st_page_03.py

#--------------------             
# Author : Serge Zaugg
# Description : some info 
#--------------------

import streamlit as st

c00, c01  = st.columns([0.6, 0.4])
with c00:
    with st.container(border=True) : 
        st.subheader("Short methods")
        st.markdown(''' 
            The training set consists of 900 mp3 recording from Northern Europe that were downloaded from xeno-canto.
            The mp3s were down-sampled to 24'000 sps, then cut in 1.0 seconds pieces and transformed to spectrograms resulting in 6714 grey-scale images of 128 x 1152 pixels (freq x time).
            These images were used to train convolutional auto-encoders (CAE) under a de-noising regime with PyTorch.
            The encoder weights were saved to be used later for feature extraction.              
                    
            The test set (shown here) comes from 571 mp3 recordings from South-Western Europe taken from xeno-canto.
            The recordings were down-sampled to 24'000 sps, cut into shorter 0.2 seconds pieces, and transformed to spectrograms.
            This gave 60'560 grey-scale mini-images of 128 x 256 pixels (freq x time).
            These images were piped through the pre-trained encoder to obtain a feature matrix (60'463 rows x 256 columns) 
            The 256 columns correspond to the dim of the latent vector in the CAE, not the 256 time-pixels of the images ;-)
            The feature matrix can be interactively dim-reduced with UMAP to the final matrix, e.g. to 60'463 x 32. 
            The final matrix can be interactively processed with DBSCAN and the mini-images can be plotted grouped by clusters.                         
            ''')
        
        st.subheader("Short discussion")
        st.markdown('''                     
            * Audio files were sequentially cut into pieces without any selection.
            * This is challenging because many pieces contain only nuisance sounds or background noise.
            * In some cases, consistent patterns coming from different mp3 recordings are clustered together.             
            * To be continued ...                             
            ''')
        
        st.subheader("Links")
        st.page_link("https://pytorch.org", label="PyTorch")
        st.page_link("https://umap-learn.readthedocs.io", label="UMAP")
        st.page_link("https://scikit-learn.org/stable/modules/generated/sklearn.cluster.DBSCAN.html", label="DBSCAN")