Neuron Plugin for TensorBoard Release Notes¶
Table of Contents
Known Issues and Limitations - Updated 5/28/2021¶
The following are not limitations in the Neuron plugin, but may affect your ability to use TensorBoard.
In the DLAMI TensorFlow-Neuron Conda environment, attempting to run
tensorboardwith the existing version of tensorboard will fail. Please update the tensorboard version before installing the Neuron plugin by runningpip install tensorboard --force-reinstall.
Traceback (most recent call last):
File "/home/ubuntu/anaconda3/envs/aws_neuron_tensorflow_p36/bin/tensorboard", line 7, in <module>
from tensorflow.tensorboard.tensorboard import main
ModuleNotFoundError: No module named 'tensorflow.tensorboard'
[2.0.29.0]¶
Date: 4/30/2021
Summary¶
First release Neuron plugin for TensorBoard. Check out it out here: Getting Started: Neuron Plugin for TensorBoard.
The Neuron plugin is now compatible with TensorBoard 2.0 and higher, in addition to TensorBoard 1.15
Provides a centralized place to better understand execution using Neuron SDK.
Continues support visualization for TensorFlow graphs, with support for PyTorch and MXNet coming in future releases.
Neuron plugin for TensorBoard is supported for Neuron tools >= 1.5, which is first introduced in Neuron v1.13.0 release
TensorBoard-Neuron is deprecated, and only supported for Neuron tools <= 1.4.12.0. The final version, 1.4.12.0 is part of Neuron v1.12.2 release.
[1.15.0.1.0.615.0]¶
Date: 11/17/2020
Summary¶
Fix issue with viewing chrome trace in Neuron profile plugin in Chrome 80+.
Resolved Issues¶
Updated dependencies to polyfill missing APIs used by chrome trace in newer browser versions.
[1.15.0.1.0.491.0]¶
Date 6/11/2020
Summary¶
Fix issue where utilization was missing in the op-profile view.
Resolved Issues¶
The op-profile view in the Neuron Profile plugin now correctly shows the overall NeuronCore utilization.
[1.15.0.1.0.466.0]¶
Date 5/11/2020
Summary¶
Fix potential installation issue when installing both tensorboard and tensorboard-neuron.
Resolved Issues¶
Added tensorboard as a dependency in tensorboard-neuron. This prevents the issue of overwriting tensorboard-neuron features when tensorboard is installed after tensorboard-neuron.
Other Notes¶
[1.15.0.1.0.392.0]¶
Date 3/26/2020
Summary¶
Added ability to view CPU node latency in the Graphs plugin and the Neuron Profile plugins.
Major New Features¶
Added an aggregate view in addition to the current Neuron subgraph view for both the Graphs plugin and the Neuron Profile plugin.
When visualizing a graph executed on a Neuron device, CPU node latencies are available when coloring the graph by “Compute time” using the “neuron_profile” tag.
The Neuron Profile plugin now has an overview page to compare time spent on Neuron device versus on CPU.
Other Notes¶
Requires Neuron-RTD config option “enable_node_profiling” to be set to “true”
[1.15.0.1.0.366.0]¶
Date 02/27/2020
Summary¶
Reduced load times and fixed crashes when loading large models for visualization.
Resolved Issues¶
Enable large attribute filtering by default
Reduced load time for graphs with attributes larger than 1 KB
Fixed a fail to load graphs with many large attributes totaling more than 1 GB in size
[1.15.0.1.0.315.0]¶
Date 12/20/2019
Summary¶
No major chages or fixes. Released with other Neuron packages.
[1.15.0.1.0.306.0]¶
Date 12/1/2019
Summary¶
Major New Features¶
Resolved Issues¶
Known Issues & Limits¶
Same as prior release
Other Notes¶
[1.15.0.1.0.280.0]¶
Date 11/29/2019
Summary¶
Initial release packaged with DLAMI.
Major New Features¶
N/A, initial release.
See user guide here: https://github.com/aws/aws-neuron-sdk/blob/master/docs/neuron-tools/getting-started-tensorboard-neuron.md
Resolved Issues¶
N/A - first release
Known Issues & Limits¶
Must install TensorBoard-Neuron by itself, or after regular TensorBoard is installed. If regular Tensorboard is installed after TensorBoard-Neuron, it may overwrite some needed files.
Utilization missing in Op Profile due to missing FLOPs calculation (see overview page instead)
Neuron Profile plugin may not immediately show up on launch (try reloading the page)
Graphs with NeuronOps may take a long time to load due to attribute size
Instructions that cannot be matched to a framework layer/operator name show as “” (blank)
CPU Usage section in chrome-trace is not applicable
Debugger currently supports TensorFlow only
Visualization requires a TensorFlow-compatible graph