krupal09 / machine-learning-meets-animal-communication Goto Github PK

This is WP - unsupervised feature learning and clustering.

It is evolved from Hunaid's Lucidmonkeys repo (https://github.com/104H/lucidmonkeys). We (Krupal and Rachael) rewrote the pipeline code created by Hunaid (IDC) and Krupal. The improvements lies in the following aspects:
(1). we modularized the pipeline to enhance readability and re-usability and included 'argparse' to enable training the networks with different parameters easily (instead of having to dive into the code to make changes).
(2). We integrated commonly used signal preprocessing techniques which have been tested on killer whales and other animal species and part of detection and denoising WPs.
(3). We also visualized inputs, reconstructions and activations of intermediate layers to gain a more intuitive understanding of the work of the networks. During the IDC presentation, the feedback from the supervisors is the lack of examples to show the performance of autoencoders and clustering part.\newline (4). We implemented a more detailed and standard logging (as opposed to 'sys.stdout.flush()') and utilized tensorboard for sanity check and keeping tracking of the training .

Work-in-progress: early stopping and checkpoints

Usage (you need to update the paths and batch_size, max_epochs, and augmentation to your specific experiments): python3 pipeline_reconstruct.py --data_dir /.../ae/train_data/ --noise_dir /.../ae/noise/ --model_dir /.../ae/model --decod_dir /.../ae/decod_img/ --cache_dir /.../ae/cache --batch_size 24 --augmentation False --max_train_epochs 100

As a starting point, it is important to understand how preprocessing in the current workflow, it is recommended to put the ORCA-Spot Preprocessing.pdf from Felix (could be found on slack) and the two .py files under 'data' side-by-side and read through the commented codes. The preprocessing is adapted from Orcaspot. It would also be helpful to read the 'Data preprocessing and training' section of the publication to gain an overall understanding of the workflow (https://www.nature.com/articles/s41598-019-47335-w).

Below are the old readme from Lucidmonkeys.

To Those Who Will Come After Me,

I started this study project with the vision of understanding how Deep Neural Networks work, and bridging the gap between human and artificial intelligence. The task taught me a lot, about my skills and the naivete of my vision. There is only so much I achieved, but in this repository, I have laid the groundwork of taking this vision foward.

The idea was to use shallow autoencoders to learn a representation of spectrograms of chimpanzee vocalization segments, and use these low dimensional representations to cluster the audio to identify components of chimpanzee vocalizations.

You will find a pipeline.py file. This file performs the entire process of reading the data, reducing its dimensions and clustering those dimensions. I have provided an explanation of the code and commented it thoroughly in order to facilitate users in tweaking the model and running their own experiments. I have taken extra pain to reduce redundant work.

We were able to produce some strong results. You are welcome to clone this repository and carry the baton forward. I will be available to answer you queries.

Regards

Hunaid Hameed

Whatever you can do or dream you can, begin it. Boldness has genius, power, and magic in it. -Goethe

Building a Chimpanzee classifier alone does not inform us on their vocal communication system, neither does it inform us on how the classifier work. However, if we keep our models small, we can apply techniques in interpretability to interpret their inner workings. This, in turn, can refine our thinking.

To learn more, read the whitebox-ai.pdf presentation file. It also has a list of suggested reading material.

At our disposal, is a dataset of (approximately) 2s recordings of chimpanzees which contain a:

• Pant Hoot
• Scream
• Buttressed Drumming
• Bark
• Grunt
• Food Grunt
• Pant Grunt

The .spec files are in the folder, /net/projects/scratch/winter/valid_until_31_July_2021/0-animal-communication/data_grid/Chimp_IvoryCoast/detector_train_ds_spec/ on the grid (the grid is explained in the infrastructure section), and which of the above mentioned calls they contain is mentioned in the file labelsfromfilename in this repository. Those .spec files were originally produced during training the detector, using a pipeline built for killer whale detection (i.e. orcaspot by Bergler et al. (2019)). The team is working on a stand-alone module to prepare the .spec of new audio files for autoencoders.

The contents of the file (i.e. call types) are also embedded in the filename. The labels of the call types were manually annotated by Dr. Ammie Kalan from EVA and the 2s snippets were automatically extracted and manually verified (at noises vs. call level) by Rachael from SP.

1. Read spectrograms from a dataset
2. Train Autoencoder on it
3. Pass the data through the Autoencoder another time
4. Save the activations on the bottleneck layer in a csv file
5. Cluster these activations
6. Visualize the clusters (an interpret the visualization)
7. Compare the cluster assignment with the label assignment done by the human beings
   1. If they are similar, the model works and has the same discrimination ability as a person. Now test it on unseen/new/novel data.
   2. If they are not similar, you need reevaluate you approach and try again (or quit)
   3. If there some are consistent with human annotations while others don't fit, it would also be a good opportunity to revisit the spectrograms & .wav of those seemingly outliers, together with reexamining the original call type definitions.

The experiments are run on the grid. The grid is a technical term for a collection of computers which do not share RAM within themselves. Instructions and FAQs on how to use the grid and work with and around it are in the grid.md file.

The most important file is the pipeline.py file. It is the pipeline through which your data goes and outputs a trained model, downsampled calls of chimpanzees and clusters this downsampled information. Pytorch is used.

The code is thoroughly commented, if you still don't get it, email me.

This file will:

1. Create a folder with the date and time as its name, e.g, 24Mar2021-1243. All output produced by the file is saved here. This adds the convenience of keeping a log of your experiments in separate folders.
2. Read a dataset
3. Define an Autoencoder
4. Define a Dataset Loader
5. Load the Dataset
6. Define Training Parameters
7. Check if training can be done on a GPU, if not, then CPU it is
8. Write down the structure of the model in a file called output
9. Define optimizer and loss function of the Autoencoder
10. Normalize a spectrogram and train the Autoencoder on it
11. Save the model
12. Regenerate spectrogram images from their downsampled features and save them in a folder called regen-samples
13. Pass the entire dataset through the trained Autoencoder and write down the bottleneck layer features in a csv file
14. Cluster these features and save the cluster assignments in the csv file

Keep the folders for the results you like, delete the rest. The code saves all interim results, therefore, if the code crashes, you have something.

1. Play some of the 2s recording and develop intuition of the data. Being oblivious to the data is a detriment.
2. Read and understand the code.
3. Run an experiment. This just means, run the pipeline.py file. View and interpret the results. This is your starting point.
4. Adjust the model and view and interpret the results.
5. Repeat Step 4 till you are satisfied, disheartened or exhausted.
6. Share or report your results

Now the ball is rolling..

We clustered the activations of an Autoencoder with 7 bottleneck neurons into 7 clusters. The number 7 was chosen because there are 7 different types of calls. Gaussian Mixture Model clustering was used. The mean values of the clusters are visualed below:

This was our last most promising result.