This repository contains the submissions, by team_zhaw, to various AutoDL challenges. All the submissions are in the respective challenge format and can easily be run using the starting kit.
- Based on MobileNetV2
- Uses bloated classifiers for increased sample efficiency
- Temporal processing for videos is achieved using a singular 3D convolution before global pooling
- Uses SVM with many N-grams and TF-IDF
- In each iteration a new SVM with a specific N-gram range starts
- The prediction is composed of predictions of each iteration (voting)
- Preprocessing for Chinese and English is different (Tokenizer, N-gram word vs N-gram char)
- HashVectorizer is used for speed
- Semi-supervised Learning Task: Iterative training on the already labelled data and labelling of unlabelled data to train LightGBM models.
- Positive-Unlabelled(PU) Learning Task: same as Semi-supervision but choosig of negative samples differ.
- Learning on Noisy Labels: Extremely week GBM classifiers and using a BoostedEnsemble of these classifiers
- Based on baseline 3
- Uses bloated classifiers for increased sample efficiency for the vision tasks