AtharvIITB commited on
Commit
7f22769
·
verified ·
1 Parent(s): a66896a

Classifier heads for BEATs and AST encoders

Browse files
Files changed (3) hide show
  1. README.md +41 -0
  2. ast.npz +3 -0
  3. beats.npz +3 -0
README.md ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: mixed-see-card
4
+ library_name: meow-decoder
5
+ pipeline_tag: audio-classification
6
+ tags: [cats, bioacoustics, meow, audio-classification]
7
+ datasets: [CatMeows]
8
+ ---
9
+
10
+ # Meow Decoder classifier heads
11
+
12
+ Small logistic-regression heads that turn a meow's audio embedding into the likeliest situation:
13
+ waiting for food, enjoying attention, or stressed somewhere new. Used by
14
+ [meow-decoder](https://github.com/ATHARV-IIT-BOMBAY/meow-decoder), which downloads them automatically.
15
+
16
+ | File | Encoder | First guess right (unseen cats) | Top two | Calibration error | License |
17
+ |---|---|---|---|---|---|
18
+ | `beats.npz` | Earth Species Project `esp_aves2_sl_beats_all` | 65.9% | 89.1% | 2.2% | CC BY-NC-SA 4.0 (follows the encoder) |
19
+ | `ast.npz` | MIT AST fine-tuned on AudioSet | 61.8% | 84.1% | 4.7% | CC BY 4.0 (follows the CatMeows data) |
20
+
21
+ Always guessing the most common situation scores 50.2%. Scores are 5-fold cross-validation over whole cats
22
+ (`splits.json` in the code repo), so every test meow comes from a cat the head never saw.
23
+
24
+ ## Format
25
+
26
+ Plain NumPy arrays, no pickle: `classes`, `mean`, `scale`, `coef`, `intercept` (standardize, then multinomial
27
+ logistic regression), plus `X`, `y` (the 440 training embeddings, kept so the app can refit with your own cat's meows).
28
+
29
+ ```python
30
+ from meow_decoder import Decoder
31
+ Decoder("beats").predict("meow.wav")["results"]
32
+ ```
33
+
34
+ ## Training data
35
+
36
+ [CatMeows](https://zenodo.org/records/4008297) (Ludovico et al., 2020): 440 meows from 21 Maine Coon and European
37
+ Shorthair cats, CC BY 4.0. Audio is 8 kHz, so nothing above 4 kHz was heard.
38
+
39
+ ## Limits
40
+
41
+ Three situations only. It cannot detect pain or illness. Pet cats from one dataset; other cats may sound different.
ast.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0978d3ae2960cf914cfaf3bb8bc85f997a0e26fd20dd5d9e68782ebcdc0338ab
3
+ size 1266514
beats.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4c247a5b586a0f2c730a4ee8bd7ad60683c2e373ea4bd0cb65c845c50edf6d1c
3
+ size 1272342