pdjota/arty-cnn-rnn

Arty CNN-RNN

Multi-head ResNet-50 backbone with column pooling, a bidirectional LSTM over spatial strips, and linear heads for genre, style, and artist on a WikiArt subset (pdjota/artyset).

Evaluation (test split)

Metric	Value
Genre (top-1)	67.80%
Style (top-1)	65.05%
Artist (top-1)	57.71%
Artist (top-5)	88.47%

• Checkpoint (local eval): checkpoints/cnnrnn/best.pt — on Hub this repo typically ships as best_model.pt.
• Arch: cnnrnn
• Epoch (from checkpoint): 0
• Test images: 2438

Files on this model repo

Typical layout after upload:

• best_model.pt — PyTorch checkpoint (model_state_dict, n_genre / n_style / n_artist, optional arch)
• genre_id2label.json, style_id2label.json, artist_id2label.json — class index → label for demos

Limitations

Not for production attribution or forensic ID; academic / demo use.