Memory Network

This project is an implementation of the Memory Network (see PAPER).

People : David Panou

Organization : UPMC - Master Data Science

• ✅ Embeddings (inputs & memory )
• ✅ Relational Inference
• ◻️ Results and Model Analysis
• ◻️ Handling other entries than OneHot (Look-up Table for pretrained embeddings)
• ◻️ Adding Optim package for better sgd (momentum and others)

Memory Network

---

Create the dataset

python preprocessing/preprocessing.py

You can adjust the level of detail in the generated answer using the level argument. (default is -level 0) -level 1 compute sentence instead of words.

transform csv data into torch tensor (only if recomputed csv)

th main.lua -generate_dataset 1

Perform Memory Network Training

th main.lua -train_mem_net 1 To run with default parameters

Go into further parametring

th main.lua -train_mem_net 1 -num_mem 10 -feature_dim 174 -voc_size 58

The written parameters above are the default one.

Be careful that : - voc_size is >= to the number of word in the dataset

---

The different modules can be found under models/*_module.lua

The current main.lua replicates the Basic Question-Answering task described in the paper, but the implementation should be able to handle other tasks.

Results are stored in results

Results

Upcoming