Room Type Classifier

Русская версия: README_ru.md

Image classification project for room-type recognition. Case: Avito, predicting one of 19 room or property photo categories from an image. Primary quality metric: Macro F1. Best single model: ConvNeXt Nano. Final submission solution: an ensemble of ConvNeXt Nano, ResNet50, and ResNet18.

What is included

• preprocessing for train, validation, and test splits
• removal of an extra train-only class that is not part of the task or validation set
• class ids normalized to the 0-18 range
• several CNN models trained and compared
• experiments logged to MLflow through DagsHub
• final three-model ensemble assembled
• test submission generated
• Streamlit prototype for checking images interactively

Data

Raw data is expected in this layout:

data/raw/
  train_df.csv
  val_df.csv
  test_df.csv
  train_images/
  val_images/
  test_images/

Preprocessing:

just prepare-data-with-heuristics

Processed files are saved to:

data/processed/

Setup

The project expects Python 3.12, uv, and just.

pip install uv
uv tool install rust-just

Install dependencies:

just install-all

Install Streamlit dependencies:

just install-streamlit

Training

Example commands:

just train-resnet18 30
just train-resnet50 30 32
just train-efficientnet-b0 30 32
just train-efficientnet-b1 30 32
just train-densenet121 2 8 5 32
just train-convnext-nano 25 32
just train-convnext-tiny

Compare models through MLflow:

just compare-models

Evaluate the final ensemble:

just eval-ensemble

Submission

Generate the submission:

just make-submission

The file is saved locally:

data/submissions/submission_ensemble.csv

Streamlit

Run the Streamlit interface:

just run-streamlit

The interface lets you upload an image and compare predictions from selected models. It includes the final ensemble as well as individual models.

Experiments

Experiments are logged to MLflow through DagsHub.

Link:

https://dagshub.com/YashinSergey/room_type_classifier/experiments

Main metrics:

• best_macro_f1
• best_accuracy
• best_train_loss
• best_val_loss
• best_epoch

Project structure

src/
  dataset.py
  dataloaders.py
  preprocess_data.py
  transforms.py
  infer_ensemble.py
  evaluate_ensemble.py

models/
  convnext_nano/
  convnext_tiny/
  densenet121/
  efficientNet/
  resnet18/
  resnet50/
  yolo/

streamlit/
  app.py

reports/metrics/
outputs/models/
data/

Model checkpoints are saved to outputs/models/.

Metrics are saved to reports/metrics/.