- This project is for classification of inter-floor noise (SNU-B36-50) in a building using VGG16
- VGG16 is fine-tuned on SNU-B36-50 without freezing any weights
- The model is evaluated using 5-fold cross-validation
- The following confusion matrix shows the evaluation results
- VGG16-SNU-B36-50 will be merged into indoor-noise repository in near future
- Python (version 3.5.2)
- Python modules : TensorFlow (version 1.2), Numpy, Scipy, Pandas, matplotlib, librosa, and Pickle
- Pretrained weights of VGG16
-
audio: This folder includes inter-floor noises (SNU-B36-50) for training and validation
-
dataset: This folder includes metadata. Also, when the audio clips are converted to log scaled Mel-spectrograms they are saved into this folder
-
result: Cross-validation accuracy and confusion matrix are saved into this folder
-
cfmtx.py: This includes a confusion matrix drawing function
-
feature.py: This includes
melspec2a function which converts the audio clips to log scaled Mel-spectrograms using LibROSA -
gen_data.py: This reads the metadata and the audio clips are converted to log scaled Mel-spectrograms using feature.py. The Mel-spectrograms are saved as .p and saved in "dataset"
-
load_data.py: This can load training data and validation data (Currently, this supports batch and mini-batch)
-
vgg16_adap.py: This builds network architecture of VGG16 with an adaptation layer. Also, this supports
save_weights()which saves weights as .npz after a training -
main.py: You need to set several parameters. The current settings in the code are for IWAENC 2018 submission
gpu_device: Select a gpu deviceis_transfer_learn: Transfer the pretrained weights or notgen_data: If this is TRUE, then the audio clips are converted to log scaled Mel-spectrograms and they are saved as .pfreeze_layer- True: Freeze the weights except fc3w, fc3b, fc4w, and fc4b
- False: Do not freeze all the weights
bn: If this is TRUE, turn on batch normalizationsaver: If it is TURE, the weights at the last epoch are saved as .npzfold: Use k-th subsample as the validation set
-
results: Training loss, validation loss, training accuracy, and validation accuracy are saved to here
-
Clone this project
git clone https://github.com/yodacatmeow/VGG16-SNU-B36-50 -
Download the pretrained weights to this project path
-
Start a process
CUDA_VISIBLE_DEVICE=0 python3 main.py
@inproceedings{choi2018floornoise,
title={Classification of noise between floors in a building using pre-trained deep convolutional neural networks},
author={Choi, Hwiyong and Lee, Seungjun and Yang, Haesang and Seong, Woojae},
booktitle={2018 16th International Workshop on Acoustic Signal Enhancement (IWAENC)},
pages={535--539},
year={2018},
organization={IEEE}
}