steren’s avatarsteren’s Twitter Archive—№ 3,799

              1. Back to our weekend project with AnneMenini: Deep-learning based enhancer for underwater pictures
            1. …in reply to @steren
              After an afternoon, we arrived at some encouraging results with a convolutional neural network: top picture is the input, middle is result output, bottom is reference. Our problem with this approach is that the result loses its sharpness.
              oh my god twitter doesn’t include alt text from images in their APIoh my god twitter doesn’t include alt text from images in their API
          1. …in reply to @steren
            The GeForce GTX 1080 GPU is working hard :)
            oh my god twitter doesn’t include alt text from images in their API
        1. …in reply to @steren
          We wanted to convert our TensorFlow model to a format that can be used in Tensorflow.js for a web demo, but failed because we are using some operations not supported by the JS version: "MirrorPad, DecodeRaw, RandomUniformInt, ParseSingleExample" :(
      1. …in reply to @steren
        New approach: let's try a fully connected neural network and as output just a few image manipulation parameters (contrast, brightness, saturation). Adding more layers seems to help.
        oh my god twitter doesn’t include alt text from images in their API
    1. …in reply to @steren
      More layers seem to do the trick (top is input, middle is output, bottom is reference).
      oh my god twitter doesn’t include alt text from images in their APIoh my god twitter doesn’t include alt text from images in their API
  1. …in reply to @steren
    🎉 And we are now successfully loading and executing our model on a webpage in the browser with TensorFlowJS! (and all done on a Chromebook 💻). Python -> SavedModel -> tensorflowjs_converter -> stored on GCS -> JS Now, lets understand why it's not so great on a real picture.
    oh my god twitter doesn’t include alt text from images in their API
    1. …in reply to @steren
      🐟Back to our "deepwater" project: The previous model was fitting the training dataset too much, AnneMenini generalized it a bit (looking for an arbitrary red saturation value). We are now quite satisfied with the results. (btw, code is at github.com/annemenini/deepwater)
      OpenGraph image for github.com/annemenini/deepwateroh my god twitter doesn’t include alt text from images in their API
      1. …in reply to @steren
        AnneMenini The web app is now on Firebase Hosting, re-deployed automatically when we push to the GitHub repo (with Google Cloud Build triggers). Model is in a public Cloud Storage bucket (for now). Webapp code: github.com/annemenini/deepwater/tree/master/web Will share link when it looks better (UI and results).
        OpenGraph image for github.com/annemenini/deepwater/tree/master/web
        1. …in reply to @steren
          AnneMenini Profiling the web app that freezes when displaying the output image of the neural net. 7s spent in getBufferSubData()...
          oh my god twitter doesn’t include alt text from images in their API