Doodle Prediction using CNN, Tensorflow, Keras.
https://www.youtube.com/watch?v=SHCC5mYxUUI&feature=youtu.be
preprocess.ipynb
preprocess_analysis.ipynb
data_model.ipynb
Data used for this project is the crowdsourced Quick-Draw data set which Google amassed using their online game called Quick Draw. The data set contains over 50 million drawings or doodles submitted by users. The drawings are classified into 345 different classes that range from 'aircraft carrier' to 'zig-zag'.
For data analysis and wrangling, only 20 classes of drawings have been considered, and first 10000 image samples were used for each category. This makes the analysis phase a bit quicker.
For the actual CNN data modelling, we decided to be ambitious and use all 345 classes to build a Convolutional Neural Network. However, only the first 10000 image samples were used for each category. As we will see later, 10000 images is not enough samples to make a prediction with great accuracy.
Entire dataset is available here: https://github.com/googlecreativelab/quickdraw-dataset
https://console.cloud.google.com/storage/browser/quickdraw_dataset/full/numpy_bitmap
For this project, we are working with preprocessed dataset provided by Google Creative Lab.
All the drawings have been rendered into a 28x28 grayscale bitmap in numpy .npy format.
Step1: Install gsutil: https://cloud.google.com/storage/docs/gsutil_install?hl=fr
Step2: gsutil -m cp gs://quickdraw_dataset/full/simplified/*.npy
Since the goal of this project is to build a CNN model using all 345 image classes, the data wrangling steps to achieve this were as follows:
- Created a dictionary for url and class labels to dowbload the entiry dataset in .npy files.
- Then the dataset was downloaded into a dictionary.
- Normalized the image pixels between 0 and 1, and added class labels.
- Aggregate the first 10,000 arrays for every doodle into one array of doodles.
- Split the data into features and target labels.
- Save the train and test data as serialised data using pickle files.
The steps listed above are described in the jupyter notebook titled preprocess.ipynb
I kept getting a 'Memory Error' trying to create a a dictionary object using the downloaded data. See below:
Running the code in colab crashed the application and produced a RAM error as well. We will have to take another approach to set up data for 345 samples to create our CNN model.
Following piece of code generates png image files from the .npy files and splits the image files into train and test data:
These images will be used to generate our CNN model to classify 345 images.
Ammount of test data:
Plotting an intensity histogram for some of the classes:
We attempted to build a CNN model with 345 classes. Architecture of our CNN model is below:
Model summary:
Model was fit as follows:
I did not have a GPU and as a result the model fitting took about 1 day and 50min to complete.
Performance:
As we can see, the CNN model did not perform that great. We ended with a validation accuracy of ~70%. Which isn't that great. Looks like including all 345 classes might have been too ambitious for 10,000 sample images.
One way to possibly improve the existing CNN model is by using Transfer Learning. More data can be fed to the existing CNN model to retrain it using saved weights. This will make training more efficient and produce better accuracy.
Another approach to making a better performing model is by using Recurrent Neural Networks to build the model instead of CNN.
We now believe that it may have been a little over-ambitious trying to train a model to classify 345 images with only 10000 samples each and no GPU available to quickly train models.