Deep-Learning-FPT

Projects on class

Convolution step by step

• This notebook is based on an assignment from the Convolutional Neural Networks course by DeepLearning.AI.
• I completed the implementation as part of my own learning process.
• Modifications, comments, and explanation are my own. For educational purposes only.
• This notebook implements forward propagration, backpropragation of volume convolution: 1 layer (Input -> Conv -> Pooling -> Output)

Sentiment Analysis - Binary classification

• Using the dataset ibmdb_reviews which containts 50,000 reviews films, splitted into 25k train and 25k test, labels are 0 or 1 as negative or positive
• Techniques used: Word embedding, bidirectional LSTM, Dropout, L2 Regularization
• Steps:

1. Load, process data:
   1.1. Use tensorflow_datasets to load the "ibmdb_reviews"
   1.2. Split into train, test
   1.3. Tokenize, truncating, padding sentences
2. Build sequential model and train:
   2.1 Embedding, Bidirectional(LSTM), Dropout, Dense(ReLU), Output(Sigmoid)
   2.2.Loss: binary_crossentropy, optimizer: Adam
3. Evaluation uses accuracy

Project 1: Hyperparameters tuning

• I try to achieve better metrics by using algorithms that search for hyperparameters, such as Bayes Optimization, Genetic Algorithm
• This project is comparison on optimizers (Bayes Optimization and Genetic algorithm) on the Fashion MNIST dataset
• Genetic Algorithm requires a really long time for searching and a strong device to train, recommend for the Bayes

Project 2: Deepfake detection

• My team and I used pre-trained ResNet50 to distinguish between real and deepfake images with a pretty number ~ 81% on test data
• Architecture: Based Model (ResNet50) -> Pooling -> Densely -> Output
• For more details, check on Project 2 - DPL.pdf (Vietnamese), slide: Deepfake detection.pdf
• Link dataset: https://drive.google.com/file/d/1HnrO5sGg8gc2VuC1dUHE_62WIbYMSXid/view?usp=sharing
• Link model.h5: https://drive.google.com/file/d/1LF-fYG8r5gXuvHmEEA2y8xLyKeE7Z5pB/view?usp=sharing

Project 3: Speech emtion recognition

• With the same architecture (2 x CNN -> LSTM -> Output), same distribution of train - test set, this aims to compare performance model using normal data vs augmentation data
• Datasets used: Crema_D, RAVDESS, SAVEE, TESS. These 4 datasets express emotion through voice, not through meaning of words.
• For more details, check on DPL-Project-3- Report.pdf (Vietnamese), slide: Project 3 - DPL.pdf
• Link dataset: https://drive.google.com/file/d/1jS8GiWf2PMS_91gGcjCwESsA0MSuyOT9/view?usp=sharing