This is a demo for CNN models training on Quick, Draw! dataset. Implemented with pytorch 🔥.
Quick, Draw! dataset is a collection of 50 million drawings across 345 categories, provided by googlecreativelab. The demo only uses at most 5000 samples from each of the 345 categories. In total, it is trained with 1380000 samples.
👇 Here are the step-by-step tutorials.
-
Clone the repo to your local device.
git clone https://github.com/XJay18/QuickDraw-pytorch.git
-
Download data from google and generate train&test dataset. You can run this command for example:
python ./DataUtils/prepare_data.py -c 10 -d 1 -show
💡 hint:
-c
for how many categories you want to download, available choices:10
,30
,100
,all
. Note thatall
is 345 categories.-d 1
means that download the data from internet, and-d 0
means that not download data and just generate train and test dataset from your pre-download data.-show
means that show some random images while generating the dataset.
-
Start training and evaluating for example.
python main.py --ngpu 0 -m convnet -e 5
🔑 Please refer to main.py to see the detailed usage.
-
Train a model in tf.keras with Colab, and run it in the browser with TensorFlow.js
-
pytorch2keras (may be used in future since the current demo is not deployed on web using the first Reference)
-
Devise or revise the current model to achieve higher accuracy on Quick, Draw!.
-
Enlarge the used dataset (i.e, choose more samples from each categories of the dataset).
-
Deploy the demo on web.
😃 I started this project with the purpose of improving my ability in coding quickly 🚀. Also the project will serve as a push on my way to learning more knowledge and experience from others⭐.