One-click setup and run

Clone the repo and run setup.sh - it should set up the repo by downloading the dataset, installing requirements and running the demo training and inferencing file with 0.05 of entire dataset.

Architectures Compared:

• MobileNetV1 (YAMNET, Transfer learned)
• MobileNetV2
• MobileNetV3
• VGG16 (VGGish, Transfer learned)
• EfficientNetV1B0
• EfficientNetV2B0
• Conv-Mixer (WIP)

Setup

• Run setup.sh to download dataset and install necessary packages

Basic Train

• Run train_all.py to iteratively train all networks, do edit the dataset path within the script
• Run train_hp_tuners.py to perform basic hyperparameter search

Single Model Train

• train.py requires several arguments. Check train_all.py for default values

  python train.py -d <dataset_path> -u <number of dense units> -a <model architecture> -e <epochs> 
  -s <random_seed> -p <patch_hop_distance> -t <time_now (or other argument for marking training log)> -r <ratio of files> 
  

Real time inference

• Run record_audio.py on terminal to record wave files
• Run display_spec_rt_2.ipynb to visualize mel-spec and predictions in real time

Visualize Weights

• For conventional DNN models, run display_weights.ipynb to see convolutional filters
• For Attention based based models, attention visualization is done within their respective notebooks