🐀 happy-szczurki

🚀 Installation

1. Create virtualenv (with python 3.8.1):

   pyenv install 3.8.1
   pyenv virtualenv 3.8.1 <name>
   pyenv activate <name>

2. Install proper version of pytorch (https://pytorch.org/): For CPU version:

   pip3 install torch==1.4.0+cpu torchvision==0.5.0+cpu -f https://download.pytorch.org/whl/torch_stable.html

   For GPU version:

   TBA

3. Install requirements:

   pip3 install -r requirements.txt

4. Install happy_szczurki package:

   pip3 install -e .

5. Optional hotfix in skorch library.

🎓 Knowledge

Types of Ultrasonic Vocalizations

Ultrasonic vocalizations (USVs) have been observed in a number of rodent species (Sales 1972). In adult laboratory rats, two main types of USVs have been described: 22-kHz and 50-kHz calls (see Brudzynski 2009 for review). The 22-kHz call type has been termed a distress or “alarm” vocalization (Litvin et al. 2007), as it can be elicited by the presentation of a predator, painful stimuli, startling noises, and intermale aggression (Blanchard et al. 1991; Calvino et al. 1996; Han et al. 2005; Kaltwasser 1991; Thomas et al. 1983). In contrast, calls of the 50-kHz category have been detected in naturalistic appetitive contexts, such as during play, mating behavior, exploratory activity, or in anticipation of food reward (Burgdorf et al. 2000; Knutson et al. 1998; Sales 1972). 50-kHz calls have also been elicited by several non-natural appetitive stimuli, particularly rewarding electrical brain stimulation and amphetamine (AMPH) administration (Ahrens et al. 2009; Burgdorf et al. 2000, 2001a, 2007; Simola et al. 2009; Thompson et al. 2006; Wintink and Brudzynski 2001). Of note, the 50-kHz class of calls encompasses a wide frequency range (30–90 kHz) (Kaltwasser 1990; Sales and Pye 1974), and these calls vary considerably in spectrographic structure

Identification of multiple call categories within the rich repertoire of adult rat 50-kHz ultrasonic vocalizations: effects of amphetamine and social context

• https://link.springer.com/article/10.1007%2Fs00213-010-1859-y
• https://www.nature.com/articles/s41598-019-44221-3

code name	meaning	description
SH	Short	USVs with duration of less than 12 ms
FM	???
RP	Upward/Downward Ramp ?	USVs displaying a monotonic increase/decreasing in frequency
FL	Flat	USVs bearing a near-constant frequency
ST	???
CMP	???
IU	Inverted U / chevron	USVs possessing a monotonic increase in frequency followed by a monotonic decrease in frequency, resembling the shape of an inverted U
TR	Trill	USVs displaying a rapid frequency oscillation, usually appearing as a sinusoid oscillation
RM	???

• https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4598429/

📈 Results

🎯 USVs detection

Detecting time intervals which contains USVs.

model type	train set	test set	f1-score	precision	recall
SVC	ch1-2018-11-20_10-29-02_0000012.wav.npz	ch1-2018-11-20_10-26-36_0000010.wav.npz	0.92	0.92	0.91
SVC	ch1-2018-11-20_10-29-02_0000012.wav.trimed.npz	ch1-2018-11-20_10-26-36_0000010.wav.trimed.npz	0.93	0.94	0.93
RandomForestClassifier
XGBClassifier
CNN
LSTM

• TODO: share pickle models
• TODO: what is used weighting function?
• TODO: masking training ?

🎯 USVs' boundaries detection

Detecting time and frequency intervals (bounding box) which contains USVs.

TBD

🎯 USVs classification

Assigning one of USV types to each detected box.

TBD

😢 Existing solutions

• https://github.com/DrCoffey/DeepSqueak

📚 Materials to read

General

1. https://medium.com/@etown/great-results-on-audio-classification-with-fastai-library-ccaf906c5f52
2. https://www.kaggle.com/maxwell110/beginner-s-guide-to-audio-data-2 / https://www.kaggle.com/c/freesound-audio-tagging-2019/notebooks

Models to use

Hidden Markov Models (HMM)

1. http://lasa.epfl.ch/teaching/lectures/ML_MSc_Advanced/Slides/Lec_XI_HMM_extensions.pdf
2. https://blog.goodaudience.com/music-genre-classification-using-hidden-markov-models-4a7f14eb0fd4

Support Vector Machine (SVM)

1. https://www.academia.edu/1549822/Audio_Classification_Using_Support_Vector_Machines_and_Independent_Component_Analysis
2. https://www.researchgate.net/publication/267782696_Non-speech_environmental_sound_classification_using_SVMs_with_a_new_set_of_features

Parallel Recognition

1. https://www.sciencedirect.com/science/article/pii/S0003682X16305254

LSTM

1. https://www.depends-on-the-definition.com/guide-sequence-tagging-neural-networks-python/