Code repository for Gulordava, Brochhagen & Boleda (2020): Deep daxes: Mutual exclusivity arises through both learning biases and pragmatic strategies in neural networks
Get in touch if you have any questions!
The code is written in python 3.6
Packages required:
- torchvision 0.2.1
Installing torchvision will also install other required dependencies
The scripts expect the symbolic and visual data to be placed in appropriate mutual-exclusivity/data
-(sub)folders
- For example, the data for symbolic experiments using the transcriptions of CHILDES used in Frank et al. (2009) may be placed in
data/frank2009/all_words
; - and the pre-processed bounding boxes of Flickr30K objects in
data/flickr
Refer to the paper for details
Results are written to mutual-exclusivity/results
python mutual-exclusivity/src/train_symbolic.py --data mutual-exclusivity/data/frank2009/all_words --seed 1008 --loss maxmargin_words
Evaluate on dogs (from Flickr training set) without competition
python mutual-exclusivity/src/train_flickr.py --data mutual-exclusivity/data/flickr/ --debug --lr 0.1 --loss maxmargin_words --novel_set dogs