Skip to content

Latest commit

 

History

History
24 lines (12 loc) · 1.9 KB

README.md

File metadata and controls

24 lines (12 loc) · 1.9 KB

Wildlife tracker

Wildlife are a very important factor in our ecosystem as they help balance natural ecosystems. They also offer revenues for many countries through tourist attraction sites, game drives and conservation hubs. One problem evident is how sometimes the ever changing climate sometimes drives them out from the forests thereby resulting into human-wildlife conflict. This often results into destruction of property and loss of human life in some instances.

Training

Training deep learning models for wildlife tracking on CPUs can be time-consuming. Utilizing GPUs, such as those provided by Google Colab, is essential for efficient training. Models will be trained on GPUs and later saved for deployment using torch on CPUs. This documentation[https://pytorch.org/tutorials/beginner/saving_loading_models.html] comes in handy to understand loading of GPU trained models on the CPU.

The optimum epochs used for training that saw an improvement in the precision, recall and F1 scores was 25. It is more evident here

The chosen metrics seem to be improving with every iteration as evident in the Precsion-Recall curve and F1 Scores.

The Precision and Recall curves independently also show the same characteristic in the combined curve.

Deployment

The saved model weights will be deployed using Streamlit because of its simple UI. The saved model can also be deployed on Neural magic.

The optimum model weights used in the training stage to improve the precision and recall scores using the GPUs can be found here.

Future steps

Deploy the model on the Neural Magic platform because of its scalability properties and efficiency in handling large volumes of data. As this is a deep learning model, deploying it uisng Neural magic in future will offer the efficiency required.