Skip to content

Parameter generation - #98

Merged
bmcfee merged 7 commits into
mainfrom
parameter-generation
Jun 29, 2022
Merged

bmcfee merged 7 commits into
mainfrom
parameter-generation

Conversation

@bmcfee

@bmcfee bmcfee commented Jun 28, 2022

Copy link
Copy Markdown
Owner

This PR resolves #96 by including a reproducible script to construct filters subject to specified constraints. (Note to self: this should seed the RNG.)

It also addresses #97 somewhat by updating our pre-computed filters.

@bmcfee

bmcfee commented Jun 29, 2022

Copy link
Copy Markdown
Owner Author

Quick check on the performance for the new filters using the sine sweep benchmark code from #75 .

Here's the 0.2.2 performance for best and fast:

image
image

And for the updated 0.3.0 versions:

image
image

Colormaps are designed to be neutral hue at -96dB, anything colored blue is below that level.

@bmcfee
bmcfee merged commit 5641595 into main Jun 29, 2022
@bmcfee
bmcfee deleted the parameter-generation branch June 29, 2022 13:52
plakal added a commit to tensorflow/models that referenced this pull request Jul 26, 2023
Fixed vggish_smoke_test failures due to resampy changes.

bmcfee/resampy#98 seems to have updated
resampy's default filters in a way that made the VGGish smoke
test's embedding outputs drift too far from the expected values.

Fixed the test by not resampling in the critical path of generating
an embedding.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Improve default filters and make the whole process reproducible

1 participant