aek-img-trainer 0.5.3

Image classification trainer using OpenCV and timm

AEK Image Trainer

AI image preprocessing library

Installation

pip install aek-img-trainer

For more secure way after doing that you can upgrade via:

pip install --upgrade aek-img-trainer

Usage

Create object

Now you can just use Trainer class methods example usage is shown below.

from aek_img_trainer import Trainer, Preprocessor

model = Trainer(train_root="root/trainset",
                val_root="root/valset",
                num_classes=16,
                img_size=224,
                batch_size=4,
                val_reach=0.9999,
                num_epochs=150,
                learning_rate=1e-3,
                checkpoint_path="efficientnet_b0_best_model.pth",
                model_name="efficientnet_b0",
                device=None,
                augment=True,
                scheduler=None,
                scheduler_params=None,
                pretrained=True)

Those hyperparameters without train and val dataset path are default if you want to use default parameters you can just give your train and val datasets' path.

model = Trainer(train_root="root/trainset",val_root="root/valset")

Training

You can train your model with parameter that created earlier.

model.train()

Prediction

You can use your model in test with below code.

model.predict(checkpoint_path="example_model.pth",
               test_path="root/test.png")

Or if you give the test set path this method predict all images in the dataset folder. Your dataset should be labelled like training dataset.

Example usage:

model.predict(checkpoint_path="example_model.pth",
                test_set_path="root/testset")

Save OpenVINO IR

You can save the pth model file to xml and bin files

Example usage:

model.save_openvino(
    pth_path="efnet.pth",
    output_path="efnet.xml",
    class_number=16,
    model_name="efficientnet_b0"
)

Prediction on OpenVINO IR Model

You can predict the OpenVINO IR model. Like above predict function predict_xml takes either test_path or test_set_path and changes the result with the given path, either predict one image or images that inside the test set. Test set should be labelled like a training dataset.

Example usage:

model.openvino_predict(
    xml_path="efnet.xml",
    bin_path="efnet.bin",
    test_path=None,
    test_set_path="root/testset"
)

Information

You can see your model's parameters and architecture.

model.print_model_info()

Help function for Trainer class

You can use help() function for get more information about functions that inside the Trainer class.

model.help()

Timm models

You can get models that inside the timm library you can use with their name in string format inside the Trainer() 'model_name' paramaters. ATTENTION: You can just use the models whose head layers are 'fc', 'head' and 'classifier'

model.list_all_timm_models()