Seasonal Color Classifier

Image credit: The Concept Wardrobe

This project implements a 12-class seasonal color analysis classifier specialized for face images, using state-of-the-art models via the timm library. It is an ongoing struggle on finding a model architecture that will fit the Deep Armocromia Dataset and might one day be the basis for a reliable armocromia classification app.

Project Structure

• data/: Data loading and augmentation modules.
• models/: Model architecture definitions (using timm).
• training/: Training loop, loss functions, and checkpointing (with optional fine-tuning).
• evaluation/: Evaluation metrics and analysis.
• visualization/: Plotting functions for training history, confusion matrices, etc.
• utils/: Helper functions.
• config.py: Configuration parameters (updated hyperparameters, augmentation settings, etc.).
• main.py: Main script for training, evaluation, and optional hyperparameter optimization.
• requirements.txt: Project dependencies (includes timm).
• viz/: Directory for generated plots.
• checkpoints/: Directory for model checkpoints.

Supported Models

The following timm model keys are supported (if you pass a partial key, it may be automatically completed):

• hgnetv2_b5.ssld_stage2_ft_in1k
• vit_base_patch16_clip_224.openai_ft_in12k_in1k
• tf_efficientnetv2_l.in21k_ft_in1k
• hgnetv2_b5.ssld_stage1_in22k_in1k
• hgnet_base.ssld_in1k
• coatnet_2_rw_224.sw_in12k_ft_in1k
• convformer_m36.sail_in22k_ft_in1k
• maxvit_base_tf_512.in1k
• tf_efficientnetv2_xl.in21k_ft_in1k
• convnextv2_huge.fcmae_ft_in1k
• vit_base_patch8_224.augreg2_in21k_ft_in1k
• vit_mediumd_patch16_reg4_gap_256.sbb_in12k_ft_in1k

The model is instantiated via timm with pretrained weights (if enabled) and its classifier head is adjusted to output 12 classes.

Usage

1. Install dependencies:

   pip install -r SeasonalColourClassification/requirements.txt
   

2. Prepare your dataset in dataset/images/train and dataset/images/test (or update paths in config.py).
3. Set your NEPTUNE_API_TOKEN as an environment variable.
4. Run training (with optional HPO):

   python -m SeasonalColourClassification.main --model tf_efficientnetv2_l.in21k_ft_in1k --epochs 250
   

   or with HPO:

   python -m SeasonalColourClassification.main --model vit_base_patch16_clip_224 --epochs 250 --hpo
   

5. Monitor your Neptune dashboard for training parameters, plots, and checkpoints.

License

This project is open source and available under the MIT License.

Acknowledgements

• PyTorch & Torchvision: For providing deep learning frameworks and pre-trained models.
• NumPy, Matplotlib, Seaborn, Scikit-Learn: For data manipulation and visualization.
• TQDM: For progress bar integration during training.
• Special thanks to lorenzo-stacchio for the creation of this dataset.