Self Help

• • • Specific values

for var in {list}

do

#body

done

• Ranges

for var in {start..stop..step}

do

#body

done

Use this if most of the original IP is the same. This may happen if there's a power outtage and the IPs within a subnet have been reassigned.

nmap -p <original port> <sub.net.mask>.*

These instructions are reproduced from here

1. Create an .ssh directory on the remote machine. This is where the 'authorized keys' for the remote machine will live.

a@A:~> ssh b@B mkdir -p .ssh

Use the -p option to avoid any errors if the .ssh directory already exists.

2. Create a public key if required and copy it into the remote .ssh directory.

a@A:~> cat .ssh/id_rsa.pub | ssh b@B 'cat >> .ssh/authorized_keys

3. Now, sign-ins can be completed without a password.

a@A:~> ssh b@B

The experimental compiler interface in tensorflow only dumps graphs in Dot format after optimization. There doesn't seem to be a way to get the dotfile for the unoptimized graph. This is a workaround.

The XLA compiler ships with a bunch of tools that help with inspection. One of these is the interactive_graphviz utility.

Pre-reqs

• bazel

• A cloned tensorflow repo

• An HLO dump in a .txt file (hlo.txt)

Steps

1. Navigate to the xla folder

cd path/to/tensorflow/compiler/xla

2. Use bazel to build the tools you need. (interactive_graphviz) in this case. A list of available tools here.

bazel build tools:interactive_graphviz

This step may take a while

3. Use the utility (it should be in the bazel-bin directory)

tensorflow/bazel-bin/tensorflow/compiler/xla/tools/interactive_graphviz --hlo_text="opt_pool.txt"

4. This will launch a command line utility which lists available commands. My most used ones are:

   1. show_fusion_subcomputations [on|off]

   2. list computations

   3. <name of computation to render>