Thanks Abrita for your answer. I do indeed refer to the slide you mention. But I had a completely different meaning with my question. I'll try to be more clear.

First of all, forget what everybody thinks of doing with NeuralNets.

Second, I am totally not interested in the result this Net gives.

I want to use it completely different : I want to use the Net for creating abstractions of the original image.

I want to look at the different intermediate layers and see what kind of 'abstractions' they have made from an image that I input. Then I want to select one or more of these abstractions (results in one of the layers) and artistically work with that abstraction from there.

But I need a high resolution. I cannot blow up an abstraction (image) of 224x224 pixels to 6000x6000 pixels without ugly artifacts. So I want to adatp the layers to accept and work with images of a as big as possible size. Adapting the layers, so that they do not reduce the size of the image and thus that the different abstractions in the different layers are of a big pixel size.

I do not need to train or retrain the net.

Ok, I realise that I am probably asking for too much here, as the number of connections will explode very quick. Already a doubling in size might be hard.

So I was thinking of the following : I run the Net on a small version of my selected image. Peep into Layer 'X' and select an abstraction that I like from that layer.

Now I extract the workings(e.g. the Convolution) and parameters (weights) from that layer + previous layers and apply this to the big version of my image. I should get more or less the same result in the big image, as in the small abstraction from the layer that I have selected.

So my question would be : How to extract the workings (layer algorithm) and weights from the leyers involved? It might be that I'll have to scale up certain parameters when appying them to the big image (like maybe the Convolution kernel and other stuff). That I can try out myself of course.

Maybe it helps to realise that I am a visual artist who uses programming, among other techniques, for his work.

Thanks for your help.

Thanks a lot :-) Also, the links that are posted during the webinar in the chat, are unavailable for those who cannot attend at the live webinar, but like me, have to watch later. Please provide us as well with the link to the VideoClass "building apps with Neuralnet Depositery" with Megan Sheer. Thanks.

@B. Cornas are you refering to the example using NetModel["Wolfram ImageIdentify Net V1"] on slide 22 (that classifies the image of a tiger and looks at the output from the intermediate layers of the neural network)?

If yes, then it is possible to perform similar Image Classification tasks with out-of-core images. The Neural Net Framework is able to work with large images that may need to stored out-of-core.

An example can be found here: https://www.wolfram.com/language/11/neural-networks/out-of-core-image-classification.html where you will notice that the training data is an association of File expressions (a symbolic representation of a location in the local file system) and labels instead of the actual image itself.

The trained network is also used to classify images directly from file paths:

Sorry for the delay. The recording from the May 6 webinar has now been published. Use your unique link from your webinar email notifications, or click here: Data Science Friday webinar (May 6)

Hello,

Thank you for organizing this series!!

I was wondering whether there is a NN model that is trained to scan text from a pdf and do character recognition to obtain a text document from it. More specifically, what I would like to do is scan old typewritten mathematical text and produce a latex document from it. But a first step is to be able to do character recognition from a pdf'ed text. Or is there a function in Mathematica that can do this? For the kind of mathematical text I am trying to recognize and put into LaTex, I am uploading a little piece of it (source: AMS)

Attachments:

example.PNG

Here is the blog post (sorry not community post) I was referring to in today's session: https://blog.wolfram.com/2021/01/07/deploy-a-neural-network-to-your-ios-device-using-the-wolfram-language/

Looking forward to today's session on the Wolfram Data repository and Wolfram Neural Net Repository with @Bob Sandheinrich

Join us at https://www.bigmarker.com/wolfram-u/data-science-friday-ready-to-use-datasets-and-neural-net-models

Hi Patrick, The "lenet-CIFAR-10.wlnet" file has been included in the zip file that you downloaded. Could you please confirm that after you unzipped the download, the notebook "ExploringNNFramework-BuildingToTraining.nb" and the file "lenet-CIFAR-10.wlnet" appear in the same location?

If yes, then

Import[FileNameJoin[{NotebookDirectory[],"lenet-CIFAR-10.wlnet"}]

should be able to load the simple network modeled after Lenet into your session and you can work further with it.

Yes the model is slightly different from the network shown on the slide 4. Here is the code to create it yourself:

lenet = 
   NetChain[{
      ConvolutionLayer[20, 5], 
      Ramp, 
      PoolingLayer[2, 2], 
      ConvolutionLayer[50, 5], 
      Ramp, 
      PoolingLayer[2, 2], 
      FlattenLayer[], 
      500, 
      Ramp, 
      10, 
       SoftmaxLayer[]
   }, 
  "Output" -> NetDecoder[{"Class", classes}], 
  "Input" -> NetEncoder[{"Image", {32, 32}}]]

You can also find it in the tutorial at https://reference.wolfram.com/language/tutorial/NeuralNetworksComputerVision.html#280210622 within the example demonstrating "CIFAR-10 Object Classification".

Hi! 1. I'm using 12.0.0.0 version. 2. Yes, as you can see in the picture attached 3. The formatting you are suggesting is working 4. Can I ask one more question. What is the difference between LearnDistribution and Predict? Can I get the same probability distribution contour plot (for 2D data) with Predict as for LearnDistribution?

Hi Damir,

Two questions:

1. What version of the Wolfram Language are you using?

2. Can you confirm, the following lines of code have been successfully run prior to this evaluation?

   wine = RandomSample[ResourceData["Sample Data: Wine Quality"]];
   {winetest, winetrain} = TakeDrop[wine, 100];
   predictor = Predict[winetrain -> "WineQuality", Method -> "LinearRegression"]
   

If you continue to face issues, we may have to just reformat the data into one of the other formats acceptable as input to Predict - like this:

wine = Rest[#] -> First[#] & /@ 
  Normal[Values@
    RandomSample[ResourceData["Sample Data: Wine Quality"]]]
{winetest, winetrain} = TakeDrop[wine, 100];
predictor = Predict[winetrain, Method -> "LinearRegression"]
wineMeasurements = PredictorMeasurements[predictor, winetest]

Hello @Tsai Ming-Chou,

No you do not need to check the data distribution beforehand and there is no requirement for the data to be from a normal distribution when you use a functions like Classify or Predict. The automated machine learning functions are capable of handling any type of data from any distribution. You do have the option of specifying a FeatureExtractor with Classify (or Predict), if you have some specific thoughts about preprocessing the input features. Else Classify (or Predict) will automatically choose a suitable feature extractor based on the data.

LinearModelFit is specific in that it can only fit a linear regression model and can only work with numeric data. Classify or Predict can work on any type of data (numbers, text, images etc.) and has more options available in terms of the type of algorithm you want to use (DecisionTree, NaiveBayes, etc.)

Just want to let everyone know that we've posted an improved and complete recording of @Jon McLoone's presentation from last week. To access recordings, go to the series page, scroll down to find the title you want to see, and click on "Start this section." The video is embedded on each webinar page. Abrita also plans to revisit this final section of Jon's presentation at the end of this week's webinar. Hope to see you there!

Let us know if you have not received the links to the download notebooks over email. Best, Abrita

Sorry for the confusion about where you can download presentation notebooks. We will include the link in all upcoming reminders, so it's easy to find. A reminder will be sent today, so please watch for that in your inbox. All the notebooks for the full series will be placed in a single location, so you can download the latest each week.

We want to be sure everyone who is interested is signed up for the series, so that you receive all the details you need. You will receive reminder emails, recording notifications and further resources. Sign up for Data Science Friday Webinars.

@Charles Glover and @John Snyder we will send out an event reminder email with the link to the download folder containing all the notebooks from the entire webinar series.

@Abrita Chakravarty, I missed the link to the notebook for this presentation. Would you please post it.

Thanks

We did run into some issues with the presenter's internet connection towards the end of the session. We are working on preparing a recording of the presentation (without the audio dropout and screenshare issues) to share with our attendees.

Looking forward to the "Machine Learning and Statistics: Better Together Talk" by @Jon McLoone on Friday April 1 at 1:00 PM CT.

If you have any questions from last Friday's talk "Getting your Data Ready for Automated Machine Learning", please post them here.

Hi. @AbritaChakravarty I registered for the seminar. I already saw the first video that has the name Getting Your Data Ready for Automated Machine Learning.

could you share the mathematica notebook?