Awesome

Voice2Series-Reprogramming

Voice2Series: Reprogramming / Prompting Acoustic Models for Time Series Classification

• Paper | Colab Demo | Video | Slides

• We provide an end-to-end approach (Repro. layer) to reprogram on time series data on raw waveform with a differential mel-spectrogram layer from kapre.

• No offiline acoustic feature extraction and all layers are differentiable.

• Pytorch version of reprogram layer could be found out in ICASSP 23 Music Reprogramming.

• updated: if you have used the ECG 200 dataset in this code, please git pull and refer to the issue for one reported label loading error. (has been fixed)

Environment

<img alt="Keras" src="https://img.shields.io/badge/Keras%20-%23D00000.svg?&style=for-the-badge&logo=Keras&logoColor=white"/> <img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow%20-%23FF6F00.svg?&style=for-the-badge&logo=TensorFlow&logoColor=white" />

Tensorflow 2.2 (CUDA=10.0) and Kapre 0.2.0.

• PyTorch noted: Echo to many interests from the community, we will also provide Pytorch V2S layers and frameworks, incoperating the new torch audio layers. Feel free to email the authors for further reprogramming collaboration.

• option 1 (from yml)

conda env create -f V2S.yml

• option 2 (from clean python 3.6)

pip install tensorflow-gpu==2.1.0
pip install kapre==0.2.0
pip install h5py==2.10.0
pip install pyts

Training

• Random Mapping

Please also check the paper for actual validation details. Many Thanks!

python v2s_main.py --dataset 0 --eps 20 --mod 2 --seg 18 --mapping 1

• Result

Epoch 14/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.4493 - accuracy: 0.9239 - val_loss: 0.4571 - val_accuracy: 0.9106
Epoch 15/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.4297 - accuracy: 0.9306 - val_loss: 0.4381 - val_accuracy: 0.9265
Epoch 16/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.4182 - accuracy: 0.9247 - val_loss: 0.4204 - val_accuracy: 0.9205
Epoch 17/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.3972 - accuracy: 0.9320 - val_loss: 0.4072 - val_accuracy: 0.9242
Epoch 18/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.3905 - accuracy: 0.9303 - val_loss: 0.4099 - val_accuracy: 0.9242
Epoch 19/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.3765 - accuracy: 0.9320 - val_loss: 0.3924 - val_accuracy: 0.9258
Epoch 20/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.3704 - accuracy: 0.9300 - val_loss: 0.3816 - val_accuracy: 0.9250
--- Train loss: 0.36046191089949786
- Train accuracy: 0.93113023
--- Test loss: 0.38329164963780027
- Test accuracy: 0.925
=== Best Val. Acc:  0.92651516  At Epoch of  14

• Many-to-one Label Mapping

python v2s_main.py --dataset 0 --eps 20 --mod 2 --seg 18 --mapping 18

• Results

3601/3601 [==============================] - 4s 1ms/sample - loss: 0.8762 - accuracy: 0.9231 - val_loss: 0.8479 - val_accuracy: 0.9182
Epoch 12/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.8360 - accuracy: 0.9236 - val_loss: 0.8191 - val_accuracy: 0.9152
Epoch 13/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.7920 - accuracy: 0.9242 - val_loss: 0.7693 - val_accuracy: 0.9273
Epoch 14/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.7586 - accuracy: 0.9228 - val_loss: 0.7358 - val_accuracy: 0.9235
Epoch 15/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.7265 - accuracy: 0.9270 - val_loss: 0.7076 - val_accuracy: 0.9205
Epoch 16/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.6980 - accuracy: 0.9247 - val_loss: 0.6707 - val_accuracy: 0.9295
Epoch 17/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.6650 - accuracy: 0.9281 - val_loss: 0.6473 - val_accuracy: 0.9250
Epoch 18/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.6444 - accuracy: 0.9286 - val_loss: 0.6270 - val_accuracy: 0.9303
Epoch 19/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.6194 - accuracy: 0.9286 - val_loss: 0.6020 - val_accuracy: 0.9318
Epoch 20/20
3601/3601 [==============================] - 4s 1ms/sample - loss: 0.5964 - accuracy: 0.9275 - val_loss: 0.5813 - val_accuracy: 0.9227
--- Train loss: 0.5795955053139845
- Train accuracy: 0.93113023
--- Test loss: 0.5856682072986256
- Test accuracy: 0.92651516
=== Best Val. Acc:  0.9318182  At Epoch of  18

Class Activation Mapping

python cam_v2s.py --dataset 5 --weight wNo5_map6-88-0.7662.h5 --mapping 6 --layer conv2d_1

Theoretical Discussion

• For sliced wasserstein distance mapping and theoretical analysis, we use the POT package (JMLR 2021).

• The population risk for the target task via reprogramming a K-way source neural network classifier is upper bounded by equation above.

FAQ

• 1. Tips for tuning the model?

I would recommend using different label mapping numbers for training. For instance, you could use --mapping 7 for ECG 5000 dataset. The dropout rate is also an important hyperparameter for tuning the testing loss. You could use a range between 0.2 to 0.5 with --dr 4 for 0.4 dropout rate.

• 2. Masking the target sequence is important?

V2S mask is provided as an option, but the training script is not using the masking for forwarding passing. From our experiments, using or not using the masking only has small variants on the performance. This is not in conflict with the proposed theoretical analysis on learning target domain adaption.

• 3. Can we use Voice2Series for other domains or collaberate with the team?

Yes, you are welcome. Please send an email to the author for potential collaberation.

Pre-trained models and training

• VGGish AudioSet

cd weight
pip install gdown
gdown https://drive.google.com/uc?id=1mhqXZ8CANgHyepum7N4yrjiyIg6qaMe6

Additional Questions

Please open an issue here for discussion. Thank you!