Thesis can be found here:
bs_thesis_en.pdf
Диплом расположен тут:bs_thesis_ru.pdf
This repository contains the code used to accomplish N.S. Detkov’s bachelor’s thesis.
Topic: Influence of the label smoothing of pseudo-labeled data on training Convolutional Neural Networks
Steps to reconstruct the solution:
- Download the data from the site and extract to the
input/
folder - (Optional) Exploratory Data Analysis
Further, you need to open the
cd src juputer lab
EDA.ipynb
file — it contains an overview and visualization of the presented data - Data pre-processing, which generates reduced resolution images and splits the training data into folds
cd src bash preprocess.sh
- Launch of the training process
python train_infer.py -c exp_train_02.yaml
- Review the statistics on the predictions of the test set
python show_prediction_stats.py -f exp_train_02.csv
- Generation of pseudo-labeled datasets for the fine-tuning
ipython Create_Datasets.ipynb
- Generation of pseudo-labeled datasets applying label smoothing, each of which is "experiment"
ipython Create_Experiments.ipynb
- Launch fine-tuning experiments on pseudo-label with label smoothing
ipython Run_Experiments.ipynb
- Overview of the experiments results
Further, you need to open the
juputer lab
Analyse_Experiments_Results.ipynb
file — it contains an overview and visualization of the presented data