Skip to content

Latest commit

 

History

History
executable file
·
97 lines (65 loc) · 2.85 KB

EXAMPLE_APPS.md

File metadata and controls

executable file
·
97 lines (65 loc) · 2.85 KB
Raw

Example Applications and Training

apps folder has some example applications which explains entire process to work with your dataset.

Install Before Start Training

Run followings to download python modules for using training.

cd ext
./download.sh

A. FSDKaggle2018

This trains a model for Freesound Dataset Kaggle 2018 dataset.

  • mobilenetv2_fsd2018_41cls.h5 is a trained model for 500 epochs, almost competitive even in the Kaggle competition.
  • It has useful representation for using as pretrained model for other tasks.
  • 44.1 kHz, 1 second duration, 128 n_mels and 128 time hops.

Followings will train MobileNetV2 model:

cd apps/fsdkaggle2018
python train.py

Followings will train AlexNet based model:

cd apps/fsdkaggle2018/alexnet
python train.py

Running this will convert trained model to Tensorflow .pb file.

  • FSDKaggle2018-TF-Model-Coversion.ipynb

B. FSDKaggle2018small

This trains a model for handing smaller audio data. Model is not small but audio, FS is 16 kHz for example.　�This sample is less computationally expensive for audio processing.

  • mobilenetv2_small_fsd2018_41cls.h5 is the model created by this.
  • 16 kHz, 1 second duration, 64 n_mels and 64 time hops.

Followings will train model:

cd apps/fsdkaggle2018small
python train_this.py

Followings will train AlexNet based model:

cd apps/fsdkaggle2018/alexnet
python train.py

C. CNN Laser Machine Listener

This is a experimental application example to github/Laser Machine Listener. Classification problem of sounds in hardware laboratory.

Originally simple NN was applied successfully for this classification problem, then what if we apply CNN?

Simple answer is too much for the provided dataset as is. It might be effective if:

  • We need a single model that needs to work fine in many different FabLabs. Then the model needs to generalize well.
  • And we have enough data from variety of FabLabs and machines.

This example has three notebooks to make it work.

  1. Run followings. 
    cd apps/cnn-laser-machine-listener
./download.sh
  2. Preprocess data by CNN-LML-Preprocess-Data.ipynb.
  3. Train model by CNN-LML-Train.ipynb.
  4. Convert model to .pb by CNN-LML-TF-Model-Conversion.ipynb.
  5. Then you can predict in realtime as follows. 
    python ../../realtime_predictor.py

cnn-model-laser-machine-listener.pb is ready in this repository for your quick try.

C.1 Another attempt for CNN Laser Machine Listener

AlexNet based model is also applied to this problem. It not only shows the better result, but also runs much faster.

  • CNN-LML-Another-Attempt-AlexNetBased.ipynb is the notebook that shows how to train, and visualization of results.
  • cnn-alexbased-laser-machine-listener.pb is also ready for your quick try.