HamzaHawk/cnn_classification

Image classification using a convolutional neural network

Image Classification using a Convolutional Neural Network

This attemps to classify different types of birds coming from the Caltech-UCSD Birds-200-2011 dataset. The dataset includes 200 classes of different birds. The implementation uses Tensorflow compiled for use on a GPU. Rudementary training is complete, although validation is not.

Data Management

The scripts in the utils folder help organize the data. The data must be stored in a directory (known as data_dir) that is specified in utils/config.py. The variables defined in config.py are also used in other files throughout the project. In short, in data_dir there should be folder seperated datasets, each with a folder called images that contains folder seperated sets of images, the folder mapping to the label for the images in it. Below is an example for how the mnist dataset would be structured.

data_dir:

• mnist:
  • README.txt
  • images:
    • number_1:
      • num_1.png
      • image_name.png
      • ...
    • number_2:
      • ...
    • ...

create_test_train_val.py

This seperates your data using symlinks into train, test, and val sets in a 80:10:10 split respectively. These can be edited in config.py. The script will create folders train, test, and val if already not created in the root directory of your dataset. The mnist example would look like this. Run an ls -l on a file in one of the folders to see that it symlinks to the correct file.

data_dir:

• mnist:
  • README.txt
  • train:
    • number_1
      • image.png
    • number_2:
      • image.png
      • ...
  • test:
    • same as train
  • val:
    • same as test and train
  • images:
    • number_1:
      • num_1.png
      • image_name.png
      • ...
    • number_2:
      • ...
    • ...

createBirdRecords.py

Tensorflow expects its data in the form of TFRecords. This script will go through the train, test, and val folders and create a TF Record for each folder. The resulting records will be stored in the records directory in data_dir as train.tfrecord, test.tfrecord, and val.tfrecord respectively.

Tensorboard

Tensorboard can be viewed at http://localhost:6006 by running tensorboard --logdir=/home/fabbric/data_dir/birds_2011/models