diff options
Diffstat (limited to 'tensorflow/contrib/tfprof/README.md')
-rw-r--r-- | tensorflow/contrib/tfprof/README.md | 46 |
1 files changed, 22 insertions, 24 deletions
diff --git a/tensorflow/contrib/tfprof/README.md b/tensorflow/contrib/tfprof/README.md index 4fa1ccea69..eefd88793e 100644 --- a/tensorflow/contrib/tfprof/README.md +++ b/tensorflow/contrib/tfprof/README.md @@ -1,26 +1,24 @@ # tfprof: TensorFlow Profiler and Beyond -# Full Document in tensorflow/tools/tfprof/README.md - -Author: Xin Pan (xpan@google.com, github: panyx0718), Jon Shlens, Yao Zhang - -Consultants: Jon Shlens, Pete Warden - -###Major Features - -1. Measure model parameters, float operations, tensor shapes. -2. Profile op execution times, requested memory size and device placement. -3. Inspect checkpoint tensors' shapes and their values. -4. Selectively group, filter, account and order ops. - -####tfprof supports 3 views to organize TensorFlow model profiles - - * code view: Stats are associated your Python codes and organized as call stacks. - * scope view: Stats are organized as name scope hierarchies. - * graph view: Stats are organized as Tensorflow Op graph. - -####For each view, there are 3 ways to display outputs: - - * stdout: Results are written to stdout. - * timeline: Visualized in chrome browser as time series. - * file: Results are dumped to file. +<h1>Please use `tf.profiler.xxx` instead of `tf.contrib.tfprof.xxx`</h1> +<h1>Full Document in tensorflow/core/profiler/README.md<h1> + +###Features + +* Profile model architectures + * parameters, tensor shapes, float operations, device placement, etc. +* Profile model performance + * execution time, memory consumption + * Profile multiple steps. +* Auto profile and advise. + * accelerator utilization check + * expensive operation check + * operation configuration check + * distributed runtime check (Not OSS) + +###Interfaces + +* Python API +* Command Line +* Visualization +* C++ API (Not public, contact us if needed.) |