- resized _devlog images to reduce file sizes
- areas into the target of Piece Detector
- switched from roboflow annotator to VIA
- renamed folders to because ultralytics yolo package names clashed
- training piece detector without warping
- CHANGELOG.md
- _devlog folder to blog everything
- resize.py to resize images in a folder
- New images data with higher resolution
- git is going to ignore data folder, due to GitHub size restrictions
- hough.py added to test board detections without CNN
Focused on overfitting board_detector data. After [0.3.1] (changes to board detector dataset), there was a major issue that was found. Overfitting only happened to particularly same few images while rest of the images were completely off.
After a ton of debugging, from ensuring the data is being shuffled, to checking individual images and figuring out exact overfitting images. The problem was simply with incorrect length function (len) in board dataset module.
- correcting len function in board dataset class.
used force push to push changes since local git and github did not have same branch. Had to figure out some trickery to finally sync local git and github.
updated .gitignore file to ignore checkpoints because of issues with pushing large trained files from Piece Detector.
- updated .gitignore file to ignore checkpoints
converted from labelbox to roboflow for ease of use and coco.json annotations output.
Learned about mixed precision training and incorporated it to both piece and board trainers to speed up training.
- dataloader/data.py
- converted from labelbox annotations to roboflow annotator
- testing.py file
- show_board_detector_results function to visualize output from board detector
- show_piece_detector_results function to visualize output from piece detector
- Mixed precision training to ensure faster training on GPU
Trainer added for Piece Detector using Adam optimizer and learning rate of 3e-4. the loss is a summation of losses dict output from fasterrcnn().
Clean up code and organized folders along with clear documentation of functions
created Piece Dataset module to prepare image data for training
- Piece Detector Dataset
Renamed skew to warp because warp sounds better. Importantly vectorized warp function using Kornia library.
- new file warp.py using kornia libray
- skew.py
readme file updated
small bug fix with utils init.py file
Organized code by using best practices and documentation along with .gitignore.
Created Board Detector model with Densenet201() with 8 outputs for detecting each corner of the board. Using 11 images of chessboards taken with an iPhone 13 Pro camera with square ratio, the model was trained with MSELoss() with 3e-4 learning rate. The goal was to overfit the 11 images data to ensure good model performance on large data.
Skew function added to be used after detecting the 4 corners. The skew function will skew the image to fit and fill the entire board perfectly in the center of the new image. Which will then be sent into Piece Detector to detect pieces
Created Piece Detector model which was just fasterrcnn.
- Board detector model created with basic Densenet201() with 8 outputs for detecting each corner of the board.
- Trainer for the model is created with MSELoss()
- Dataset of 11 images
- Created piece detector model
- skew function