Developed classification algorithm for different gestures based on time-series data collected using single-channel EMG electrode on flexor muscle.
Self-contained Pipeline
I have created a fully self-contained automatic pipeline for every step/component of the project so that other team members who are not familiar with data processing, programming, and deep learning can directly operate. The functionalities that the pipeline offers are broken down into the three main components below. Access the Google Colab notebook of the full pipeline using the button below:
01
Automated data preprocessing & labeling pipeline
The code pipeline can transform raw files into formatted datasets tailored for deep-learning model training. By opening up interactive windows showing the raw signal and smoothed signal, users can label time series data, distinguishing between gesture regions and relaxed segments. These user-defined labels are then encoded directly into the dataset files, in preparation for the model training process.
02
Automatic modeling pipeline
The automatic modeling pipeline provides interactive dialogues for users to specify the parameters, model storing directory, and log file directory. All the related files (model file, log file, training history file, etc.) are automatically sorted and stored in designated places that are easy to find. The training files are automatically scaled using Standard scalar, and the scaling metadata (mean and variance) are automatically stored as JSON file and will be updated when new data are included.
03
Automatic model validation pipeline
This automatic validation pipeline allows calling an already trained model and validating on a specified validation set (not seen by the training pipeline) and producing a confusion matrix.
