elanzini / SketchGAN

Teaching a computer to sketch objects using GANs and Quick, Draw! dataset

SketchGAN

Check out the article on Medium

Environment

Set up a new environment with the necessary dependencies:

# Create a new environment
conda create -n py37 python=3.7

# Activate the environment
conda activate py37

# Install dependencies from requirements.txt file
pip install -r requirements.txt

Data

The Quick, Draw! dataset is publicly available and here you can find the instructions on how to download the data locally. The load_data module is available in order to handle the .ndjson drawing files.

Here is a quick step-by-step guide to download the data locally:

1. Download gsutil following the instructions here

2. Find the dataset on Google Cloud Console and navigate to /full/simplified/

3. Browse all the different categories available and pick one

4. Run gsutil -m cp gs://quickdraw_dataset/full/simplified/[category].ndjson [path-to-store]

   For instance, the follwing command will download the circle.ndjson file in your current directory:

   gsutil -m cp gs://quickdraw_dataset/full/simplified/circle.ndjson .

Make sure the .ndjson file corresponding to the category is in your root folder.

Model

GANs implementation in Keras, available here.

Here you can see a summary of the architecture:

Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
flatten_1 (Flatten)          (None, 784)               0         
_________________________________________________________________
dense_1 (Dense)              (None, 512)               401920    
_________________________________________________________________
leaky_re_lu_1 (LeakyReLU)    (None, 512)               0         
_________________________________________________________________
dense_2 (Dense)              (None, 256)               131328    
_________________________________________________________________
leaky_re_lu_2 (LeakyReLU)    (None, 256)               0         
_________________________________________________________________
dense_3 (Dense)              (None, 1)                 257       
=================================================================
Total params: 533,505
Trainable params: 533,505
Non-trainable params: 0
_________________________________________________________________
Model: "sequential_2"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_4 (Dense)              (None, 256)               25856     
_________________________________________________________________
leaky_re_lu_3 (LeakyReLU)    (None, 256)               0         
_________________________________________________________________
batch_normalization_1 (Batch (None, 256)               1024      
_________________________________________________________________
dense_5 (Dense)              (None, 512)               131584    
_________________________________________________________________
leaky_re_lu_4 (LeakyReLU)    (None, 512)               0         
_________________________________________________________________
batch_normalization_2 (Batch (None, 512)               2048      
_________________________________________________________________
dense_6 (Dense)              (None, 1024)              525312    
_________________________________________________________________
leaky_re_lu_5 (LeakyReLU)    (None, 1024)              0         
_________________________________________________________________
batch_normalization_3 (Batch (None, 1024)              4096      
_________________________________________________________________
dense_7 (Dense)              (None, 784)               803600    
_________________________________________________________________
reshape_1 (Reshape)          (None, 28, 28, 1)         0         
=================================================================
Total params: 1,493,520
Trainable params: 1,489,936
Non-trainable params: 3,584

Usage

python gan.py

Results

Circle

Airplane

Cat