backprop 0.0.0.1 → 0.0.1.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 30d4ab63e0502df289e6e648ad5c04c5d0ffe4b29a0bd5fce2053809c4879ddd
-  data.tar.gz: 943142da82fb2a4fd4adad13f55a1fe1f1e1713e71128a5c02c48887cc7aa4cb
+  metadata.gz: 2c263e60c633a0cad1e917d5bc3754e9c642c7389de1b3aded81dc120bedf163
+  data.tar.gz: 3a1ebc7367aa0ba51ba30b8a88bfdcc71bb24726123c49091b43e9465a93c44a
 SHA512:
-  metadata.gz: fedcb937e83efec000f8cc944e53b8cf8e61eb5e2ea4eccb60d4431ae230e9a7607adf4172a61cc14e32bfa41344538439c3538d5f4f52b2936df6ddf70827c1
-  data.tar.gz: 8d75ed673305ba9213840974a314caef65029156a23f56040dd99035166b8bb357fc44adff9f79a640d343fd84788a98422d7099b49e651016fa18aa48385108
+  metadata.gz: 16fb1e7ae73410ac405934c103606f723677eaa481e9d993b98caaad14de99fc1cbc4cbc1f3b21c64fd630c9b3cba608b8005fd120f2410e5fd2a0e81cd1bdde
+  data.tar.gz: 02600d7f6ef729f60285b94d033c875aa1c067df52ffff7350ffbe253ebaa35f5bee08efecfba46d57d85eb3012b21c76f995212e4e478466ad30da3d6d03a9a

data/README.md

@@ -1,3 +1,5 @@
+[![Tests](https://github.com/rickhull/backprop/actions/workflows/test.yaml/badge.svg)](https://github.com/rickhull/backprop/actions/workflows/test.yaml)
+
 # Backward Propagation
 
 This is a reimplementation of Andrej Karpathy's
@@ -202,12 +204,12 @@ puts output
 
 Loop:
 
-1. Backward propagate the gradients
+1. Run the network forward to generate a new output.
+2. Determine the loss; it should be smaller over time
+3. Backward propagate the gradients
    (derivatives for each value with respect to the output value)
-2. Adjust all weights slightly, according to their gradients.
-3. Run the network forward to generate a new output.
-   The loss should be smaller.
-   The new output should be closer to the desired output.
+4. Adjust all weights slightly, according to their gradients.
+
 
 ## Further Reading
 

data/VERSION

@@ -1 +1 @@
-0.0.0.1
+0.0.1.1

data/demo/loss.rb

@@ -55,7 +55,7 @@ gets
   end
 
   # 4. adjust all weights and biases towards minimizing loss function
-  n.descend(gradient_step)
+  loss.descend_recursive(gradient_step)
 }
 
 p outputs.map(&:value)

data/lib/backprop.rb

@@ -46,8 +46,17 @@ module BackProp
     def +(other)
       other = Value.wrap(other)
       val = Value.new(@value + other.value, children: [self, other], op: :+)
+
+      # What we're about to do here is pretty twisted.  We're going to refer
+      # to this execution context in the definition of a lambda, but we'll
+      # evaluate it later.
+      # Backstep is a lambda attached to val, which will be the return value
+      # here. When val.backstep is called later, it will update the gradients
+      # on both self and other.
       val.backstep = -> {
-        # gradients accumulate to handle a value used multiple times
+        # gradients accumulate for handling a term used more than once
+        # chain rule says to multiply val's gradient and the op's derivative
+        # derivative of addition is 1.0; pass val's gradient to children
         self.gradient += val.gradient
         other.gradient += val.gradient
       }
@@ -58,6 +67,7 @@ module BackProp
       other = Value.wrap(other)
       val = Value.new(@value * other.value, children: [self, other], op: :*)
       val.backstep = -> {
+        # derivative of multiplication is the opposite term
         self.gradient += val.gradient * other.value
         other.gradient += val.gradient * self.value
       }
@@ -65,15 +75,19 @@ module BackProp
     end
 
     # Mostly we are squaring(2) or dividing(-1)
+    # We don't support expressions, so Value is not supported for other
+    # This will look like a unary op in the tree
     def **(other)
       raise("Value is not supported") if other.is_a? Value
       val = Value.new(@value ** other, children: [self], op: :**)
       val.backstep = -> {
+        # accumulate, chain rule, derivative; as before
         self.gradient += val.gradient * (other * self.value ** (other - 1))
       }
       val
     end
 
+    # e^x - unary operation
     def exp
       val = Value.new(Math.exp(@value), children: [self], op: :exp)
       val.backstep = -> {
@@ -84,6 +98,7 @@ module BackProp
 
     #
     # Secondary operations defined in terms of primary
+    # These return differentiable Values but with more steps
     #
 
     def -(other)
@@ -96,6 +111,7 @@ module BackProp
 
     #
     # Activation functions
+    # Unary operations
     #
 
     def tanh
@@ -125,22 +141,37 @@ module BackProp
     # Backward propagation
     #
 
+    # Generally, this is called on the final output, say of a loss function
+    # It will initialize the gradients and then update the gradients on
+    # all dependent Values via back propagation
     def backward
-      self.reset_gradient
-      @gradient = 1.0
-      self.backprop
+      self.reset_gradient # set gradient to zero on all descendants
+      @gradient = 1.0     # this node's gradient is 1.0
+      self.backprop       # call backstep on all descendants
     end
 
+    # recursive call; visits all descendants; sets gradient to zero
     def reset_gradient
       @gradient = 0.0
       @children.each(&:reset_gradient)
       self
     end
 
+    # recursive call; visits all descendants; updates gradients via backstep
     def backprop
       self.backstep.call
       @children.each(&:backprop)
       self
     end
+
+    def descend(step_size = 0.1)
+      @value += -1 * step_size * @gradient
+    end
+
+    def descend_recursive(step_size = 0.1)
+      self.descend(step_size)
+      @children.each { |c| c.descend_recursive(step_size) }
+      self
+    end
   end
 end

data/lib/perceptron.rb

@@ -25,11 +25,8 @@ module BackProp
       sum.send(@activation)
     end
 
-    def descend(step_size)
-      (@weights + [@bias]).each { |p|
-        p.value += (-1 * step_size * p.gradient)
-      }
-      self
+    def parameters
+      @weights + [@bias]
     end
 
     def to_s
@@ -56,9 +53,8 @@ module BackProp
       @neurons.map { |n| n.apply(x) }
     end
 
-    def descend(step_size)
-      @neurons.each { |n| n.descend(step_size) }
-      self
+    def parameters
+      @neurons.map { |n| n.parameters }.flatten
     end
 
     def to_s
@@ -87,9 +83,8 @@ module BackProp
       x
     end
 
-    def descend(step_size)
-      @layers.each { |l| l.descend(step_size) }
-      self
+    def parameters
+      @layers.map { |l| l.parameters }.flatten
     end
 
     def to_s

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: backprop
 version: !ruby/object:Gem::Version
-  version: 0.0.0.1
+  version: 0.0.1.1
 platform: ruby
 authors:
 - Rick Hull