A simple neural network module for relational reasoning
Knet implementation of Relational Networks - A simple neural network module for relational reasoning
Implemented & tested on CLEVR from state descriptions.
Usage
- Download clever data (without images) and preprocess the data.
cd relational-network
bash prepare_data.sh
- Run experiment script to train the model on state descriptions
bash run.sh
- The model and log file will be generated inside
src/saved_models
Results
| Epoch | Accuracy (Val Set) |
|---|---|
| 1 | 44.07% |
| 5 | 47.50% |
| 15 | 57.69% |
| 25 | 79.60% |
| 40 | 93.21% |
| 65 | 94.50% |
- Training Time: On Tesla K80 GPU, 1 epoch takes ~ 12 minutes.