Repository files navigation

Classifying images from Google Quick Draw dataset using customConvolutional Neural Network.

About the Google Quick Draw:

Quick, Draw! is an online game developed by Google that challenges players todraw a picture of an object or idea and then uses a neural network artificial intelligenceto guess what the drawings represent. The AI learns from each drawing, increasing itsability to guess correctly in the future. The game is similar to Pictionary in thatthe player only has a limited time to draw (20 seconds). The concepts that it guessescan be simple, like 'foot', or more complicated, like 'animal migration'. This game isone of many simple games created by Google that are AI based as part of a project knownas 'A.I. Experiments'.

Open Soure dataset is availablehere.

First make sure that you have the correct data. You can downloadit from the link above. I have used theNumpy bitmap files (.npy) format.

For my experiment I have chosen the following10 classes:

• Bear,
• Bee,
• Camel,
• Cat,
• Cow,
• Crab,
• Crocodile,
• Dog,
• Dolphin,
• Duck.

For our training we need the PNG files so we need to convert them from the numpy format. To do this,firstly create a virtualenv:

$ virtualenv venv

Enter created virtualenv:

$source venv/bin/activate

Install needed dependencies:

$ (venv) pip install -r requirements.txt

Now we can useconvert_to_png.py script. It takes the datafromin folder and converts the defined number of numpy filesto the PNG images.

For my experiment I have decided to have:

• 20 000 images intraining set,
• 4 000 images invalidation set,
• 1 000 images intesting set.

To convert our data, simply run:

$ (venv) python convert_to_png.py

To start training the network simply run:

$ (venv) python training/convolutional.py<batch_size>

You can view the training and validation accuracies andlosses in the TensorBoard. During each training the datais saved to theruns directory.

During the training of after it, you can simply start up theTensorBoard with the following command:

$ (venv) tensorboard --logdir runs

To get this kind of visualisation, simply head to theClassificationVisualisation.ipynb

I have conducted three experiments to see how would the network train with different values of thelearning rate andbatch size.

Tested the following combinations:

• Learning rate=0.01 witch batch sizes=32, 64, 128, 256, 512
• Learning rate=0.001 witch batch sizes=32, 64, 128, 256, 512
• Learning rate=0.0001 witch batch sizes=32, 64, 128, 256, 512

Shared parameters during the experiments:

• Optimizer:Adam
• Epochs:50
• Dropout:0.5

Test accuracy:

Summary: In this experiment we can see that choosing such large large learning rate makes itunable to actually hit a global minima. We can then observe the outcome of this: every model below the 512batch size was returning the same class which (when we have 10 classes) results in exactly 10% accuracy.

Test accuracy:

Summary: In this experiment we can see that decreasing the learning rate 10x have a massive effect for the testaccuracy.

Test accuracy:

Summary: In this experiment we can observe that decreasing learning rate even more did not improved the testaccuracy dramaticaly, but it did change the plots. They are much smoother due to fact that we are making much smallersteps than before when searching for a global minima.