ruby-dnn 0.2.0 → 0.2.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: c62e4bbbd5aa1e89dbff3e112c161f14fef9b4165721fed81815b6103af184f5
-  data.tar.gz: d5b46dfde07d33f9309adb4af6256fa404218b9ad1d1cfc82a79b6752fc4549e
+  metadata.gz: ccb761989aa380096c4b16db46da21a4fb74e95bb8b15ed4564f4da6818bf04f
+  data.tar.gz: 2764cc949b09a42b2bbb7ccee7504e62055b33318cf5f3beb4977eb50c9b86ff
 SHA512:
-  metadata.gz: 49515dd6ed15f07b551b88d3b724c280f417a07b12c869974d5b4f34a2c2bc6ff4747ba6c853840588b365e3d77ee50cebded9d82f739e79142aebaf505313b5
-  data.tar.gz: dbd7084f28a5a1c9b5675eadb9c20e76c42897ba32a20a552ed183ada75dc83aa7953718e853a9a658dcdb794317943c56f4e46ebeedf8d6a663eb02f90832c0
+  metadata.gz: da160fcb2a10367a916dbb020b10ecf6b7cc8eba0e1aa30d68e0f788c06a5a5ff29122a758308666384921f49866a493230aae527e343bec1b3470f6de849fa2
+  data.tar.gz: 6016b580542781fae7e1b07e2cb3aac219070a6428d0d02505fe833a1a54564947dbb196afeea76122fb964f761c7b23aca2513162d5b14be64113cddeac4f2b

data/README.md

@@ -1,7 +1,7 @@
 # ruby-dnn
 
 ruby-dnn is a ruby deep learning library. This library supports full connected neural network and convolution neural network.
-Currently, you can get 99% accuracy with MNIST and 60% with CIFAR 10.
+Currently, you can get 99% accuracy with MNIST and 74% with CIFAR 10.
 
 ## Installation
 

data/lib/dnn/core/layers.rb

@@ -314,11 +314,16 @@ module DNN
       include Convert
 
       def initialize(pool_width, pool_height, strides: nil, padding: false)
+        super()
         @pool_width = pool_width
         @pool_height = pool_height
         @strides = strides ? strides : [@pool_width, @pool_height]
         @padding = padding
-      end 
+      end
+
+      def self.load_hash(hash)
+        MaxPool2D.new(hash[:pool_width], hash[:pool_height], strides: hash[:strides], padding: hash[:padding])
+      end
 
       def build(model)
         super
@@ -386,6 +391,7 @@ module DNN
       attr_reader :shape
       
       def initialize(shape)
+        super()
         @shape = shape
         @x_shape = nil
       end
@@ -421,16 +427,21 @@ module DNN
     
     class Dropout < Layer
       def initialize(dropout_ratio)
+        super()
         @dropout_ratio = dropout_ratio
         @mask = nil
       end
 
+      def self.load_hash(hash)
+        self.new(hash[:dropout_ratio])
+      end
+
       def self.load(hash)
         self.new(hash[:dropout_ratio])
       end
     
       def forward(x)
-        if @model.training
+        if @model.training?
           @mask = SFloat.ones(*x.shape).rand < @dropout_ratio
           x[@mask] = 0
         else
@@ -443,17 +454,48 @@ module DNN
         dout[@mask] = 0 if @model.training
         dout
       end
+
+      def to_hash
+        {name: self.class.name, dropout_ratio: @dropout_ratio}
+      end
     end
     
     
     class BatchNormalization < HasParamLayer
+      def initialize(momentum: 0.9, running_mean: nil, running_var: nil)
+        super()
+        @momentum = momentum
+        @running_mean = running_mean
+        @running_var = running_var
+      end
+
+      def self.load_hash(hash)
+        running_mean = SFloat.cast(hash[:running_mean])
+        running_var = SFloat.cast(hash[:running_var])
+        self.new(momentum: hash[:momentum], running_mean: running_mean, running_var: running_var)
+      end
+
+      def build(model)
+        super
+        @running_mean ||= SFloat.zeros(*shape)
+        @running_var ||= SFloat.zeros(*shape)
+      end
+
       def forward(x)
-        @mean = x.mean(0)
-        @xc = x - @mean
-        @var = (@xc**2).mean(0)
-        @std = NMath.sqrt(@var + 1e-7)
-        @xn = @xc / @std
-        @params[:gamma] * @xn + @params[:beta]
+        if @model.training?
+          mean = x.mean(0)
+          @xc = x - mean
+          var = (@xc**2).mean(0)
+          @std = NMath.sqrt(var + 1e-7)
+          xn = @xc / @std
+          @xn = xn
+          @running_mean = @momentum * @running_mean + (1 - @momentum) * mean
+          @running_var = @momentum * @running_var + (1 - @momentum) * var
+        else
+          xc = x - @running_mean
+          xn = xc / NMath.sqrt(@running_var + 1e-7)
+        end
+        @params[:gamma] * xn + @params[:beta]
       end
     
       def backward(dout)
@@ -466,7 +508,16 @@ module DNN
         dvar = 0.5 * dstd / @std
         dxc += (2.0 / batch_size) * @xc * dvar
         dmean = dxc.sum(0)
-        dxc - dmean / batch_size 
+        dxc - dmean / batch_size
+      end
+
+      def to_hash
+        {
+          name: self.class.name,
+          momentum: @momentum,
+          running_mean: @running_mean.to_a,
+          running_var: @running_var.to_a,
+        }
       end
     
       private

data/lib/dnn/core/model.rb

@@ -8,7 +8,6 @@ module DNN
     attr_accessor :layers
     attr_reader :optimizer
     attr_reader :batch_size
-    attr_reader :training
   
     def initialize
       @layers = []
@@ -86,6 +85,10 @@ module DNN
     def compiled?
       @compiled
     end
+
+    def training?
+      @training
+    end
   
     def train(x, y, epochs,
               batch_size: 1,
@@ -136,7 +139,13 @@ module DNN
     end
   
     def accurate(x, y, batch_size = nil, &batch_proc)
-      @batch_size = batch_size if batch_size
+      unless batch_size
+        if @batch_size
+          batch_size = @batch_size >= x.shape[0] ? @batch_size : x.shape[0]
+        else
+          batch_size = 1
+        end
+      end
       correct = 0
       (x.shape[0].to_f / @batch_size).ceil.times do |i|
         x_batch = SFloat.zeros(@batch_size, *x.shape[1..-1])

data/lib/dnn/core/version.rb

@@ -1,3 +1,3 @@
 module DNN
-  VERSION = "0.2.0"
+  VERSION = "0.2.1"
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: ruby-dnn
 version: !ruby/object:Gem::Version
-  version: 0.2.0
+  version: 0.2.1
 platform: ruby
 authors:
 - unagiootoro
 autorequire: 
 bindir: exe
 cert_chain: []
-date: 2018-07-11 00:00:00.000000000 Z
+date: 2018-07-12 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: numo-narray