt

For compability issues, please check the next section.

Issues Part

Issues with Aligment_Model/requirements.txt

The problem:

They provide package versions that are not compatible with eachother. Pandas 1.2.3 is not compatible with Python 3.7

The fix:

Just download Pandas 1.1

Issues with CUDA compilation:

CUDA would complain since the original paper is using a super old version of torch==1.7.1 and the new generation of GPUs can not be compiled with old version torch, this will cause torch.device("cuda") unusable.
To solve this error, please run the following commands:

pip uninstall torch
pip3 install torch==1.7.1  --force-reinstall  --extra-index-url https://download.pytorch.org/whl/cu110

Data

To get started run the following commands to get the data:

git clone https://github.com/MikeySaw/ara_seminar_research
git clone https://github.com/interactive-cookbook/alignment-models
mv alignment-models/data/ ara_seminar_research/

Training

To train the model, run the following command from root directory:

python train.py [model_name] --embedding_name [embedding_name] --fold [fold]

where [model_name] could be one of the following:

Sequence : Sequential Ordering of Alignments
Cosine_similarity : Cosine model (Baseline)
Naive : Common Action Pair Heuristics mode (Naive Model)
Alignment-no-feature : Base Alignment model (w/o parent+child nodes)
Alignment-with-feature : Extended Alignment model (with parent+child nodes)

and [embedding_name] could be one of the following:

bert : BERT embeddings (default)
elmo : ELMO embeddings

additionally [fold] is a number between 1 and 10: Here you specify on which fold to run the validation on

Right now the Naive and Cosine-similarity models don't work.

Testing

To test the model, choose the application from the following:

Run the following command from this directory:

python test_best_alignment.py [model_name] --embedding_name [embedding_name]

As output, a prediction file named after the test dish(es) will be created. Here the best alignment computed for each action of the test recipes is saved.

The testing script is not working according to what is described in the paper and the original repo.