Deep Learning Projects - [Python v3.7.6, PyTorch v1.5, CUDA v10.1]:

1 .Convolutional Neural Network (CNN)

• Classification of images containing spectrograms from audio files
• 10 classes total:
  • 1 anechoic class
  • 3 classes with different levels of noise
  • 3 classes with different levels of reverb
  • 3 classes with different levels of noise and reverb
• Data augmentation via delta-Mel spectrograms
• Transfer learning from the pretrained ResNet50 model
• Replacement of ResNet50's last layer with a sequence of layers and train only them
• Systematic search of hyperparameters:
  • Tuning
  • Fine tuning

2. Recurrent Neural Network (LSTM)

• Binary Classification of 1-dimensional sequences extracted from the history values of four cryptocurrency coins
  • Class "0" : falling price expected
  • Class "1" : rising price expected
• Data cleaning
• Class Imbalance removed
• Data scaling
• Automated best Learning Rate region extraction
• Systematic search of hyperparameters:
  • Tuning
  • Fine tuning

3. Deep Reinforcement Learning (DRL)

• Mountain Car problem of OpenAI Gym library
  • Agent with three available actions
  • Train an agent via rewards, on how he can gain momentum using a sequence of actions, in order to surpass a steep mountain
• Systematic search of hyperparameters:
  • Hyperparameter Tuning
  • Architecture - Policy tuning
• Visualization of learning via a gif of Q-tables (every 100 episodes)