word-learning.cogsci2020
@inproceedings{tsutsui2020wordlearning,
Author = {Satoshi Tsutsui and Arjun Chandrasekaran and Md Reza and David Crandall and Chen Yu},
Booktitle = {Annual Conference of the Cognitive Science Society (CogSci)},
Title = {A Computational Model of Early Word Learning from the Infant's Point of View},
Year = {2020}
}
Environment
See env.yml for the exact environment. As a shortcut, you can use the following python binary as is if you have access to salk.psych.indiana.edu.
/data/stsutsui/public/word-learning.cogsci2020/miniconda/bin/python
Alternatively, you can do
export PATH="/data/stsutsui/public/word-learning.cogsci2020/miniconda/bin:$PATH"
Make sure that which python will give the above python binary.
Image Data
I copied the necessary image files into the following.
./data/naming_3s_imgs./data/test_images
The naming rule is same as experiment 13 where cam07 is child view and cam08 is parent view, and frames are raw frames, and acuity is after applied acuity filter. These are not included in the github but is in the salk.
Training Data
The point of this work is to make a training set based on a criteria, and then train CNNs. I represent each set as a txt file. These files are inside ./data/dataset_txt. The name is sort of informative enough to figure out the corresponding set mentioned in the paper. It's a simple csv file to list path to images and corresponding labels.
CNN training
To train cnns, you need to go to the ./train_cnn directory. main.py is the training script. This script should be readable. You can see the help to get the meanings of args.
cd train_cnn
/data/stsutsui/public/word-learning.cogsci2020/miniconda/bin/python main.py --help
An example command to train is:
cd train_cnn
/data/stsutsui/public/word-learning.cogsci2020/miniconda/bin/python main.py --saveroot ../experiments/cogsci2020/ --train ../data/dataset_txt/naiming_3s_whole_img_acuity.txt --seed 1 --gpu 1
This trains with the subset of naiming_3s_whole_img_acuity.txt.
All the training I did for this paper is all_trainings.sh . The training results are in ./experiments/cogsci2020-reported/, which is included in the salk, but not in this github repository.
Results
The results plot used in the paper is in ./results. The code to make these plots are ./ipython/results-cnn-train-reported-cogsci2020.ipynb