Fire Image Classification's Project

Its an image classification model (CNN) to classify fire images from nonfire images.

Challenges

1. the most data in our datset are images based on outdoor environment so it's not suitable for classifying indoor environments.

Training the model

for train a model, one can easily run the python3 train.py with this argument:

python3 train.py -c train_config.json

train parameters

The paths of all data and backbone weights are in Aparat2 Server. For more details refer to fire_image_classification/train/data_path/data_path.txt

1. train_path : tain path of data wich includes fire and Nonfire folders.

2. test_path : validation path of data wich includes fire and Nonfire folders.

3. num_epoch : an integer number of epochs.

4. learning_rate : a float number (for example: 0.001).

5. backbone_name : a string value based on the existence models (for example: "resnet18", "mobilenet")

6. model_json_file : a json model architecture path based on the existence models(for mobilenet its valus in None)

7. path_weight : path of weights.

8. checkpoint_path : path for saving checkpoints.

9. num_freeze : an integer number to freeze backbone layers (for mobilenetv2 its value is 15 but for resnet18 its value is 78).

10. target_shape : a tuple of integer numbers to set inputs shape of model (for example [448, 448])

11. batch_size : an integer number of batch size of images.

12. early_stopping_mode : a string value based on the existence models (for example: empty string "" for resnet18, any string value for mobilenetv2).

13. lr_reduction_mode : a string value based on the existence models (for example: empty string "" for resnet18, any string value for mobilenetv2).

To run this product you can activate this virtual environment:

source /home/bahrami/fire/new_git_FC/bin/activate

Running the example

You can easily run the python3 example.py with no argument.

python3 example.py