This repository contains analysis done in study Zhu, Ji, Zhang, Xu, Zhou, Chan, Arxiv, 2019. The study uses several state-of-the-art CNN architectures to explore how a machine may help perceive and quantify art styles, and explores (1) How accurately can a machine classify art styles? (2) What may be the underlying relationships among different styles and artists?
- Python 3.7
- Tensorflow (2.0.0)
- Numpy (1.18.1)
- Pandas (1.0.3)
- Matplotlib (3.1.3)
- Scikit-learn (0.22.1)
- Pillow (7.1.2)
- Download data from https://www.wikiart.org/ with relabeled train.csv and test.csv
First, put all downloaded Wikiart images at data/ALL. To train specific CNN models with specified hyperparameters, check out table below or python train_models.py --help for complete list of arguments.
| Arguments | Description |
|---|---|
| --input_dir | Input directory (str, required) |
| --output_dir | Output directory (str, required) |
| --model_type | Type of model used (str, required) |
| --batch_size | Batch size used (int, default = 150) |
| --learning_rate | Learning rate for training the model (float, default = 1e-5) |
| --image_size | Image size of model (int, default = 300) |
| --num_epochs | Number of epochs to train (int, default = 100) |
| --subset | Length to subset a portion of total data for training (int, default = 1000000) |
| --add_FC | Add fully connected layer before softmax for obtaining features? (store_true, default = False) |
| --n_threads | Number of threads (int, default = 1) |
Choices of CNN architectures:
- VGG16
- ResNet152
- ResNet152V2
- ResNeXt101
- InceptionV3
- InceptionResNetV2
- EfficientNet (B7)
- NASNet (Large)
We also implemented Grad-CAM for direct visualzation of where the model is paying attention to when making classification decision. Once the model is trained, one can run python vis_cam.py for generation of class activation maps. Table below lists some arguments to specify for generating the maps: