thepetrmarek / automaticneuralimagecropper

Neural network making the best looking crops of images

Automatic Neural Image Cropper

Neural network trained to make the best looking crops of images. It is described in the blog .

How to run

1. Download checkpoint of Inception v4 from https://github.com/tensorflow/models/tree/master/slim#pre-trained-models and unpack it to root.

2. Install all requirements by running

   pip install requirements.txt

3. Find some data (you have to do it on your own. AVA dataset maybe...). It is JSON containing array with objects: {"picture":"path/to/picture","good_example":True} Method to decide if picture is good or bad is described in blog . You can take inspiration from .

4. Open . There are several parameters at the beginning of the file.

Parameter	Function
TRAIN	Run training or run cropping (Bool)
EPOCHS	How many epochs will we train? (Int)
BATCH_SIZE	Size of batch (Int)
TRAIN_ACCURACY	Do we want to evaluate accuracy on trainig set? This will make training A LOT slower. (Bool)
MODEL_NAME	Name of model (String)
RESTORE	Will we restore the model from checkpoint? (Bool)
CHECKPOINT	Path to checkpoint (String)
EPOCH	From which epoch to continue? (Int)
[training|validation|testing]_dataset	Path to folder with dataset and name of json file (String, String)
IMAGE_FOLDER	Path to folder with photos to crop (String)

4. Set parameters to train and run training by

   py -3 Main.py

5. Set parameters to cropping and run cropping by

   py -3 Main.py

Examples