Google Colab

Processing Power for Deep Learning

• GPUs (graphical processing units) allow parallel processing that greatly enhances the speed of neural networks.

• We will use the GPU's available on Google Colab to train DeepLabCut (DLC) neural nets against labeled training sets and to analyze novel video against the trained network. 

• We will also use Colab to run OpenPose (unsupervised) to compare its output to the supervised output of DLC.

Using Colab

• HUGS interns will have access to a Colab Pro account through CV@hugs-lab.org  

• You can learn about Colab  by loading the notebook linked below and exploring its resources.

Making the Most of Your Colab Subscription

Training in Colab With DeepLabCut

• As described in the DeepLabCut (DLC)  installation instructions, we will use the DLC GUI to label infant points of interest then we will use that labeled file to train the algorithm using the GPU power of Colab. 

• See the link below for a  demo DLC Jupyter Notebook that you can copy into CV's Colab folder (or a Colab folder on your personal Google Drive) and run against your data.

https://github.com/DeepLabCut/DeepLabCut/blob/master/examples/README.md

Working Smart with Colab

• Even given a fairly powerful GPU,  running deep learning code on Colab can take a long time. 

• We will want to adopt some standard working procedures such as saving training iterations so they can be picked back up again and perhaps editing videos to smaller chunks to make sure a process you start (particularly if you are physically in the lab) finishes before you are ready to leave. 

• With Colab Pro, your process will abort when you close your browser so you need to think ahead and plan your work so that doesn't happen!