Trash Identification

A deep learning model capable of identifying various solid waste items such as glass, paper, cardboard, plastic, metal, and trash.

Requirements

• Pytorch 1.13+
• Numpy 1.16.4+
• Matplotlib 3.1.1+

Dataset

Trashnet: 2527 Images - 2274 Training (90%) & 253 Validation (10%)

• 501 Glass
• 594 Paper
• 403 Cardboard
• 482 Plastic
• 137 Trash
• 410 Metal

Process

1. Reorganize Dataset to have a training and validation directories with subdirectories containing each category
2. Initalize the pretrained model
3. Reshape the final layer(s) to have the same number of outputs as the number of classes in the new dataset (6)
4. Define for the optimization algorithm which parameters we want to update during training
5. Run the training step

Model

Model Name	Accuracy	Time*
Resnet	0.953125	26m 50s
Alexnet	0.914062	13m 23s
VGG	0.960938	59m 23s
Squeezenet	0.941406	15m 10s
Densenet	0.964844	29m 34s
Inception	0.972656	40m 56s

*Can be reduced using more powerful GPUs, reducing its importance for now.

To Do

• Try with various architectures
• Save and export model for post-training evaluation to finetune hyperperameters

Acknowledgements

Thank you @garythung for the dataset and @mpcrlab for the help