Painting Date Estimator

Goal of the project is developing a machine learning algorithm which is able to estimate the year in which a painting has been created, based on its appearance only.

This code has been developed as Capstone Project for the Coursera's Advanced Data Science with IBM Specialization.

Visualizing the results

Results can be visualized using this Notebook.

Installing the dependencies

From the command prompt, type:

pip install -r requirements.txt

Testing the model

Pretrained model can be tested downloading the weights from this link. The weights should be unzipped in ./models_pretrained folder.

Then:

cd src
python ./test.py

Input image path, pretrained model path and model architecture can be modified. Please take a look at:

python ./test.py --help

Training the model

Downloading the data

Data can be downloaded from Painter by Numbers Kaggle competition, after registering to Kaggle and joining the competition.

Copy all_data_info.csv in ./data/ folder;
Extract the images in ./data/train and ./data/test folders.

Running the training script

Model can be trained calling the train.py script in ./src/ folder:

cd src
python ./train.py

Training parameters, input and output directories can be modified. Please take a look at:

python ./train.py --help

In particular, model can be trained fine tuning Resnet152 (default) or SqueezeNet. In case SqueezeNet is preferred, run:

python ./train.py --select_squeezenet=True

Monitoring training curves

Network training can be monitored via Tensorboard and it is updated once per epoch:

tensorboard --logdir=runs

matteobustreo/painting-date-estimator