machine_learning_workbench 0.5.1 → 0.6.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA1:
-  metadata.gz: afdd4f1bf99c2abbe10f4c348531f4efabde3c73
-  data.tar.gz: 2334e9b7e5c4f276f94e75dd50164f8fa78cc699
+  metadata.gz: 0677a5bd2af361c57794f13c2c62ceb353bab987
+  data.tar.gz: 160f5d7f06abc3c4520a7753cd4fa0f155bbb09e
 SHA512:
-  metadata.gz: 4f51ca6077627401bb1c27dfaed60cb84e5d70bdebe2f7c6d366abae299040e32ec7a4a1ad1ee3dc8ccc9aaa536a4125e91c2e97eaf043115d21e5e41d6a8e70
-  data.tar.gz: 884493c20fde5e8ac707f6c00b442d924a7dab5d783cf011926c2425b92116eea2bd1f10116a026b9b55db2210d918aa533eb0a6bc259b7bd0704df217b0985b
+  metadata.gz: d7e2872adb965b394cfc23c0cdd16f905e336c17ef0f91f7f53a05abc32a5e176720e9adaa163430898b80f873213dacda8fbfe346fc49c0cee589f1b6830cee
+  data.tar.gz: 49671433cb861a17a73414c0cd1c5b4f3249122ea809dec4ed2c7c14ddd4587cdbb1f8c688d5bdb59ab0d97adf4ff81ac0b4c3911c3344c5c5677a26e96942cf

data/README.md

@@ -58,4 +58,4 @@ Please feel free to contribute to this list (see `Contributing` above).
 - **UL-ELR** stands for Unsupervised Learning plus Evolutionary Reinforcement Learning, from the paper _"Intrinsically Motivated Neuroevolution for Vision-Based Reinforcement Learning" (ICDL2011)_. Check [here](https://exascale.info/members/giuseppe-cuccu/) for citation reference and pdf.
 - **BD-NES** stands for Block Diagonal Natural Evolution Strategy, from the homonymous paper _"Block Diagonal Natural Evolution Strategies" (PPSN2012)_. Check [here](https://exascale.info/members/giuseppe-cuccu/) for citation reference and pdf.
 - **RNES** stands for Radial Natural Evolution Strategy, from the paper _"Novelty-Based Restarts for Evolution Strategies" (CEC2011)_. Check [here](https://exascale.info/members/giuseppe-cuccu/) for citation reference and pdf.
-- **Online VQ** stands for Online Vector Quantization, from the paper _"Intrinsically Motivated Neuroevolution for Vision-Based Reinforcement Learning" (ICDL2011)_. Check [here](https://exascale.info/members/giuseppe-cuccu/) for citation reference and pdf.
+- **DLR-VQ** stands for Decaying Learning Rate Vector Quantization, from the algorithm originally named _*Online VQ*_ in the paper _"Intrinsically Motivated Neuroevolution for Vision-Based Reinforcement Learning" (ICDL2011)_. Check [here](https://exascale.info/members/giuseppe-cuccu/) for citation reference and pdf.

data/lib/machine_learning_workbench/compressor.rb

@@ -1,2 +1,3 @@
 require_relative 'compressor/vector_quantization'
-require_relative 'compressor/online_vector_quantization'
+require_relative 'compressor/decaying_learning_rate_vq'
+require_relative 'compressor/copy_vq'

data/lib/machine_learning_workbench/compressor/copy_vq.rb

@@ -0,0 +1,47 @@
+module MachineLearningWorkbench::Compressor
+  # Train-less VQ, copying new images into centroids
+  # Optimized for online training.
+  class CopyVQ < VectorQuantization
+
+    attr_reader :equal_simil, :next_train
+
+    def initialize **opts
+      puts "Ignoring learning rate: `lrate: #{opts[:lrate]}`" if opts[:lrate]
+      puts "Ignoring similarity: `simil_type: #{opts[:simil_type]}`" if opts[:simil_type]
+      # TODO: try different epsilons to reduce the number of states
+      # for example, in qbert we care what is lit and what is not, not the colors
+      @equal_simil = opts.delete(:equal_simil) || 0.0
+      super **opts.merge({lrate: nil, simil_type: nil})
+      @ntrains << 0 # to count duplicates, images we skip the train on
+      @next_train = 0 # pointer to the next centroid to train
+    end
+
+    def ntrains; @ntrains[0...-1]; end
+    def ntrains_skip; @ntrains.last; end
+
+    # Overloading lrate check from original VQ
+    def check_lrate lrate; nil; end
+
+    # Train on one vector:
+    # - train only if the image is not already in dictionary
+    # - find the next untrained centroid
+    # - training is just overwriting it
+    # @return [Integer] index of trained centroid
+    def train_one vec, eps: equal_simil
+      mses = centrs.map do |centr|
+        ((centr-vec)**2).sum / centr.size
+      end
+      # BEWARE: I am currently not handling the case where we run out of centroids!
+      # => Will be addressed directly by dynamic dictionary size
+      # return -1 if mses.min < eps
+      return -1 if mses.min < eps || next_train == ncentrs
+      trg_idx = next_train
+      @next_train += 1
+      # require 'pry'; binding.pry if next_train == ncentrs
+      puts "Overwriting centr #{next_train}"
+      centrs[trg_idx] = vec
+      trg_idx
+    end
+
+  end
+end

data/lib/machine_learning_workbench/compressor/{online_vector_quantization.rb → decaying_learning_rate_vq.rb}

@@ -1,7 +1,7 @@
 module MachineLearningWorkbench::Compressor
-  # Online Vector Quantization: VQ with per-centroid decaying learning rates.
+  # VQ with per-centroid decaying learning rates.
   # Optimized for online training.
-  class OnlineVectorQuantization < VectorQuantization
+  class DecayingLearningRateVQ < VectorQuantization
 
     attr_reader :lrate_min, :lrate_min_den, :decay_rate
 

data/lib/machine_learning_workbench/compressor/vector_quantization.rb

@@ -2,17 +2,20 @@ module MachineLearningWorkbench::Compressor
 
   # Standard Vector Quantization
   class VectorQuantization
-    attr_reader :ncentrs, :centrs, :dims, :vrange, :init_centr_vrange, :lrate, :simil_type, :rng, :ntrains
-    Verification = MachineLearningWorkbench::Tools::Verification
+    attr_reader :ncentrs, :centrs, :dims, :vrange, :init_centr_vrange, :lrate,
+      :simil_type, :encoding_type, :rng, :ntrains, :utility, :ncodes
+    attr_writer :utility, :ncodes # allows access from outside
+
+    def initialize ncentrs:, dims:, vrange:, lrate:, simil_type: nil, encoding_type: nil, init_centr_vrange: nil, rseed: Random.new_seed
+
+      @rng = Random.new rseed # TODO: RNG CURRENTLY NOT USED!!
 
-    def initialize ncentrs:, dims:, vrange:, lrate:, simil_type: nil, init_centr_vrange: nil, rseed: Random.new_seed
-      # TODO: RNG CURRENTLY NOT USED!!
-      @rng = Random.new rseed
       @ncentrs = ncentrs
       @dims = Array(dims)
-      check_lrate lrate # hack: so that we can overload it in online_vq
+      check_lrate lrate # hack: so that we can overload it in dlr_vq
       @lrate = lrate
       @simil_type = simil_type || :dot
+      @encoding_type = encoding_type || :ensemble_norm
       @init_centr_vrange ||= vrange
       @vrange = case vrange
         when Array
@@ -24,10 +27,12 @@ module MachineLearningWorkbench::Compressor
       end
       init_centrs
       @ntrains = [0]*ncentrs # useful to understand what happens
+      @utility = NArray.zeros [ncentrs] # trace how 'useful' are centroids to encodings
+      @ncodes = 0
     end
 
     # Verify lrate to be present and withing unit bounds
-    # As a separate method only so it can be overloaded in `OnlineVectorQuantization`
+    # As a separate method only so it can be overloaded in `DecayingLearningRateVQ`
     def check_lrate lrate
       raise ArgumentError, "Pass a `lrate` between 0 and 1" unless lrate&.between?(0,1)
     end
@@ -57,30 +62,45 @@ module MachineLearningWorkbench::Compressor
     # Computes similarities between vector and all centroids
     def similarities vec, type: simil_type
       raise NotImplementedError if vec.shape.size > 1
-      centrs.map { |centr| SIMIL[type].call centr, vec }
+      simil_fn = SIMIL[type] || raise(ArgumentError, "Unrecognized simil #{type}")
+      NArray[*centrs.map { |centr| simil_fn.call centr, vec }]
       # require 'parallel'
-      # Parallel.map(centrs) { |c| c.dot(vec).first }
+      # NArray[*Parallel.map(centrs) { |c| c.dot(vec).first }]
     end
 
     # Encode a vector
-    # TODO: optimize for Numo
-    def encode vec, type: :most_similar
+    # tracks utility of centroids based on how much they contribute to encoding
+    # TODO: `encode = Encodings.const_get(type)` in initialize`
+    # NOTE: hashes of lambdas or modules cannot access ncodes and utility
+    def encode vec, type: encoding_type
       simils = similarities vec
       case type
       when :most_similar
-        simils.index simils.max
+        code = simils.max_index
+        @ncodes += 1
+        @utility[code] += 1
+        code
       when :ensemble
-        simils
+        code = simils
+        tot = simils.sum
+        tot = 1 if tot < 1e-5  # HACK: avoid division by zero
+        contrib = code / tot
+        @ncodes += 1
+        @utility += (contrib - utility) / ncodes # cumulative moving average
+        code
       when :ensemble_norm
-        tot = simils.reduce(:+)
-        tot = 1 if tot == 0  # HACK: avoid division by zero
-        simils.map { |s| s/tot }
-      else raise ArgumentError, "unrecognized encode type: #{type}"
+        tot = simils.sum
+        tot = 1 if tot < 1e-5  # HACK: avoid division by zero
+        code = simils / tot
+        @ncodes += 1
+        @utility += (code - utility) / ncodes # cumulative moving average
+        code
+      else raise ArgumentError, "Unrecognized encode #{type}"
       end
     end
 
     # Reconstruct vector from its code (encoding)
-    def reconstruction code, type: :most_similar
+    def reconstruction code, type: encoding_type
       case type
       when :most_similar
         centrs[code]
@@ -98,14 +118,13 @@ module MachineLearningWorkbench::Compressor
     #   followed by the corresponding similarity
     def most_similar_centr vec
       simils = similarities vec
-      max_simil = simils.max
-      max_idx = simils.index max_simil
-      [max_idx, max_simil]
+      max_idx = simils.max_index
+      [max_idx, simils[max_idx]]
     end
 
     # Per-pixel errors in reconstructing vector
     # @return [NArray] residuals
-    def reconstr_error vec, code: nil, type: :most_similar
+    def reconstr_error vec, code: nil, type: encoding_type
       code ||= encode vec, type: type
       (vec - reconstruction(code, type: type)).abs.sum
     end
@@ -129,7 +148,7 @@ module MachineLearningWorkbench::Compressor
       vec_lst.each_with_index do |vec, i|
         trained_idx = train_one vec
         print '.' if debug
-        ntrains[trained_idx] += 1
+        @ntrains[trained_idx] += 1
       end
     end
   end

data/lib/machine_learning_workbench/neural_network/base.rb

@@ -162,31 +162,23 @@ module MachineLearningWorkbench::NeuralNetwork
 
     ## Activation functions
 
-    # Traditional sigmoid with variable steepness
+    # Traditional sigmoid (logistic) with variable steepness
     def sigmoid k=0.5
       # k is steepness:  0<k<1 is flatter, 1<k is flatter
       # flatter makes activation less sensitive, better with large number of inputs
-      -> (x) { 1.0 / (NMath.exp(-k * x) + 1.0) }
-    end
-
-    # Traditional logistic
-    def logistic
-      -> (x) do
-        exp = NMath.exp(x)
-        # exp.infinite? ? exp : exp / (1.0 + exp)
-        exp / (1.0 + exp)
-      end
+      -> (vec) { 1.0 / (NMath.exp(-k * vec) + 1.0) }
     end
+    alias logistic sigmoid
 
     # LeCun hyperbolic activation
     # @see http://yann.lecun.com/exdb/publis/pdf/lecun-98b.pdf Section 4.4
     def lecun_hyperbolic
-      -> (x) { 1.7159 * NMath.tanh(2.0*x/3.0) + 1e-3*x }
+      -> (vec) { 1.7159 * NMath.tanh(2.0*vec/3.0) + 1e-3*vec }
     end
 
     # Rectified Linear Unit (ReLU)
     def relu
-      -> (x) { (x>0).all? && x || x.class.zeros(x.shape) }
+      -> (vec) { (vec>0).all? && vec || vec.class.zeros(vec.shape) }
     end
 
 

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: machine_learning_workbench
 version: !ruby/object:Gem::Version
-  version: 0.5.1
+  version: 0.6.0
 platform: ruby
 authors:
 - Giuseppe Cuccu
 autorequire: 
 bindir: bin
 cert_chain: []
-date: 2018-04-06 00:00:00.000000000 Z
+date: 2018-04-12 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: bundler
@@ -202,7 +202,8 @@ files:
 - examples/neuroevolution.rb
 - lib/machine_learning_workbench.rb
 - lib/machine_learning_workbench/compressor.rb
-- lib/machine_learning_workbench/compressor/online_vector_quantization.rb
+- lib/machine_learning_workbench/compressor/copy_vq.rb
+- lib/machine_learning_workbench/compressor/decaying_learning_rate_vq.rb
 - lib/machine_learning_workbench/compressor/vector_quantization.rb
 - lib/machine_learning_workbench/monkey.rb
 - lib/machine_learning_workbench/neural_network.rb