ResNet50 on Art Composition Attributes

Fine-tunes a ResNet50 (pretrained on imagenet) network by training on WikiArt images labeled with eight art composition attributes. Used with https://github.com/hollygrimm/cyclegan-keras-art-attrs to generate art.

Please read the accompanying blog post: https://hollygrimm.com/acan_final

Requirements

• keras
• scikit-learn
• pillow

AWS Install

• Select Deep Learning AMI (Ubuntu) Version 13.0
• Instance Type GPU Compute such as p2.xlarge
• 125GB sda1

Connect to instance, copy contents of acan-aws-setup.sh to file in /home/ubuntu and run:

vi acan-aws-setup.sh
chmod +x acan-aws-setup.sh
./aws-setup.sh

Manual Install

Download Dataset

download test.tgz and train.tgz from https://github.com/zo7/painter-by-numbers/releases/tag/data-v1.0

cd data
tar -xvf test.tgz
tar -xvf train.tgz

Label Data with Attributes

Example attribute data has been supplied for four examples in all_domain.csv. For best results, modify all_domain.csv and label more images with attributes.

input_params.json Configuration

• base_lr: float learning rate default is 1e-04
• optimizer: either adam or adagrad
• batch_size: integer batch size appropriate for your GPU size
  • 30 for 8GB GPU
• nb_epoch: integer number of epochs
• validation_split: float split training and validation set

Run Training

source activate tensorflow_p36
cd art-composition-cnn/
python main.py -c input_params.json

Tensorboard

source activate tensorflow_p36
cd art-composition-cnn/experiments/
tensorboard --logdir=.

Run Inference on Validation Samples

Update weights_path with selected hdf5 from training:

vi input_params_for_inference.json

Run inference:

python
import main
main.infer()

Run Tests

cd tests
python art_composition_cnn_tests.py