Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/ilirosmanaj/detect_kermit/llms.txt

Use this file to discover all available pages before exploring further.

Training uses ImageAI’s ModelTraining class with a ResNet backbone to learn a binary classifier that separates kermit from no-kermit frames. The script reads your prepared image directories, runs for a configurable number of epochs with optional built-in data augmentation, and saves the resulting model weights to an HDF5 file ready for inference. Training time scales with dataset size, epoch count, batch size, and the compute hardware available.

​
Training script

The full training script is concise by design — ImageAI abstracts away the ResNet architecture, data loading, and checkpoint saving:
imageai_build_model.py
from imageai.Prediction.Custom import ModelTraining


def main():
    model_trainer = ModelTraining()
    model_trainer.setModelTypeAsResNet()
    model_trainer.setDataDirectory("data/images")
    model_trainer.trainModel(num_objects=2,
                             num_experiments=20,
                             enhance_data=True,
                             batch_size=16,
                             show_network_summary=True)


if __name__ == '__main__':
    main()

​
Training parameters

ParameterValueDescription
num_objects2Binary classification — the two classes are kermit and no-kermit.
num_experiments20Number of full training epochs over the dataset.
enhance_dataTrueEnables ImageAI’s built-in data augmentation (additional flips, shifts, and zoom transforms applied on the fly).
batch_size16Number of samples used per gradient update. Lower this value if you run into GPU memory errors.
show_network_summaryTruePrints the full ResNet layer-by-layer architecture to stdout before training begins.

​
Running training

1

Prepare the dataset

Make sure your data/images/ directory is fully populated with train/ and test/ subdirectories for both classes before starting. Follow the Dataset Preparation guide if you have not done this yet.
2

Start training

From the repository root, run:
python imageai_build_model.py
3

Wait for training to complete

Training time varies depending on the number of training images, the value of num_experiments, batch_size, and your hardware. With 20 epochs on a mid-sized dataset, expect anywhere from several minutes on a GPU to multiple hours on CPU only.
4

Locate the trained model

Once training finishes, the saved model weights are written to:
data/images/models/kermit_finder.h5
This file is referenced by the prediction script when running inference.

​
Output files

After a successful training run, two key files are produced: data/images/models/kermit_finder.h5 — The trained ResNet weights saved in HDF5 format. This is the file loaded by kermit_model_evaluation.py at inference time. data/images/json/model_class.json — A JSON file that maps integer class indices to human-readable class names: 
{
    "0": "kermit",
    "1": "no-kermit"
}
Both files are required together for predictions; the model file provides the weights, and the JSON file provides the label mapping.
For faster training, run on a CUDA-enabled GPU. If TensorFlow complains about a limited number of available devices, set the CUDA_VISIBLE_DEVICES environment variable before launching the script:
export CUDA_VISIBLE_DEVICES=''
python imageai_build_model.py
This forces TensorFlow to fall back gracefully and avoids device enumeration errors.
requirements.txt lists both tensorflow==1.12.0 and tensorflow-gpu==1.12.0. Install only the package that matches your hardware — installing both in the same environment will cause conflicts. Use tensorflow for CPU-only machines and tensorflow-gpu for machines with a CUDA-compatible GPU.

Build docs developers (and LLMs) love

Get started for freeTalk to us