Install the following in a python environment:
$ pip install -r requirements.txt
pip install torchvision
pip install -U albumentations
pip install pandas
pip install -U scikit-learn
pip install pycocotools
pip install ensemble_boxes
pip install labelbox
pip install nicegui (only for QR code extraction)
Using the UI
Start the UI by running python UI.py in the root directory. Select commands to run from the list. If using a fresh install they should be run in order.
Multiple commands can be run by entering numbers seperated by commas in order. Ex: 0,1,5 will download images from Labelbox, make the rgbJson, then exit.
UI options
- (0) Download images from Labelbox Downloads mask images using the labelbox api as well as a json file with the image annotation information. Images are stored in Data/raw/segmentedImages. The export Json is stored in Data/raw/exportProject.ndjson.
- (1) Make the rgbJson file
Extracts rgb values and external IDs associated with each annotation in each image from data/raw/exportProject.ndjson. These values are stored in data/processed/rgbPairs.json. - (2) Make the bounding box csv file
Uses the segmented images and rgbPairs.json to define a bounding box around each annotation in the format x1,y1,x2,y2 and store it in data/processed/boundingBoxes.csv - (3) Train a model
Starts the training process using settings form configs/trainConfig.json. - (4) Run model
Runs a model using settings from configs/runConfig.json - (5) exit
Training configuration
General settings regarding training are set in configs/trainConfig.json. Changes to the model must be made in src/models.py.
The GetTrainingModel function should be edited to make modifications to the model.
Modifications to model saving should be made at the bottom of the TrainModel function.
The model saves as a .pt file on intervals and at the end of training. It also saves a state dictionary as a .tar at the end of training.
configs/runConfig.json-- configuration for running models.configs/trainConfig.json-- configuration for training models.data/processed/boundingBoxes.csv-- bounding boxes extracted from the image masks.data/processed/rgbPairs.json-- rgb values associated with each annotation generated from exportProject.ndjson.data/raw/origionalImages/-- place the original images here.data/raw/run/-- place images here to run inference on.data/raw/segmentedImages/-- location that labelbox mask images are downloaded to.data/raw/exportProject.ndjson-- annotation data downloaded from labelbox.model_checkpoints/-- location model checkpoints are saved to.outputs/-- location images are saved to after inference.qrDetection/input/-- place images here to run qr code extraction on.qrDetection/output/-- renamed images are saved here.src/dataPrep.py-- downloads images from labelbox and makes exportProject.ndjson, rgbPairs.json, and boundingBoxes.csv.src/datasets.py-- splits the training images for training and validation and defines a custom dataset that is used to feed training data to the model.src/models.py-- creates, trains, loads, and runs models.src/utils.py-- Utilities mostly related to file IO.modelNotes.txt-- notes about several models trained on beocat.qrcode.py-- runs a GUI for extracting qr codes from images and renaming the images.UI.py-- runs a CLI tool for setting up and running models.
