A naukri.com initiative
Medium
10h
3
Image Credit: Medium
Garbage Classification with FastAI: Training, Interpreting, and Deploying to Hugging Face
- Recycling plays a crucial role in sustainable living, starting with sorted waste, motivating the project to classify household garbage into 12 categories.
- Fastai library was utilized for training in garbage classification using models like ResNet34 and ResNet50 pretrained on ImageNet, and the model output was saved as a .pkl file.
- The Gradio interface was implemented for user image testing, and the project was deployed on Hugging Face for accessibility.
- Data collection was mainly from web scraping and open-source datasets, forming the basis of the classification project.
- Data augmentation techniques were employed, such as image resizing and transformation pipelines to preprocess the images.
- The DataBlock was established for image and label processing, followed by model training and val ratio split.
- The trained model was fine-tuned with accuracy metrics and 2 epochs of training to build the initial classification model.
- ClassificationInterpretation was used to analyze model errors and improve accuracy by data cleaning and retraining.
- Optimal learning rate determination through learn.lr_find() and model training for improved performance.
- Models were saved at different stages, including the ResNet34 base model and the ResNet34 in freeze-unfreeze method.
Read Full Article
Like
For uninterrupted reading, download the app