shresthakamal / custom-image-classification

Custom Image classification using multiple DL architectures

lglut-classify

Build Instructions

Make sure the folders A and U containing images are place inside data folder like shown below:

       ROOT
        ├── api
        │   ├── app.py
        │   ├── routes.py
        │   ├── services
        │   │   └── predictor.py
        │   ├── static
        │   │   └── banner.jpg
        │   └── templates
        │       ├── index.html
        │       └── predict.html
        ├── data
        │   ├── A
        │   │   ├── A8.jpeg
        │   │   └── A9.jpeg
        │   ├── data.csv
        │   └── U
        │       ├── U8.jpeg
        │       └── U9.jpeg
        ├── lglutide
        │   ├── architectures
        │   │   ├── alexnet.py
        │   │   ├── cnn.py
        │   │   ├── densenet.py
        │   │   ├── nn.py
        │   │   └── resnet.py
        │   ├── dispatcher.py
        │   ├── logs
        │   ├── main.py
        │   ├── make_data.py
        │   ├── models
        │   │   └── 8
        │   │       ├── config.json
        │   │       └── Mar25_08_30_fold_4.pth
        │   ├── predict.py
        │   ├── runs
        │   └── utils
        │       ├── augmentations.py
        │       ├── logger.py
        │       └── options.py
        ├── Makefile
        ├── notebooks
        │   └── playground.ipynb
        ├── poetry.lock
        ├── pyproject.toml
        ├── README.md
        ├── requirements.txt
        └── tests
            └── test_lglutide.py

Install Dependencies:

Before installing the project dependencies, make sure you are working on a virtual environment (conda/venv).

make install

or

pip install -r requirements.txt

Training the model

All the training arguments and hyperparameters are specified in the lglutide/config module.

python3 -m lglutide.main --experiment <experiment_no> --epochs <EPOCHS> --batch <BATCH>

eg:

python3 -m lglutide.main --experiment 6 --epochs 1 --fold 2 --batch 64

Training Flags:

• Seed (int)
• Training Batch Size (int)
• Training Epochs (int)
• Image Channel (int)
• Image Width (int)
• Image Height (int)
• Learning Rate (float)
• Weight Decay Value (float)
• Gradient Accumulation (int)
• Dropout (float)
• Number of k-folds (int)
• Experiment No (int)
• Data Augmentation (bool)
• Usage of GPU (bool)
• --densenet(default) or --resnet to specify training model (bool)

Experimental Tracking sheet, here.

Checkpoints are uploaded here.

Evaluating the Training metrics

The training logs like training loss, validation/test loss, F1 scores for individual epochs are all recorded using tensorboard writers. All the event logs are recorded under runs/<experiment_no>. To view the metrics, execute the following command:

tensorboard --logdir=lglutide/runs/<experiment_no>

eg:

tensorboard --logdir=lglutide/runs/6

Inference

All the inferences are made through config.json files which contains the details about the trained model checkpoints, hyperparameters used to train the model, model names and more.

The config.json file generated after each training can be found in: lglutide/models/<experiment_no>/config.json, depending on the experiment that we ran.

eg: lglutide/models/6/config.json

1. Running Inference through CLI

• To run inference from command line MAKE SURE you have the following architecture:

    ROOT
      └──lglutide
          └── models
              └── <experiment_no>         # Training Experiment No
                  ├── config.json         # Training Hyperamaters
                  └── <saved_model>.pth   # Trained model checkpoint
  

• Download the checkpoints corresponding to a specific experiment and run inference.

• Run inference using the following command:

python3 -m lglutide.predict --config <CONFIG FILE PATH>

eg:

python3 -m lglutide.predict --config "lglutide/models/6/config.json"

• This will prompt you to type the image path, remeber the root foler before specifying the path, eg: data/A/A1.jpeg

2. Running Inference through APP

• To run the inference through UI application, make sure you specify the config.json file on .env file (found in root directory) like shown below:

CONFIG="lglutide/models/<experiment_no>/config.json"

eg:

CONFIG="lglutide/models/6/config.json"

• Now execute the following command to start the application:

python3 -m api.app

• You can now open http://127.0.0.1:5000/ on your browser to make the inferences