Issues with ImageClassificationTransformer

[yaeldekel]

• The Options class contains two string array fields: one for input column names and one for output column names. However, it always uses exactly two input columns: a label column and a features column, and it always outputs exactly two output columns: score and predicted label. This discrepancy should be addressed.
• The transformer has a field called _labelColumnName that is serialized and deserialized and otherwise does nothing. Since after deserialization we cannot rely on having data that has a label column anyway, this field should be removed.
• GetOutputSchema() in the estimator does not check that the type of the label column is correct (what other types, if any, should be allowed other than key type?), and neither does the constructor (it doesn't check the type of the features column either).
• I think the constructor of the transformer should not have a boolean flag to indicate whether the transformer is being instantiated from file or from the estimator. Instead, the estimator should do the required training (including figuring out the class count), and there should be one constructor for the transformer, that is called both from the estimator's Fit() method, and from the transformer's Create(ModelLoadContext) method.
• The estimator has a _transformer field that is not needed.
• The transformer ctor has an unused argument: batchSize - don't know if the error is in not using it or in not needing it.

[codemzs]

These are good to be addresses but not critical. I’ll consider this as a P3 and resolve them right before GA since this API undergoes changes every release to make it better.