repeating results of audioset paper

[yunbin]

1. Is your result close to what published by the audioset paper? How do you define hit rate? Is it accuracy?
2. Is the feature you generated using vggish model close to their released feature? It seems very different. 3. They have quantized 8-bit feature and you used floating number feature? Is your model trained with their 8-bit feature?

[igor-panteleev]

Hi,

1. Can you provide link to "audioset paper" i will take closure look.
2. Model is trained with audioset 8-bit features. But I should mention that youtube-8m lib dequantize features back before feed them to model. So for final script i have modified vggish feature extractor to prevent unnecessary quantize/dequantize process and feed features directly.
   vggish quantize - here
   youtube-8m dequantize - here

[yunbin]

Thanks for the info. It turns out the release features don't match the code generated feature, such that the model doesn't work very well with audio data.

…

On Mon, Apr 2, 2018 at 8:26 AM, Igor Panteleev ***@***.***> wrote: Hi, 1. Can you provide link to "audioset paper" i will take closure look. 2. Model is trained with audioset 8-bit features. But I should mention that youtube-8m lib dequantize features back before feed them to model. So for final script i have modified vggish feature extractor to prevent unnecessary quantize/dequantize process and feed features directly. vggish quantize - here <https://github.com/tensorflow/models/blob/d640ab9cb1c06e7da39cca0399f14f8c6a54eb9c/research/audioset/vggish_postprocess.py#L84> youtube-8m dequantize - here <https://github.com/igor-panteleev/youtube-8m/blob/master/utils.py#L28> — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#7 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AE9rS9Tr3aFO-YOhxCw4JVPzL0-Y94I9ks5tkhkSgaJpZM4S8HeG> .

-- Yunbin