Our task was to train a convolutional neural network (CNN) to recognize the style of selected painters, and later on to determine the origin of a painting received as input. As for the second part of the semester using the received results, we decided to recreate the learnt painter styles in a way, that the network is able to produce style from noise defined by the obtained features of given paintings.
- Python 3.5
- Keras (1.1.1)
- TensorFlow
- Numpy
- Sklearn
- Scipy
- preprocess2.py - Processing the csv input, scaling and transforming the input images into tensors
- learn.py - Train the neural network using the input data
- train_info_modified_fix.csv - Data containing painter details
- /train_sample - Image set for painters
Parameters are modifiable from code only.
python learn.py
- csv_file_name - defined in preprocess2.py | cs_load function argument | path to the csv containing meta data
- location - defined in preprocess2.py | load_images function argument | path to the directory containing the training set images
The output is an array with the size equal to the input size, containing 10 floats for each element. These floats represent the probability that the painter on a given index is the creater of the painting.
[
[0.9, ... , 0.01],
[ ... ],
...
[ ... ],
[0.07, ... , 0.2]
]
- Python 3.5
- Keras (1.1.1)
- TensorFlow
- Numpy
- Sklearn
- Scipy
- hftest2.ipynb - Generate an image from the given style
Parameters are modifiable from code only.
Run hftest2.ipynb
- train_sample - defined in hftest2.ipynb | load_images function argument | path to the train images
- filename - defined in hftest2.ipynb | preprocess_image function argument | name of the base image for style transfer
- iteration number - the number of style transfer iterations
The output is an image whose base is the filename image and got some style features from the train images iteration number times.