Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
    Yad
    @0bserver07
    sweet, yeah well lets see who else can be around (even though I've a conf to attent to on Saturday)
    Pommier
    @jeanpat
    @xennygrimmato @0bserver07 At what time do you plan it?
    Yad
    @0bserver07
    Cool I'm in UTC-08:00 time zone, Perhaps my morning time would work, since it will your evening (for both, if i'm not mistaken)
    Vaibhav Tulsyan
    @xennygrimmato
    How does tomorrow look for you guys?
    Jean-Patrick Pommier
    @dip4fish_twitter
    Hi, sorry for tomorrow (open school day)
    Yad
    @0bserver07
    @xennygrimmato I'm doing some reading, i have time, but do you wanna do sometime when Jean would be around as well?
    Vaibhav Tulsyan
    @xennygrimmato
    Sure :)
    Pommier
    @jeanpat
    Have an issue with jakeret/tf_unet#32
    Pommier
    @jeanpat
    Sourav Singh
    @souravsingh
    I have now added a dataloader for the Enet model. I have to now test the model.
    Pommier
    @jeanpat
    @souravsingh great, so impatient to see the result! Do plan to push your code on the repo?
    On my side I tried tf_unet, but it is not ready to train for multilabel ground truth images. The developper said it should be possible to tweak this problem by writing a custom dataloader...
    @souravsingh You may consider the new dataset : https://github.com/jeanpat/DeepFISH/tree/png-dataset/dataset . The groundtruth labels are clean.
    Yad
    @0bserver07
    Looks great :)
    Jean-Patrick Pommier
    @dip4fish_twitter
    Yes, I am still looking for a ready to train implementation 😀
    Yad
    @0bserver07
    Oh you mean a gray-scale model to train the data on?
    Pommier
    @jeanpat
    Yes, I found some for training on rgb images. There's an implementation with torchnet which could possibly fit.
    Sourav Singh
    @souravsingh
    I believe the Enet works with RGB images( I tested the model with CamVid and LSUN datasets.)
    Yad
    @0bserver07
    Wait but you're looking for gray-scale right? since the dataset doesn't contain rgb?
    Pommier
    @jeanpat
    @0bserver07 absolutely, this means trying to modify some code ...
    Pommier
    @jeanpat
    @0bserver07 For example, how many files should be modified in the One-Hundred-Layers-Tiramisu to handle greyscaled images?
    Yad
    @0bserver07
    I can definitely help, I didn't know that was the issue.
    Just the the model-file, it's as simple as changing the channel sizes:
    Here is an example of VGG-net Grayscale:
    https://gist.github.com/aaronpolhamus/034cfe4dd8e2cc26ccb24ad6616bf7b4
    as you can see the input is:
    model.add(Convolution2D(64, 3, 3, border_mode='valid', input_shape=(1, 128, 128))) it's 1 in the channel rather than 3
    Pommier
    @jeanpat
    thanks so much !!
    Yad
    @0bserver07
    For sure! I will edit the SegNet to do this and send it your way yeah?
    Pommier
    @jeanpat
    You would do that? Thanks Yad!
    Yad
    @0bserver07
    Absolutely! Tonight, I will do this!
    Also, we should touch-base on the paper status :)
    I have a draft and include all the proper work contributed here.
    Sourav Singh
    @souravsingh
    PyTorch is not running on my machine
    Would anyone be willing to test the model?
    Yad
    @0bserver07
    Hey Sourav, I have played around pyTorch, what's your error exactly?
    Sourav Singh
    @souravsingh
    This-
    Scanning dependencies of target TH [ 5%] Building C object CMakeFiles/TH.dir/THGeneral.c.o [ 10%] Building C object CMakeFiles/TH.dir/THHalf.c.o [ 15%] Building C object CMakeFiles/TH.dir/THAllocator.c.o [ 21%] Building C object CMakeFiles/TH.dir/THStorage.c.o [ 26%] Building C object CMakeFiles/TH.dir/THTensor.c.o cc: internal compiler error: Killed (program cc1) Please submit a full bug report, with preprocessed source if appropriate. See <file:///usr/share/doc/gcc-5/README.Bugs> for instructions. CMakeFiles/TH.dir/build.make:146: recipe for target 'CMakeFiles/TH.dir/THTensor.c.o' failed make[2]: *** [CMakeFiles/TH.dir/THTensor.c.o] Error 4 CMakeFiles/Makefile2:60: recipe for target 'CMakeFiles/TH.dir/all' failed make[1]: *** [CMakeFiles/TH.dir/all] Error 2 Makefile:116: recipe for target 'all' failed make: *** [all] Error 2
    The source installation is not working
    For anyone curious, I have 16 GB RAM
    So installation shouldn't have been a problem
    Pommier
    @jeanpat
    @0bserver07 great!
    Madhur Tandon
    @madhur-tandon
    Hi Everyone! I would like to get started working on this, can anyone guide me where to start ? Specifically, I am looking for the GitHub repository link and some list of topics that are needed for this.
    Pommier
    @jeanpat
    @amrandahvur Hi Madhur! The project is to to use semantic segmentation to perform segmentation of image of chromosomes and particularly to resolve overlapping chromosomes. What have been done:
    Two generations of dataset simulating pairs of overlapping chromosomes have been generated:
    A first version with pairs of grey scaled /label images stored in a hdf5 archive. In this version, the labels are not as clean as they could be (there are some noisy pixels).
    https://github.com/jeanpat/DeepFISH/tree/master/dataset.
    A second version dataset contains much more examples of pairs of overlapping examples and the labels are clean. The dataset is provided at both low and full resolution as png images. The training and validation dataset don't contain the same chromosomes:
    http://dip4fish.blogspot.fr/2017/04/modelisation-of-overlapping-pairs-of.html
    https://github.com/chromosome-seg/DeepFISH
    https://github.com/jeanpat/DeepFISH/tree/png-dataset/dataset
    Several people investigated segmentation with U-net:
    https://github.com/rohitghosh/chromosome_segementation
    https://github.com/chromosome-seg/image_segmentation_chromosomes
    https://blog.insightdatascience.com/separating-overlapping-chromosomes-with-deep-learning-based-image-segmentation-22f97afd3283
    http://dip4fish.blogspot.fr/2017/05/resolution-of-pairs-of-overlapping.html
    Semantic segmentation was not performed on "real world" overlapping chromosomes which can be more complex than two chromosomes:
    http://dip4fish.blogspot.fr/2017/04/100-overlapping-chromosomes-isolated.html
    Pommier
    @jeanpat
    Hi, a post about how to set-up and train a unet implementation with a dataset.
    Vaibhav Tulsyan
    @xennygrimmato
    :D
    Pommier
    @jeanpat
    I also posted a model trained for 14 epochs (this is not enough) on the first version of the dataset .
    Yad
    @0bserver07
    Nice! @jeanpat let's set up a time this weekend to catch up, I have done a bunch of tests from way back and haven't pushed it anywhere. I applied ENet as well.
    Pommier
    @jeanpat
    @0bserver07 good idea for the catch-up! What about saturday?
    Vaibhav Tulsyan
    @xennygrimmato
    Did you guys happen to catch up?
    Pommier
    @jeanpat
    not yet, we didn't say which saturday
    Vaibhav Tulsyan
    @xennygrimmato
    :D
    Vaibhav Tulsyan
    @xennygrimmato
    @jeanpat How was the conference - NIPS 2017?
    Jean-Patrick Pommier
    @dip4fish_twitter
    I don't know, I am in France . Lily Hu have a poster session today