What exactly does CNN see?

Let’s start with the basic. We’re gonna need a model, test images for which we want to generate explanations and image preprocessing function (if needed).

library(sauron)

input_imgs_paths <- list.files(system.file("extdata", "images", package = "sauron"), full.names = TRUE)
model <- application_xception()
preprocessing_function <- xception_preprocess_input

There’s a ton of different methods to explain CNNs, but for now with sauron you have access to 6 gradient based ones. You can check full list using sauron_available_methods function:

sauron_available_methods
# # A tibble: 6 x 2
#   method name                  
#   <chr>  <chr>                 
# 1 V      Vanilla gradient      
# 2 GI     Gradient x Input      
# 3 SG     SmoothGrad            
# 4 SGI    SmoothGrad x Input    
# 5 IG     Integrated Gradients  
# 6 GB     Guided Backpropagation

Package is still in development so I won’t talk about the theory of those methods today. I will leave it for another post (or more probably multiple posts 🙂 ).

To generate any set of explanations simply use generate_explanations function. Beside, images paths, model and optional preprocessing function you have to pass class indexes for which explanation should be made (NULL means select class with highest probability for this image), some method specific arguments and if you want to generate grayscale o RGB explanation maps.

explanations <- generate_explanations(
  model,
  input_imgs_paths,
  preprocessing_function,
  class_index = NULL,
  methods = sauron_available_methods$method,
  num_samples = 5, # SmoothGrad samples
  noise_sd = 0.1, # SmoothGrad noise standard divination
  steps = 10, # Integrated Gradients steps
  grayscale = FALSE)

Now we can plot our results:

plot_explanations(explanations, FALSE)
# $Input

# 
# $V

# 
# $GI

# 
# $SG

# 
# $SGI

# 
# $IG

# 
# $GB

# 
# $Input

# 
# $V

# 
# $GI

# 
# $SG

# 
# $SGI

# 
# $IG

# 
# $GB

What next?

First of all, as I said in the beginning, sauron is still in development and it should be available at the end of 2020. So if this topic is interesting for you be sure to visit my github from time to time.

Second of all I’m planning to expand sauron capabilities. The first step will be to add methods like: Grad-CAM, Guided Grad-CAM, Occlusion and Layer-wise Relevance Propagation (LRP).