A comparison study of pre-trained and randomly initialized models on image and sequence data

---

• This repo contains the final project for the Master's course Deep Learning at University of Groningen

Computer Vision - Cassava Leaf disease image classification

---

Dataset

• The dataset can be downloaded from here

Baseline CNN model

• The Base model can be found in src/cv_task/models.py
• Base trained model can be found in src/cv_task/base_model

ResNet models

• The ResNet-34 implementation can be found in src/cv_task/models.py
• The trained ResNet-34 models can be found in src/cv_task/model_pretrained and src/cv_task/model_random which are pre-trained and randomly initialized models
• Use the following to list all possible commandline parameters for training script

python src/cv_task/train.py --help

• For training with default parameters, run

python src/cv_task/train.py

• Use the following to list all possible commandline parameters for testing script

python src/cv_task/test.py --help

• For testing with default parameters, run

python src/cv_task/test.py

• To generate various plots, the following jupyter notebook src/cv_task/plot_graphs.ipynb is useful

Natural Language Processing - semantic analysis on tweet dataset regarding Coronavirus

---

• In the folder of the src/nlp_task, you will find models for semantic analysis on a tweet dataset regarding the Coronavirus.
• The semantic analysis is performed using a pretrained version of the BERT model and an untrained version of it, as well as a baseline LSTM model.

LSTM model

• The baseline model can be found in the base_lstm_model.py file. The program can easily be run by running python base_lstm_model.py. It will print some data statistics, a model summary, a classification report and it will generate a confusion matrix in the directory that this file is run from.

BERT model

• The BERT model can be found in the BERT_model.py file. The model is designed using the Google colab notebook found here. This model is pretrained using the bert-base-uncased model, and finetuned according the semantic analysis task with 5 different semantic labels. Before training and testing, the tweets are cleaned up using a cleanup function as defined by Edgar Jonathan for another BERT model on the same dataset, found here.
• The BERT model in the file can be run in both the pretrained and the untrained manner by commenting or uncommenting the following lines:

if __name__ == "__main__":
    # uncomment these lines to run an untrained BERT model
    print("\n\n----------------- untrained ------------------")
    main(True)

    # uncomment these lines to run an pretrained BERT model using bert-base-uncased
    print("\n\n----------------- pretrained -----------------")
    main(False)