EffNet heads from 2022-08-25 declare input 1200; newer heads and EffNet embedding are 1280

[xiaden]

All EffNet classifier sidecars from 2022-08-25 specify a shape of 1200 while newer heads (nsynth*, approachability/engagement) specify 1280 or batch_size,1280.

EffNet embedding on my end is 1280-D. Please confirm if the 1200 in the Aug-25 sidecars is a typo. If so, could the sidecars be republished with input 1280?

Models showing 1200 (all dated 2022-08-25): mood_acoustic, mood_aggressive, mood_electronic, mood_happy, mood_party, mood_relaxed, mood_sad, genre_electronic, danceability_classifier, gender, tonal_atonal, voice_instrumental.
Models showing 1280: nsynth_reverb, nsynth_instrument, nsynth_acoustic_electronic (2023-02-05), bright/dark_timbre (2023-11-04), approachability_regression (2022-06-16), engagement_regression (2022-06-16).

Why it matters:
Custom runners that validate shapes either fail or force users to truncate embeddings; if heads were trained on 1280, truncating to 1200 drops information. Specifically, I'm developing a docker image that uses the sidecars to automatically tag songs via an API.

Raw output of the json files looking for shape, and release data:

Get-ChildItem -Recurse **redacted (folder structure)** | ForEach-Object { $content = Get-Content $_.FullName -Raw | ConvertFrom-Json; $embedding = if ($content.inference.embedding_model) { $content.inference.embedding_model.model_name } else { 'EMBEDDING' }; $shape = if ($content.schema.inputs) { $content.schema.inputs[0].shape -join ',' } else { ($content.schema.outputs | Where-Object { $_.output_purpose -eq 'embeddings' }).shape -join ',' }; [PSCustomObject]@{ Type = if ($embedding -eq 'EMBEDDING') { 'EMB' } else { 'HEAD' }; Name = $content.name; ReleaseDate = $content.release_date; Shape = $shape } } | Sort-Object Type,ReleaseDate | Format-Table -AutoSize

Type Name                           ReleaseDate Shape
---- ----                           ----------- -----
EMB  EffnetDiscogs                  2022-02-17  64,128,96
HEAD nsynth reverb                  02/05/2023  batch_size,1280
HEAD nsynth instrument              02/05/2023  batch_size,1280
HEAD nsynth acoustic electronic     02/05/2023  batch_size,1280
HEAD nsynth bright/dark timre       11/04/2023  batch_size,1280
HEAD approachability_regression     2022-06-16  1280
HEAD engagement_regression          2022-06-16  1280
HEAD genre electronic               2022-08-25  1200
HEAD mood sad                       2022-08-25  1200
HEAD mood relaxed                   2022-08-25  1200
HEAD mood party                     2022-08-25  1200
HEAD mood happy                     2022-08-25  1200
HEAD mood electronic                2022-08-25  1200
HEAD mood aggressive                2022-08-25  1200
HEAD mood acoustic                  2022-08-25  1200
HEAD gender                         2022-08-25  1200
HEAD danceability classifier        2022-08-25  1200
HEAD tonal/atonal                   2022-08-25  1200
HEAD voice/instrumental classifiers 2022-08-25  1200

[palonso]

Hi @xiaden, thanks for noticing this issue!

EffNet embedding on my end is 1280-D.

I confirm the EffNet embedding dimension is 1280.

Please confirm if the 1200 in the Aug-25 sidecars is a typo. If so, could the sidecars be republished with input 1280?

Yes, after inspecting the models, I confirm that this is a typo in the associated metadata files. The models were correctly trained on the 1280 dimensions of the EffNet embedding and don't require further modification. For example,

Mood sad architecture

I'll fix the metadata files and let you know when they're ready