HW9_code_description.txt

The script tests different models for three different kinds of precisions, namely float, half, and double. To benchmark the models some default values are used, such as number of tests = 50, batch size = 12, number of classes = 1000, number of GPU = 1, etc. A random dataset of size 3x224x224 is created to do the benchmarking.

To test the performance running time was considered for training and inference for different models along with different precisions.

This script was run on a machine with an NVIDIA TITAN Xp GPU and 8 CPUs. The CUDA version was 11.6, and cudnn version was 8303. To complete the whole run it tool=k 1 hour 8 minutes and 21 seconds. The smallest recorded time was ~ 10 ms for a model while the highest recoded time for a model was over 2000 ms.