Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Repo info
Activity
    DrCoffey
    @DrCoffey
    It will have to be re-written.
    Soengzaak
    @Soengzaak
    Hello, developer, I am a newbie who has just used DeepSqueak, thank you very much for developing this software. I'm doing an isolated ultrasonic vocal analysis of maternal pups based on the MIA model. I'm trying to use Supervised Classification to classify the vocal spectrum. Can I use MouseClassifierNet, a neural network? I noticed that this neural network divides vocalization into 5 categories (Split, Inverted U, Short Rise...). What are the specific classification criteria? Looking forward to your answer :D
    DrCoffey
    @DrCoffey
    Hey @Soengzaak, the MouseClassifierNet is a classification CNN, so it does not use specific classification criteria. It is just trained on a set of data that was scored by a person. However, I wouldn't recommend using it for classification of pup/mother calls. It was a simple proof of concept with only 5 classifications, and different mice stains/ages/etc have different and larger syllable repertoires. I would recommend using one of the unsupervised classification options to determine what kind of calls your animals are making. Our new test version has a few options, including contour-based and Variation Auto Encoder based clustering
    Soengzaak
    @Soengzaak
    Sincerely thank you for your contribution to science! This is very helpful. I can’t wait to see the new version released :D!
    DrCoffey
    @DrCoffey
    The new version is available now at the link above.
    Sean Bradley
    @seanpaulbradley
    Exporting to RavenPro currently throws an error looking for a "Begin Path" column when opened in RavenPro 1.6. This probably is a header error - the format is always changing.
    DrCoffey
    @DrCoffey
    Ya, that has always been an issue with Raven, they can never stick with 1 header format... I'll check out what the newest header format is and fix it sometime.
    Sean Bradley
    @seanpaulbradley
    I gave it a look. It wants a "Begin Path" column with the file name of the associated media file and a "File Offset" column - I don't think you can start detection at any point other than the start of the file (although you can truncate it) so this should just be identical to Begin Time. Adding these to the current export format permitted import.
    DrCoffey
    @DrCoffey
    Awesome thanks! I'll add that to the export code when I get a chance...
    hdesalvo
    @hdesalvo
    Hi again! I am experiencing a merging boxes error again only in some networks. This doesn't occur all the time either. I checked and my frequencies aren't out of range. It says something about an unrecognized variable 'h'? If this is something to put in the actual Issues section please let me know! Thank you so much for this software, it has been invaluable for my research!
    DrCoffey
    @DrCoffey
    h is the variable for the progress bar window. Sometimes it not open, and therefore the command to close it throws an error. I'm not sure why this happens, but if you can give me more info to recreate the error I might be able to figure it out.
    sometimes I just put a try; catch; end; around problem lines of code.
    that lets the error occur, but the code will keep running.
    hdesalvo
    @hdesalvo
    Thanks!
    Soengzaak
    @Soengzaak
    Hello, I am currently using the V3 version. When I use the supervised learning model to perform cluster analysis on sounds, an error message pops up after loading the neural network file, "Network not be found. Is this file a trained CNN?" The model is generated after I train the supervised learning model. How to solve this problem? Thanks for the answer: D
    DrCoffey
    @DrCoffey
    @Soengzaak I believe this is fixed in the latest release. I completely re-wrote the image scaling/generation so you will need to train a new classification network. It should work the same as version 2
    but the networks aren't cross compatible
    Marrekko
    @Marrekko
    Hey Dr Coffey & Russel Marx,
    ups :D
    Congratulations for the new Version, also to Ciszek & Valderrabano! It looks very promising and I hope to have some time in the next weeks to test it!
    Marrekko
    @Marrekko
    Saying this, it would be great, if you again could publish a compiled Windows version to allow as much scientist as possible to work with it. In 2019, I opened an issue because the compiled version would produce errors when training a network (no matter which). Last year, I had acces to the Matlab compiler myself through a promo at our university. I found, that the only thing that hinders the compiled version to work, is the training progress plot. Changed the line within the "options" command in each training function from 'Plots','training-progress' to 'Plots','none' will let the neural network training work! Hope you can adjust this for a compiled version :) I myself do not have access to the compiler anymore :S....
    Best Wishes!
    PS. Also the Screener worked perfect with this little change in the code to be compiled!
    DrCoffey
    @DrCoffey
    @Marrekko I will try to get a compiled version put together after we merge screener with the main branch, but that may still be a little while. If you have a working compiled version of the current screener code, feel free to share it here, or if you @ me on twitter I will share it there.
    Marrekko
    @Marrekko

    Hey, sure I wanted to offer that since months, but it never got on top of the agenda...
    DS-Screener can be downloaded following this link:
    https://hessenbox.uni-marburg.de/getlink/fi3MhuSEH5dfCXb74e5LvDLC/DeepSqueak_Screener_installer.exe

    The link will expire May 12th. It was compiled with DS-Screener codes from July 6th '20

    Ralph Emilio Peterson
    @ralphpeterson
    Hi all, I'm trying DS-Screener and wondering which version of matlab this has been tested with? I'm getting errors using 2018b ('Error in Network, Why Broken?', 'Error in Network Skipping Audio Chunk', no calls found in audio)
    DrCoffey
    @DrCoffey
    Hey Ralph, the screener branch (https://github.com/DrCoffey/DeepSqueak/tree/Screener) is built with some features only released recently. You will have to upgrade to 2020a or 2020b
    Matlab has made a lot of changes to Deep Learning code that kept breaking DS, but I have been talking with them about it and things should be much more stable now. I'm doing all development and testing on 2020b
    Ralph Emilio Peterson
    @ralphpeterson
    Thank you! I'll try it out
    Sdagher000
    @Sdagher000
    Hello, I have pulled the latest commit of the master branch and started exploring version 3 however I noticed a few things that I am not sure of whether it is just because of my environment or they have been changed. I cannot see the arrows needed to adjust the contour and waveform as in the previous version, in addition to that, I was trying supervised classification of rat calls but I cannot find the '' Wright model ''. Have these been changed in the new version?
    DrCoffey
    @DrCoffey
    Hey @Sdagher000. I removed the tonality slider because many people though it could be adjusted on a per-call basis. It is actually a global threshold that applies to all calls. You can still change it in the tools menu if you need to optimize contour extraction. I will likely add it back if we add manual tonality adjustments for every call.
    The supervised classification method was completely redone to improve image normalization across a wide range of audio files, so the old networks no longer function. It honestly wasn't particularly good, and biased toward labeling calls complex or trill.
    I am hoping someone will hand label calls and train a new network for Wright classification because I don't have the time right now.
    We tend to use the unsupervised clustering more anyway.
    bergagi1
    @bergagi1
    Hello,
    just a general question aftern aving used DSV2 and being impressed with it. Will there be a release of DS as a standalone executable or in some form indipendent of MatLab. Thank you for the great software!
    DrCoffey
    @DrCoffey
    Hey @bergagi1, I will try to compile a stand alone version on DeepSqueak v3 sometime. We are still in active development right now, so I will probably wait until the next stable release.
    bergagi1
    @bergagi1
    Thank you very much! Best
    weeezik
    @weeezik
    @DrCoffey Hi,
    weeezik
    @weeezik
    I am working on recording mice vocalizations, but I keep getting so much room noise that the calls are inaudible. Besides room treatment and other acoustic adjustments, which I am currently working on, would you recommend a certain frequency parameter when recording? Also, where can I find the Denoising Networks? Thank you.
    weeezik
    @weeezik
    I am using the Pettersson M500-384 ultrasosound microphone.
    DrCoffey
    @DrCoffey
    Hey @weeezik I haven't included a denoising network with the new version of DeepSqueak because we don't really need one anymore.
    You can train your own, but it rely on you capturing calls and noise. Can you see any calls in the 35-100kHz range?
    If your room is so loud that calls are completely occluded something is very off in your room, most external noise is well bellow the 35-100kHz of calls.
    hdesalvo
    @hdesalvo
    I am using the new version of DeepSqueak and noticed that some calls are detected but are not visible - yet they have a sound. Is this a graphics issue, or are the calls super quiet? I am using the Rat Detector YOLO R1 network.
    hdesalvo
    @hdesalvo
    Nevermind - resolved by changing graphic color limits. Thanks for this great update!
    hdesalvo
    @hdesalvo
    I have another question/comment. @DrCoffey I decided to check my old USV counts (for previous version) against the new update, and found that the counts are either much lower or higher than the previous version had found. I had manually checked my USV counts of the previous DS version against Raven Pro, so I am not sure why this update is not giving me the same output. Do you have any reasoning why this might be occurring?
    DrCoffey
    @DrCoffey
    Hey @hdesalvo, the new version of DeepSqueak uses a new method of spectrogram scaling and a new detection architecture (YOLO V2). In our experience the new networks have been performing better across a wider array of audio files, but there may certainly be some audio files for which the old networks worked better. Fortunately the new version of DeepSqueaks also make reviewing the files much easier, and retraining the networks much simpler.