Skip to content

TROUBLESHOOTING.md

Latest commit

 

History

History
96 lines (60 loc) · 2.34 KB

TROUBLESHOOTING.md

File metadata and controls

96 lines (60 loc) · 2.34 KB

Troubleshooting

Flask App Does Not Start

Check that dependencies are installed:

cd dermascan_app
source .venv/bin/activate
pip install -r requirements.txt
python app.py

If TensorFlow import fails, confirm that your Python version and platform are compatible with the TensorFlow package being installed.

Model File Not Found

The app expects:

artifacts/models/current/melanoma_detector.keras

If this file is missing, restore the packaged model artifact or run training after preparing the dataset.

Upload Returns "Invalid File Type"

The app currently accepts only:

  • png
  • jpg
  • jpeg

File extension checking is not a complete security control. It is only a basic prototype filter.

Upload Returns "Could Not Process The Image"

The file was saved, but OpenCV could not read it as an image.

Try:

  • using a standard RGB JPEG or PNG
  • checking that the file is not corrupted
  • avoiding screenshots or non-image files renamed with image extensions

Training Fails Because Dataset Directory Is Missing

Training expects:

data/melanoma_cancer_dataset/
  train/
  validation/
  test/

Each split must contain benign/ and malignant/ folders.

Evaluation Fails Because Test Data Is Missing

Evaluation expects:

data/melanoma_cancer_dataset/test/benign/
data/melanoma_cancer_dataset/test/malignant/

The repository does not include the dataset.

TensorFlow Or Keras Version Problems

The packaged model metadata reports Keras 3.8.0. If loading fails:

  • check the installed TensorFlow/Keras version
  • install a compatible TensorFlow release
  • regenerate the model artifact locally from the training script

TODO: Add pinned environment file after dependency versions are finalized.

Results Differ Between Runs

The current training script does not set random seeds. Differences can occur from:

  • random initialization
  • dataset shuffling
  • hardware/backend differences
  • dependency version differences

For reproducible experiments, add explicit seed control and pinned dependencies.

Runtime Prediction Differs From Evaluation Threshold

The Flask app currently uses a runtime threshold of 0.4. The evaluation pipeline reads 0.35 from ml_pipeline/config/model_parameters.json.

This mismatch is documented as a technical gap. Align thresholds before comparing app behavior to evaluation metrics.