arraypress commited on
Commit
9445269
·
verified ·
1 Parent(s): 64a505d

Card: w2 is [39x512], outputs by inputs

Browse files
Files changed (1) hide show
  1. README.md +3 -2
README.md CHANGED
@@ -28,8 +28,9 @@ sub_bass against bass_hit.
28
  (LAION, Apache-2.0), converted to Core AI with its weights unchanged; it turns
29
  the spectrogram into a 512-number embedding.
30
  3. `head-*.f32` — a small trained head on the L2-normalised embedding:
31
- 512 → 512 (GELU) → 39, softmax. Little-endian float32: `w1` [512×512], `b1`
32
- [512], `w2` [512×39], `b2` [39].
 
33
 
34
  Only the head was trained, on a licensed sample library; the audio is not
35
  distributed and cannot be recovered from these files. Apache-2.0, following CLAP.
 
28
  (LAION, Apache-2.0), converted to Core AI with its weights unchanged; it turns
29
  the spectrogram into a 512-number embedding.
30
  3. `head-*.f32` — a small trained head on the L2-normalised embedding:
31
+ 512 → 512 (GELU) → 39, softmax. Little-endian float32, row-major in PyTorch's
32
+ Linear layout (outputs × inputs): `w1` [512×512], `b1` [512], `w2` [39×512],
33
+ `b2` [39].
34
 
35
  Only the head was trained, on a licensed sample library; the audio is not
36
  distributed and cannot be recovered from these files. Apache-2.0, following CLAP.