TensorFlow Eager versus PyTorch
Quick Benchmark: Linear Regression
With the release candidate of TensorFlow 1.7, TensorFlow Eager (TFE) will soon be ready for widespread use (https://github.com/tensorflow/tensorflow/releases). TFE was likely designed in response to PyTorch's popular and user-friendly software development experience. But, how does TFE's performance compare to PyTorch? As a quick experiment, two Jupyter notebooks included in this repository, based off of a TFE example, were run on a desktop machine with the following hardware and software:
- Hardware:
- NVIDIA Titan X (Pascal)
- Software:
- Operating System: Ubuntu 16.04.4 LTS (GNU/Linux 4.13.0-37-generic x86_64)
- NVIDIA driver 390.48
- CUDA 9.1.85-1
- cuDNN 7.1.2
- TensorFlow 1.7.0 COMPILED FROM SOURCE
- PyTorch 0.3.1 COMPILED FROM SOURCE