Skip to content

Latest commit

 

History

History
112 lines (58 loc) · 5.92 KB

Readme.md

File metadata and controls

112 lines (58 loc) · 5.92 KB
Raw

Welcome to vacation2020 program

This is a program for Dachuang and design of graduation.

The project is about BirdClef . As for now, I have finished the model and preprocessing.

Here's some problems that should be handling following:

  1. how to deal with the audio that get many species make sound as same time? (cocktail party problem)

  2. Cause it's vacation time. The model cannot be too huge (though I can train the model with cpu), I still hope to decrease the data dimension.(Dose Image can do that?)

  3. Transfer learning seems like a good idea for data unbalance. The method based on model is easy, but I want use some method like TCA, JCA that based on dataset's features. Not only for transfer learning, but also for decreasing the dataset's size.

  4. the data will change through my code

  5. The problems above come out from my mind right now, not last...

2020/2/27

Uploading the basic code including data preprocessing, data generator, model.

2020/3/2

Have trained three models which are deep model, overfitting model, residual net respectively.

But the trend like a straight line, I will focus on fixing it. and improve the GPU efficiency is also important.

2020/3/3

I have trained ten models and I get two points:

  • the neural network I have made can get the point of accuracy is about 54% (So the preprocessing is import )
  • I should train a model with a lot of epochs until overfit, otherwise I cannot see whether the model is get the best point.

2020/3/4

  • complete tf.data.Dataset.from_generator script, but cannot feed it into model.fit_generator. may should use tf.estimator later
  • uploading tensorflow to 2.0.1, the office web is asked to update the CUDA, and 2.1.0cannot work though I have no idea what mistake I have made. In the end I downloading the 2.0.1

pray for @lilith

everything gonna be okay.

2020/3/10

Focus on preprocessing. Use CEEMDAN and ICA get some component, but don't know how to use it.

And CEEMDAN is slow, it's necessary use multi-threading or multi-processing

2020/3/19

Find out that STFT from matlab and python is different. And specs from python makes more sense. So generate ICA component from matlab (because it is faster). Then generating spectrums in python

2020/3/20

  • 发现鸟鸣集中在高频区域，论文部分也可以指出这点，生成的频率进行截取（NFFT要大），图片已有但最好自己再生成。

  • mobilenet模型前半段未上升，可能是之后学习到了应该在高频处学习，然后准确率迅速拉高，如果可以，可以深入这个点进行分析。

  • 各类模型要做比较，拿未做任何处理的与之比较，并提出自己的模型。

  • attention模型可以可视化特征图，看看其是否学习到了重点。

2020/3/23

  • 做了图像的截取，准确率得到提高，并且收敛的速度也更加快速。正在跑直接在fourier变换的数据
  • checkpoint保存不了optimizer导致模型加载出错。使用pickle保存模型？

2020/3/23

  • 没有找到特别好加快收敛的方法，hasbird的方案并不是特别有效。 还没做模型的eval。寻找一个可以完整保存模型的方法
  • 根据新学习的CNN位置的问题，调高NFFT，将音频放置在图像的中央进行学习。

2020/3/26

  • 更正 鸟鸣集中在低频区域，提高NFFT只是有效的拉伸了鸟鸣，但不能只靠提高NFFT就可以使信号集中在中央。

2020/3/29

  • 学习使用pytorch
  • 鸟鸣信号到底在低频还是在高频？

2020/3/31

  • 为什么mobilenetv1的准确率高于mobilenetv2和v3？原因？
  • 鸟鸣信号到底在低频还是高频？

2020/4/2

  • 学点loss相关的知识

2020/7/30

This programe is completed. Here's my summary for this program Following opnions are some problem and correspodding solution:

  • why I cannot reload model in both torch and keras => Cause when I generate label using dict like label = {label_ID:range(ID_num)} there is a random process so the label is not fixed
  • Proposing a multi-level model to fuse the family,genus,species data. The model is supposed to solve the unbalanced data problem and improve the model performance.
  • The training data is too small to reflect the model performance. And I cannot solve the overfitting problem.

In the end, hope somebody can propose better idea based on my idea

multilevel-model