Manas Sahni
Introduction If you’ve trained deep learning models, you know the process can take hours or days (weeks?) and thousands of dollars’ worth of computation. With increasing use of DNNs in common production, this problem only gets bigger – they need to be used on diverse deployment targets with widely varying latency constraints, based on hardware capabilities and application requirements. Designing…
The emergence of CNNs in mainstream deployment has necessitated methods to design and train efficient architectures tailored to maximize the accuracy under diverse hardware & latency constraints. Designing and training DNN architectures for each deployment target is not feasible. Each deployment costs training time, compute dollars, system expertise, ML expertise, CO2 emissions. In CompOFA , we…
On my not-too-shabby laptop CPU, I can run most common CNN models in (at most) 10-100 milliseconds, with libraries like TensorFlow. In 2019, even a smartphone can run “heavy” CNN models (like ResNet) in less than half a second. So imagine my surprise when I timed my own simple implementation of a convolution layer and found that it took over 2 seconds for a single layer! It’s no…
This post was featured on Heartbeat by Fritz and DL Weekly Francois Chollet puts it concisely: - make it possible - make it work - make it efficient - make it dependable and invisible - move on to the next layer and never think about it again — François Chollet (@fchollet) April 15, 2018 For many deep learning problems, we’re finally starting with the “make it efficient” stage. We had…