Posit AI Weblog: lime v0.4: The Kitten Image Version

Introduction

I’m completely happy to report a brand new main launch of lime has landed on CRAN. lime is
an R port of the Python library of the identical identify by Marco Ribeiro that enables
the person to pry open black field machine studying fashions and clarify their
outcomes on a per-observation foundation. It really works by modelling the end result of the
black field within the native neighborhood across the statement to clarify and utilizing
this native mannequin to clarify why (not how) the black field did what it did. For
extra details about the speculation of lime I’ll direct you to the article
introducing the methodology.

New options

The meat of this launch facilities round two new options which are considerably
linked: Native help for keras fashions and help for explaining picture fashions.

keras and pictures

J.J. Allaire was form sufficient to namedrop lime throughout his keynote introduction
of the tensorflow and keras packages and I felt compelled to help them
natively. As keras is by far the preferred method to interface with tensorflow
it’s first in line for build-in help. The addition of keras implies that
lime now immediately helps fashions from the next packages:

In case you’re engaged on one thing too obscure or innovative to not be capable of use
these packages it’s nonetheless attainable to make your mannequin lime compliant by
offering predict_model() and model_type() strategies for it.

keras fashions are used identical to another mannequin, by passing it into the lime()
perform together with the coaching knowledge with a purpose to create an explainer object.
As a result of we’re quickly going to speak about picture fashions, we’ll be utilizing one of many
pre-trained ImageNet fashions that’s out there from keras itself:

Mannequin
______________________________________________________________________________________________
Layer (kind)                              Output Form                         Param #        
==============================================================================================
input_1 (InputLayer)                      (None, 224, 224, 3)                  0              
______________________________________________________________________________________________
block1_conv1 (Conv2D)                     (None, 224, 224, 64)                 1792           
______________________________________________________________________________________________
block1_conv2 (Conv2D)                     (None, 224, 224, 64)                 36928          
______________________________________________________________________________________________
block1_pool (MaxPooling2D)                (None, 112, 112, 64)                 0              
______________________________________________________________________________________________
block2_conv1 (Conv2D)                     (None, 112, 112, 128)                73856          
______________________________________________________________________________________________
block2_conv2 (Conv2D)                     (None, 112, 112, 128)                147584         
______________________________________________________________________________________________
block2_pool (MaxPooling2D)                (None, 56, 56, 128)                  0              
______________________________________________________________________________________________
block3_conv1 (Conv2D)                     (None, 56, 56, 256)                  295168         
______________________________________________________________________________________________
block3_conv2 (Conv2D)                     (None, 56, 56, 256)                  590080         
______________________________________________________________________________________________
block3_conv3 (Conv2D)                     (None, 56, 56, 256)                  590080         
______________________________________________________________________________________________
block3_pool (MaxPooling2D)                (None, 28, 28, 256)                  0              
______________________________________________________________________________________________
block4_conv1 (Conv2D)                     (None, 28, 28, 512)                  1180160        
______________________________________________________________________________________________
block4_conv2 (Conv2D)                     (None, 28, 28, 512)                  2359808        
______________________________________________________________________________________________
block4_conv3 (Conv2D)                     (None, 28, 28, 512)                  2359808        
______________________________________________________________________________________________
block4_pool (MaxPooling2D)                (None, 14, 14, 512)                  0              
______________________________________________________________________________________________
block5_conv1 (Conv2D)                     (None, 14, 14, 512)                  2359808        
______________________________________________________________________________________________
block5_conv2 (Conv2D)                     (None, 14, 14, 512)                  2359808        
______________________________________________________________________________________________
block5_conv3 (Conv2D)                     (None, 14, 14, 512)                  2359808        
______________________________________________________________________________________________
block5_pool (MaxPooling2D)                (None, 7, 7, 512)                    0              
______________________________________________________________________________________________
flatten (Flatten)                         (None, 25088)                        0              
______________________________________________________________________________________________
fc1 (Dense)                               (None, 4096)                         102764544      
______________________________________________________________________________________________
fc2 (Dense)                               (None, 4096)                         16781312       
______________________________________________________________________________________________
predictions (Dense)                       (None, 1000)                         4097000        
==============================================================================================
Complete params: 138,357,544
Trainable params: 138,357,544
Non-trainable params: 0
______________________________________________________________________________________________

The vgg16 mannequin is a picture classification mannequin that has been construct as a part of
the ImageNet competitors the place the objective is to categorise photos into 1000
classes with the best accuracy. As we will see it’s pretty difficult.

With a view to create an explainer we might want to cross within the coaching knowledge as
nicely. For picture knowledge the coaching knowledge is actually solely used to inform lime that we
are coping with a picture mannequin, so any picture will suffice. The format for the
coaching knowledge is solely the trail to the photographs, and since the web runs on
kitten photos we’ll use one in every of these:

img <- image_read('https://www.data-imaginist.com/property/photographs/kitten.jpg')
img_path <- file.path(tempdir(), 'kitten.jpg')
image_write(img, img_path)
plot(as.raster(img))

As with textual content fashions the explainer might want to know learn how to put together the enter
knowledge for the mannequin. For keras fashions this implies formatting the picture knowledge as
tensors. Fortunately keras comes with numerous instruments for reshaping picture knowledge:

image_prep <- perform(x) {
  arrays <- lapply(x, perform(path) {
    img <- image_load(path, target_size = c(224,224))
    x <- image_to_array(img)
    x <- array_reshape(x, c(1, dim(x)))
    x <- imagenet_preprocess_input(x)
  })
  do.name(abind::abind, c(arrays, checklist(alongside = 1)))
}
explainer <- lime(img_path, mannequin, image_prep)

We now have an explainer mannequin for understanding how the vgg16 neural community
makes its predictions. Earlier than we go alongside, lets see what the mannequin consider our
kitten:

res <- predict(mannequin, image_prep(img_path))
imagenet_decode_predictions(res)

[[1]]
  class_name class_description      rating
1  n02124075      Egyptian_cat 0.48913878
2  n02123045             tabby 0.15177219
3  n02123159         tiger_cat 0.10270492
4  n02127052              lynx 0.02638111
5  n03793489             mouse 0.00852214

So, it’s fairly positive about the entire cat factor. The rationale we have to use
imagenet_decode_predictions() is that the output of a keras mannequin is all the time
only a anonymous tensor:

[1]    1 1000

NULL

We’re used to classifiers understanding the category labels, however this isn’t the case
for keras. Motivated by this, lime now have a method to outline/overwrite the
class labels of a mannequin, utilizing the as_classifier() perform. Let’s redo our
explainer:

model_labels <- readRDS(system.file('extdata', 'imagenet_labels.rds', package deal = 'lime'))
explainer <- lime(img_path, as_classifier(mannequin, model_labels), image_prep)

There may be additionally an as_regressor() perform which tells lime, unquestionably,
that the mannequin is a regression mannequin. Most fashions could be introspected to see
which kind of mannequin they’re, however neural networks doesn’t actually care. lime
guesses the mannequin kind from the activation used within the final layer (linear
activation == regression), but when that heuristic fails then
as_regressor()/as_classifier() can be utilized.

We are actually able to poke into the mannequin and discover out what makes it assume our
picture is of an Egyptian cat. However… first I’ll have to speak about yet one more
idea: superpixels (I promise I’ll get to the reason half in a bit).

With a view to create significant permutations of our picture (keep in mind, that is the
central thought in lime), we now have to outline how to take action. The permutations wants
to be substantial sufficient to have an effect on the picture, however not a lot that
the mannequin fully fails to recognise the content material in each case – additional,
they need to result in an interpretable consequence. The idea of superpixels lends
itself nicely to those constraints. In brief, a superpixel is a patch of an space
with excessive homogeneity, and superpixel segmentation is a clustering of picture
pixels into various superpixels. By segmenting the picture to clarify into
superpixels we will flip space of contextual similarity on and off throughout the
permutations and discover out if that space is necessary. It’s nonetheless essential to
experiment a bit because the optimum variety of superpixels depend upon the content material of
the picture. Bear in mind, we’d like them to be massive sufficient to have an effect however not
so massive that the category chance turns into successfully binary. lime comes
with a perform to evaluate the superpixel segmentation earlier than starting the
rationalization and it is strongly recommended to play with it a bit — with time you’ll
possible get a really feel for the appropriate values:

# default
plot_superpixels(img_path)

# Altering some settings
plot_superpixels(img_path, n_superpixels = 200, weight = 40)

The default is about to a reasonably low variety of superpixels — if the topic of
curiosity is comparatively small it could be crucial to extend the variety of
superpixels in order that the total topic doesn’t find yourself in a single, or a number of
superpixels. The weight parameter will mean you can make the segments extra
compact by weighting spatial distance larger than color distance. For this
instance we’ll persist with the defaults.

Bear in mind that explaining picture
fashions is far heavier than tabular or textual content knowledge. In impact it’s going to create 1000
new photographs per rationalization (default permutation measurement for photographs) and run these
by means of the mannequin. As picture classification fashions are sometimes fairly heavy, this
will lead to computation time measured in minutes. The permutation is batched
(default to 10 permutations per batch), so that you shouldn’t be afraid of working
out of RAM or hard-drive house.

rationalization <- clarify(img_path, explainer, n_labels = 2, n_features = 20)

The output of a picture rationalization is a knowledge body of the identical format as that
from tabular and textual content knowledge. Every characteristic might be a superpixel and the pixel
vary of the superpixel might be used as its description. Normally the reason
will solely make sense within the context of the picture itself, so the brand new model of
lime additionally comes with a plot_image_explanation() perform to do exactly that.
Let’s see what our rationalization have to inform us:

plot_image_explanation(rationalization)

We will see that the mannequin, for each the key predicted courses, focuses on the
cat, which is sweet since they’re each completely different cat breeds. The plot perform
received a number of completely different capabilities that will help you tweak the visible, and it filters low
scoring superpixels away by default. Another view that places extra focus
on the related superpixels, however removes the context could be seen by utilizing
show = 'block':

plot_image_explanation(rationalization, show = 'block', threshold = 0.01)

Whereas not as widespread with picture explanations additionally it is attainable to have a look at the
areas of a picture that contradicts the category:

plot_image_explanation(rationalization, threshold = 0, show_negative = TRUE, fill_alpha = 0.6)

As every rationalization takes longer time to create and must be tweaked on a
per-image foundation, picture explanations will not be one thing that you just’ll create in
massive batches as you may do with tabular and textual content knowledge. Nonetheless, a number of
explanations may mean you can perceive your mannequin higher and be used for
speaking the workings of your mannequin. Additional, because the time-limiting issue
in picture explanations are the picture classifier and never lime itself, it’s certain
to enhance as picture classifiers turns into extra performant.

Seize again

Other than keras and picture help, a slew of different options and enhancements
have been added. Right here’s a fast overview:

• All rationalization plots now embody the match of the ridge regression used to make
  the reason. This makes it simple to evaluate how good the assumptions about
  native linearity are saved.
• When explaining tabular knowledge the default distance measure is now 'gower'
  from the gower package deal. gower makes it attainable to measure distances
  between heterogeneous knowledge with out changing all options to numeric and
  experimenting with completely different exponential kernels.
• When explaining tabular knowledge numerical options will not be sampled from
  a traditional distribution throughout permutations, however from a kernel density outlined
  by the coaching knowledge. This could be certain that the permutations are extra
  consultant of the anticipated enter.

Wrapping up

This launch represents an necessary milestone for lime in R. With the
addition of picture explanations the lime package deal is now on par or above its
Python relative, feature-wise. Additional growth will concentrate on bettering the
efficiency of the mannequin, e.g. by including parallelisation or bettering the native
mannequin definition, in addition to exploring various rationalization varieties equivalent to
anchor.

Comfortable Explaining!