Forked from rosinality/mac-network-pytorch which is based on Memory, Attention and Composition (MAC) Network for CLEVR from Compositional Attention Networks for Machine Reasoning.
This fork is born after observing that existing pytorch implementations were performing weel below what they should according to the paper and the oficial tensorflow implementation. Following the later the changes done to the implementation this repository was forked from are:
- Fixed dropout values (for each of now multiple dropout sectors).
- Removed variational dropout for control.
- Fixed LSTM dimensions.
- Fixed bidirectional RNN question vector.
- Added dropout to convolutional layers.
- Fixed weight initializations.
- Added missing controlUnit projection.
- Added two missing projections to readUnit (and respective activations).
- Removed unnecesary projection in writeUnit.
- Doubled layers in classifier_out.
- Many more things I forgot to write here...
This implementation also cleans up the code and makes it more easily modified by further separating the model into components, using pytorch 1.0's device, and using yacs for all configuration parameters.
Run the following command to install missing dependencies:
pip install -r requirements.txtpython preprocess.py [CLEVR directory]
python image_feature.py [CLEVR directory]python train.py DATALOADER.FEATURES_PATH [CLEVR directory]Alternatively, change the default path to the pre-computed features in the default configuration (under config/defaults) to be able to just run python train.py for training.
To experiment with the model without changing code just create a configuration file and pass its location as an argumet to train.py:
# a sample configuration file is provided at configs/train.yaml
python train.py --config-file=configs/train.yaml