Deep Learning GPU CODER

I have question about Deep Learning Training, GPU Coder and Precision

1. We are working on a project that trains data on pre-trained models such as Resnet50 and AlexNet. I would like to use GPU CODER because I heard that using Tensor Core can achieve faster speed efficiency. However, when I looked at the examples related to GPU CODER, there were only examples of data validation using the model. I wonder if it is impossible to train data on a pre-trained model using GPU CODER. If possible, I'd like to ask if there are any relevant examples
2. I know that MATLAB uses Single Precision by default when Deep Learning Training. How to use Double Precision or Half Precision?

Answer 1

[Sayan Saha]

Hi Kyle,

GPU Coder only supports inference using deep learning networks currently. Training the network using coder products is not yet supported. Given that you are using pre-trained networks, you can leverage transfer learning capability in MATLAB to re-train the networks for your datasets.

Training the network can only be done using single precision in MATLAB currently. Can you elaborate on your requirements to train using double and half precision? I understand that half-precision is likely to improve the training time. Are there any other reasons?

For inference we do support quantized datatype with CuDNN and TensorRT libraries as documented in: https://www.mathworks.com/help/deeplearning/ug/quantize-a-residual-network-trained-for-image-classification-and-generate-cuda-code.html

https://www.mathworks.com/help/gpucoder/ug/tensorrt-target.html

With TensorRT library half-precision is also supported: https://www.mathworks.com/help/gpucoder/ug/code-generation-using-tensorrt.html#mw_c1f34b7f-7d7a-4fd8-ad98-be0d4212118e

~Sayan

[Kyle Lee]

RTX 3090 is faster than A100 on single precision. On double and half precision, however, A100 is faster. So I am thinking about which one to use between RTX 3090 and A100. Does MATLAB have a plan to supported double precision for training network?