What this project does

• Goal: Classify images as either logo or nologo and make it easy to search a folder for items like “t‑shirts with/without a logo”.
• Artifacts: A trained Keras .h5 model in models/ and two user interfaces: a chat-style CLI and a Streamlit web app.

Where the trained model is used

• The .h5 model is loaded by model_utils.load_trained_model().
• Both chat_cli.py and streamlit_app.py import load_trained_model from model_utils.py and use it for inference.

Key components

• model_utils.py

  • load_trained_model(path=None): loads the trained .h5 from models/imageclassify.h5 (default) or a path you pass.
  • predict_image(model, image_path): returns {image_path, logo_probability, predicted_label}.
  • predict_folder(model, folder, keyword_filter=None, limit=None, recursive=True): scans a folder, predicts for each image.
  • Note: The model trained in the notebook outputs a value where higher means “nologo”. The utilities invert that so logo_probability reflects the chance of a logo.

• chat_cli.py

  • Loads the model and runs an interactive prompt (chat-style) to search folders for items like “show me t-shirts with logo”.
  • Parses phrases such as "with logo" and "without a logo".
  • Scans data/ and logos/ by default (or a custom folder via --scan).

• streamlit_app.py

  • Web UI to upload a single image or scan a folder and visualize predictions.
  • Sidebar inputs let you choose model path and scan folder; results are shown with thumbnails.

How it works end-to-end

1. Your .h5 model is discovered at models/imageclassify.h5 by default.
2. Images are loaded and resized to the model’s input size, normalized to [0,1].
3. The model’s output is post-processed so that logo_probability is in [0,1] and predicted_label is either logo or nologo.
4. CLI/App filter results by label and keyword, and sort by confidence.

Quickstart

• Install deps:

  • pip install -r requirements.txt

• Chat CLI:

  • python chat_cli.py --model models/imageclassify.h5 --scan data
  • Example queries:
    • show me t-shirts with logo
    • show me t-shirts without a logo
    • top 10 caps with logo

• Streamlit app:

  • streamlit run streamlit_app.py
  • Use the sidebar to set model path and scan folder; use tabs to upload or scan.

Configuration and customization

• Model path: pass --model to CLI or set the sidebar field in the app.
• Scan path: pass --scan to CLI or set the sidebar field in the app.
• Label mapping: If you retrain a model with different output ordering, adjust _postprocess_output() in model_utils.py.

Troubleshooting

• "No matching images found":
  • Ensure the keyword matches filenames or parent folder names (e.g., use t-shirt, tshirts, or tee).
  • Check the scan folder contains images with extensions like .jpg, .png.
  • Try removing the keyword to see if images are detected at all.
• Predictions look inverted (logo vs nologo): Revisit _postprocess_output() and the logic from your training notebook.