Cloud TPU Alpha Access

rolyan_trauts · May 11, 2020, 10:45pm

If have just been running

tensorflow/docs/blob/master/site/en/r1/tutorials/sequences/audio_recognition.md

# Simple Audio Recognition

This tutorial will show you how to build a basic speech recognition network that
recognizes ten different words. It's important to know that real speech and
audio recognition systems are much more complex, but like MNIST for images, it
should give you a basic understanding of the techniques involved. Once you've
completed this tutorial, you'll have a model that tries to classify a one second
audio clip as either silence, an unknown word, "yes", "no", "up", "down",
"left", "right", "on", "off", "stop", or "go". You'll also be able to take this
model and run it in an Android application.

## Preparation

You should make sure you have TensorFlow installed, and since the script
downloads over 1GB of training data, you'll need a good internet connection and
enough free space on your machine. The training process itself can take several
hours, so make sure you have a machine available for that long.

## Training

This file has been truncated. show original

On my poor old I7-3770K @4.3Ghz which is actually still running but already hit a validation accuracy of 89%.
Forgot what time I started this thing but prob more than 12 hours ago 16300 steps of the 18000 now done.

Its after running this cloud TPU Access starts to get tempting as we could be talking minutes rather than hours.
The audio_recognition demo is pretty cool and not optimised in fact for a KWS the more labels you have the harder it is to attain accuracy and would of been interesting if the 2 main labels where ‘key word’ and ‘not key word’.
Which I will try when patience allows and with more patience see how much difference adding my voice and keyword adds to overall accuracy.

That got me thinking that it is in the realms of possibility to maybe run in a minimum donation model a online custom model trainer for a Rhasspy KWS.
You can hire TPUs hourly and its not as expensive as you might think.

PU type (v2)	v2 cores	Total memory	On-demand price (USD)	Preemptible price (USD)
v2-8	8	64 GiB	$4.50 / hour	$1.35 / hour
TPU type (v3)	v3 cores	Total memory	On-demand price (USD)	Preemptible price (USD)
v3-8	8	128 GiB	$8.00 / hour	$2.40 / hour

Dunno how long what I have done would take, but likely minutes rather than hours, but as said dunno.

rolyan_trauts · May 12, 2020, 2:40pm

PS I worked out the sign in to the nvidia cuDNN Download is free and open and now running on my old GTX780 which is ticking away much, much faster than than the I7 cpu.
I may try and pick another one up as they do turn up for approx £60.