Create Image Classification/Clustering Sample w/ DnnImageFeaturizerTransform

[justinormont]

We do have a sample of image classification w/ TF.

Instead (or in addition), we should demonstrate the use of DnnImageFeaturizerTransform. Being easier to utilize, we should be encouraging users to make use of it.

I'd recommend using the Dog Breeds vs. Fruits dataset which we used in NimbusML for its image examples. We currently host this dataset in our CDN for NimbusML.

In Python, the dataset / image loader looks like:

# Load image summary data from github
url = "https://express-tlcresources.azureedge.net/datasets/DogBreedsVsFruits/DogFruitWiki.SHUF.117KB.735-rows.tsv"
df_train = pd.read_csv(url, sep = "\t", nrows = 100)
df_train['ImagePath_full'] = "https://express-tlcresources.azureedge.net/datasets/DogBreedsVsFruits/" + \
                         df_train['ImagePath']
... load images

Purpose of the dataset is for example code & includes ~775 images of dogs & fruit:

/cc @CESARDELATORRE, @JRAlexander

(copied from PR -- dotnet/machinelearning#2372 (review))

[prathyusha12345]

I am fine with this. @CESARDELATORRE Do you recommend to add new sample or update existing one?

[developervariety]

This would be awesome, any update?

[CESARDELATORRE]

I think we could show the current plus this approach in the same sample, choosing one or the other approach, here:

https://github.com/dotnet/machinelearning-samples/blob/b40499b69fddd751f2c582e748f8f1de64497d84/samples/csharp/getting-started/DeepLearning_TensorFlowEstimator/README.md

About that dataset, we can switch to it since that one is larger than the current image dataset.

We'll add it to the backlog. Thanks! 👍

[CESARDELATORRE]

I think this should be in our backlog for samples, but not with a very high priority because:

• 1. It only supports Windows. See that limitation in this comment here:

https://github.com/dotnet/machinelearning/blob/e3830910531f00013c27391914233a085a1394a4/src/Microsoft.ML.OnnxTransform/DnnImageFeaturizerTransform.cs#L47

• 2. It is using ONNX. The feedback we're getting from ML.NET customers is that they want to see more TensorFlow samples. We mostly have no new requests about ONNX coming from customers.

In any case, I think it is a good scenario to have in the backlog and happy to discuss about its priority.
Thoughts? :)

[prathyusha12345]

Added a task to our backlog https://github.com/dotnet/machinelearning-samples/blob/master/docs/Samples-Backlog.md

[justinormont]

Thanks @CESARDELATORRE, @prathyusha12345.

@jignparm: Is the ONNX transform still restricted to only Windows?

https://github.com/dotnet/machinelearning/blob/e3830910531f00013c27391914233a085a1394a4/src/Microsoft.ML.OnnxTransform/DnnImageFeaturizerTransform.cs#L43-L48

    /// <summary>
    /// The Dnn Image Featurizer is just a wrapper around two <see cref="OnnxScoringEstimator"/>s and three <see cref="ColumnCopyingEstimator"/>
    /// with present pretrained DNN models. The ColumnsCopying are there to allow arbitrary column input and output names, as by default
    /// the ONNXTransform requires the names of the columns to be identical to the names of the ONNX model nodes.
    /// Note that because of this, it only works on Windows machines as that is a constraint of the OnnxTransform.
    /// </summary>