Project : Kidney Stones detection

Getting a working template pipeline for detection with many features in training and reporting

Description

• This project is designed to train a model for Kidney Stones detection , where:
  • input ---> is a 2D CT coronal slice .
  • output ---> is a bbox around stones
• The used model is YOLO-8

Usage

• link [https://osama-ammar-detection-app-streamlit-app-kuziji.streamlit.app/]

Configuration

The project can be configured by editing model_config.yaml file . The available configuration options are as follows:

- `input_size`: The size of input images (default: ...).
- `in_channels`: The number of input channels (default: 1).
- `out_channels`: The number of output channels (default: 2).
- `epochs`: The number of training epochs (default: 300).
- `batch_size`: The batch size for training (default: 16).
- `learning_rate`: The learning rate for optimization (default: 1e-3).
- `log_period`: The interval for printing training logs (default: 1).
- `criterion` : The Loss function used (default:....).
- `optimizer`: The optimizer (default:AdamW).
- `mode`: The mode of operation (train/test/export).
- `weights_path`: The path to load the checkpoint for testing or exporting .
- `dataset_path`: The path to the dataset directory .
- `results_path`: The path to save the result files .

Dataset

The dataset used for training and testing should be placed in the specified dataset_path directory. The dataset .......

• data set Balance :

Training

To train the model, set the mode argument to train. The training progress, including loss and accuracy, will be displayed during training.

Testing

To test the trained model, set the mode argument to test. The model will be loaded from the checkpoint_path, and the test results will be displayed.

Exporting

To export the trained model, set the mode argument to export. The model will be loaded from the checkpoint_path, and the exported model will be saved in the model_path.

Results

The results, including trained models and logs, will be saved in the specified model_path and result_path directories, respectively. Each run of the program will create a subdirectory in the result_path directory named with the model type and the current date and time.

---

Model performance

• Generally , accuracy = .......................
• Accuracy in bad Cases = .........................

---

TODO

• General steps

  • select a problem (simple one)
  • prepare data and make analysis to it , track it
  • select a model and train , fine tune or use it directly if applicable
  • view results in MLFlow or weight and bias
  • deploy it in a gui or AWS service
  • make a report with images and links

• Editing hyperparams in only ony file

• model versioning

• Applicability of different dataset

• Augmentations scripts

  • image classification
  • bboxes detection
  • segmentation
  • overall example

• automatic experiementation

• early stopping

• CI/CD

• making readme template ...including (images , tables , videos , emojies ...)

• Cleaning and organizing

• Feature engineering

• Error analysis

• Using weight and biases

• apply DVC

• [/] use MLFlow

  • log params
  • log models
  • log dataset

• making one file -scv- for tracking experiments , each expriement and its parameters and metrics

  • generating this file automatically or generating it after making experiemnts (replaced by MLFlow)

• include training time and computer resources in trial summary

• apply some performance optimization mentioned in [https://pytorch.org/tutorials/recipes/recipes/tuning_guide.html]

---

Credits