earthspecies / aves Goto Github PK

Hi,

could you please show how to extract the attention weights learned for each token in sequences ?

Suppose I want to fine-tune AVES on vocalizations of one species, giving in input a sequence of utterances, and in output multiple classes (i.e. observations).
I want to inquire the attention weights learned for the utterances.

Can you please help and show a code snippet for doing so, so that I can have a reference ?

Hi - I pulled and followed the instructions to set up this repo locally using jupyter notebooks/conda package manager. I'm able to import fairseq in my notebook, but attempting to use or import any of the modules results in a module not found error. The issue occurs when using the sample aves class provided in the repo. Thank you!

Hi! Thank you for making this code open-source!

I tried to replicate the clustering on this dataset :
https://github.com/YashNita/Animal-Sound-Dataset

But only obtained a correct cluster for birds (kush-part1) as you can see from the image.
Is it because the bio-aves pre-trained model has not been exposed to the animals included here?
(Lion / Donkey / Cow / Cat / Dog / Sheep / Frog / Bird / Monkey / Hen)

Hi @mhagiwara, I successfully installed fairseq-0.12.2, but jupyter lab cannot loading it.

Have you met this issue ?
Could you help? I am trying to use the classifier demo.

I tried to install fairseq as per instruction, but having issue in loading:
import fairseq.checkpoint_utils

ImportError: cannot import name 'metrics' from 'fairseq' (unknown location)

I tried here :
OFA-Sys/OFA#305

but no success.

Hi, I looked at BEANS as well, but not clear how to use the model for a classification tasks, and for fine tuning.

Consider the classification of wav files, with multiple classes, of a species, or of a specific specimen.

Can you show how to fine tune and classify?

Example:

Train
file | labels
file1.wav ['stress', 'mating', 'addressee_1']
file2.wav ['feeding', 'stress', 'addressee_2']

...

Predict
test.wav > ['stress', 'feeding' , 'addressee_3'] # number of labels may vary