yuan-ds / sorting-sequences-shot-change

Group Project for CMU 11-785 Deep Learning Course

Entropy-Optimal Sorting For Random-Shuffled Sequences With Shot Change Robustness

Final Project for CMU 11-785 Deep Learning
Team 17: I-Tsun Cheng, Tz-Ruei Liu, Yuanyuan Wang, Tianrun Wang

This is the official codebase for our project. The code can be structured into two components: Data and Models.

Data

The data folder contains the following three folders:

• data: the datasets used for different input sequence configurations for our main experiments (1+1, 2+1, 3+1, 4+1, 5+1)
• utilities: the data utilties used for generating the datasets

To generate the .csv files for each input configuration in the data folder, navigate to the utilities folder. Run get_shot_lists.py which will take knnw_labels.csv and output a list of shots in shot_list.csv. Then, run get_data.py to generate all the .csv dataset files by inputting shot_list.csv.

Models

The models folder contains all of the models that are implemented and tested in our experiments. Here is an overview of the models folder:

• direct: The models we implemented and trained for our main set of experiments
  • baseline_approach: The models implemented using the baseline approach
  • proposed_approach: The models implemented using our proposed approach
  • data_augmentation: The best model trained using our proposed approach with different data augmentation techniques (.README file provided in the respective folder for more details)
• pairwise: The models that we attempted using the pairwise prediction approach mentioned in our ablation study

Code Citations

Here we would like to attribute the code that we referenced for our model implementations:

EfficientNet from scratch reference: https://github.com/WZMIAOMIAO/deep-learning-for-image-processing/tree/master/pytorch_classification/Test9_efficientNet
RegNet code reference: https://github.com/facebookresearch/ClassyVision/blob/main/classy_vision/models/regnet.py