mkommar / plants Goto Github PK

Everything you need to evaluate is ready. For Apple Silicon: pip install torch --pre

Install requirements: pip install -r requirements.txt web.py

Select an image from the plant_images/train or plant_images/test to evaluate. The defualt location is http://127.0.0.1:8080

Requirements:

Python 3.10.12
Bing API key for image search

For Apple Silicon: pip install torch --pre

Install requirements: pip install -r requirements.txt

First get images from the Wikipedia page of your choice. Make sure to add the Bing API Key to the file: get_images.py

To update failing URLs: update_pending.py

To find non-JPEG files: ./verify.sh | grep 'Non-JPEG'

To convert images that can be converted to JPEG: ./convert.sh

(optional) If the testing set is small, use the move_files.sh script to move 20% of files: ./move_files plant_images/train plant_images/test

Train your model using the ResNet18 as the base (trains to 30 epochs): train.py

Example Output for 10 epochs:

Epoch 1/10 Loss: 3.4192 Acc: 0.0681
Epoch 2/10 Loss: 2.7664 Acc: 0.2893
Epoch 3/10 Loss: 2.2519 Acc: 0.4182
Epoch 4/10 Loss: 1.8852 Acc: 0.4990
Epoch 5/10 Loss: 1.6163 Acc: 0.5996
Epoch 6/10 Loss: 1.4078 Acc: 0.6488
Epoch 7/10 Loss: 1.2197 Acc: 0.6918
Epoch 8/10 Loss: 1.0836 Acc: 0.7107
Epoch 9/10 Loss: 0.9640 Acc: 0.7348
Epoch 10/10 Loss: 0.8566 Acc: 0.7579
Test Accuracy: 0.2444

Metadata and state data is saved and can be used between code and processor types.

Run the webserver and try uploading an images: web.py