ljj7975 / honk2 Goto Github PK

View Code? Open in Web Editor NEW

4.0 3.0 0.0 1.69 MB

PyTorch implementations of neural network models for keyword spotting

License: Mozilla Public License 2.0

Python 100.00%

honk2's Introduction

honk2

dataset

GoogleSpeechCommandsDataset

version1: https://ai.googleblog.com/2017/08/launching-speech-commands-dataset.html
version2: wget http://download.tensorflow.org/data/speech_commands_v0.02.tar.gz

HeySnipsDataset

https://github.com/snipsco/keyword-spotting-research-datasets

tensorboard

tensorboard --logdir=<log dir>

Performance Summary

ResNet

res8

train acc: 0.9447739768695322
dev acc: 0.9431462125379703
test acc: 0.9365293720459149

res15

train acc: 0.9701254028872457
dev acc: 0.9686191968236504
test acc: 0.9615124915597569

res26

train acc: 0.9627312369654126
dev acc: 0.9620655839651749
test acc: 0.9540850776502363

Res-narrow

res8-narrow

train acc: 0.9113783483654289
dev acc: 0.9102585567700734
test acc: 0.9140220571685798

res15-narrow

train acc: 0.9530619430676309
dev acc: 0.9516132890047597
test acc: 0.9450821516993023

res26-narrow

train acc: 0.9429863763170012
dev acc: 0.9415197684709034
test acc: 0.9419311276164753

CNN

cnn-trad-pool2

train acc: 0.8374908588608109
dev acc: 0.8369011456863354
test acc: 0.8505514292144947

cnn-trad-fpool3

train acc: 0.9124617426396902
dev acc: 0.9103810076394933
test acc: 0.8936170212765957

cnn-one-fpool3

train acc: 0.829446656374421
dev acc: 0.8299245720916739
test acc: 0.8279459901800328

cnn-one-fstride4

train acc: 0.7669077218926897
dev acc: 0.7701624601102408
test acc: 0.7700490998363339

cnn-one-fstride8

train acc: 0.5886351940629994
dev acc: 0.5926651194275215
test acc: 0.6121112929623568

cnn-tstride2

train acc: 0.8917959968581566
dev acc: 0.8924910550236921
test acc: 0.8056464811783961

cnn-tstride4

train acc: 0.8786056715690258
dev acc: 0.8808142346001354
test acc: 0.8966857610474632

cnn-tstride8

train acc: 0.8433953576555347
dev acc: 0.8475485929794023
test acc: 0.877454991816694

cnn-tpool2

train acc: 0.9132472034885296
dev acc: 0.9126776907455758
test acc: 0.9065057283142389

cnn-tpool3

train acc: 0.9036320793044609
dev acc: 0.9034184314863166
test acc: 0.9095744680851063

honk2's People

Contributors

Stargazers

Watchers

honk2's Issues

Supporting multiple metric of different type

The goal is to support multiple metrics with different goals (max/min)

Implement res-narrow model

reset metrics before evaluation

When one training epoch is done, metrics should be reset before perform evaluation, or it would cause incorrect accumulate result.
is it a bug or is there something I missed?

implement cache for audio data

store loaded audio in memory to boost up the loading process

Investigate the cause of the accuracy mismatch

The accuracy of the current implementation is slightly less than the original implementation

log the learning rate to the tensorboard

A flag for lr scheduler stepping

Add a flag for scheduler step toggling between per-epoch vs. per-optimizer-step
A detailed discussion can be found in #14

Supporting hey snips dataset

https://github.com/snipsco/keyword-spotting-research-datasets

workspace class for handling model serialization

e.g.

class Workspace(object):
    def __init__(self, model, optimizer, other_stuff):

workspace.save(...);
workspace.load(...);

CTC loss

https://pytorch.org/docs/stable/nn.html#ctcloss

combining all main runnable scripts

combine all the __main__ runnable scripts in single dir
ex. run.train, run.evaluate, run.preprocess

Implement CNN models

https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/43969.pdf

Explore pydantic for default config

https://pydantic-docs.helpmanual.io/usage/settings/

various data augmentation

Report micro/macro-level metrics

let C_i be the number of correct examples in the i-th class
let T_i be the total number of examples in the i-th class

macro : (C_1 / T_1, C_2/ T_2, ...)
micro : (C_1 + C_2 + ...) / (T_1 + T_2 + ...)

replacing getattr by find_cls

following https://termbin.com/6zyn enables find_cls based class instantiation

@register('resnet-8')
class ResNet8(object):
...
model_cls = find_cls('resnet-8')

which effectively decouples the internal class name from the moniker

Update README.md

README.md should include

how to train
how to test
list of supported model
list of supported datasets
how to adjust config