…ion.

…ng flag

Due to the difference in images resized by different libraries, DeepImageFeaturizer is no longer required to match the results from keras on raw (non-resized) images.
Instead of computing l-inf or l2 norm of the two feature vectors, we compare their cosine distance and require it to be sufficiently "low" (< 1e-2).
We also ran several transfer learning examples and ensured that the results were comparable.
These experiments were successfull and new sparkdl's features proved to be at least as good as native keras ones, however, they have not been added as automated tests.

Overall I think the combination of (1) you get exact match with no resize and (2) not too different with resize is good enough.

Cosine distance justification:
Cosine distance ensures that the resulting feature vector has similar direction. Intuitively, this is important property for the generated features and I think ensuring cosine distance is low enough gives better guarantees than computing l2 or l-inf norm and comparing with huge allowed diff.

I did comparisons to different images, various amounts of added noise and some obvious bugs I could think of such as skipping the preprocess or having the color channels flipped. Most distances came orders of magnitude higher, noise with sd = 0.01 got comparable distance. Here's the breakdown on the test images: (the distance metric is by definition from [0,1] interval):

cosine distance per image to the same image with added (normal, mean = 0) noise:
sd = 1.00: [0.69, 0.78, 0.77 0.75, 0.76]
sd = 0.10: [0.1, 0.2, 0.31, 0.12, 0.23]
sd = 0.01: [0.0078, 0.0094, 0.060, 0.0040, 0.0085]

cosine distance with no preprocessing: all ~ 0.9
cosine distance with faulty preprocess (mean of one channel is incorrect): all ~ 0.1
cosine distance with flipped channels: all ~ 0.3
cosine distance matrix for the test images:
[ 0.00 0.75 0.71 0.70 0.67]
[ 0.75 0.00 0.79 0.85 0.74]
[ 0.71 0.80 0.00 0.75 0.69]
[ 0.70 0.85 0.75 0.00 0.70]
[ 0.67 0.74 0.69 0.70 0.00]

…p_models.py script no longer modifies the Models.scala source direclty. Instead it generates a file in the current working directory and lets user copy it.

  Few minor fixes, scaleHint converter in DeepImageFeaturizer is now lazy (and params are eagerly trasnfered to jvm)
  Added licences for generated named model wrappers.

…_model.py to python/model_gen/

…pace around operators and arguments."

…t parameter.