Re: [Crm114-developers] 2 bugs in NN

[Ger H. <ge...@ho...>]

On Mon, Feb 23, 2009 at 8:01 PM, Bill Yerazunis <ws...@me...> wrote:
> I don't suppoose you could also reveal the pre-fixed code
> and line numbers, could you, with +/- line prefixes?
>
>    - Bill

Sure, here's the hacked diff file with GerH (anyway, it's just the
same as grabbing a line from those snippets and / or Ctrl-F finding
them in the source: it's safe as those bits only occur at one spot in
the source file (no [semi-]duplicates in other sections).

@@ -637,17 +1022,18 @@
   //
   //     Again, we sacrifice neuron 0 of the hidden layer to be the
   //     bias weights.
-
-  nn->output_layer[0] = *arefWout (nn, 0, 0);
-  nn->output_layer[0] = *arefWout (nn, 1, 0);
+    nn->output_layer[0] = nn->Wout[0]; // *arefWout(nn, 0, 0);
+    nn->output_layer[1] = nn->Wout[(1 * nn->hidden_layer_size) + 0];
// *arefWout(nn, 1, 0);   [i_a] bug

It's the second [0] anyway.

plus:

@@ -1523,20 +2270,26 @@
 	
 	//    Calculate the SUM  "in place" (the error mag term)
 	for (neuron = 0; neuron < nn->first_layer_size; neuron++)
+                {
 	  for (channel = 0; channel < nn->hidden_layer_size; channel++)
+                    {
 	    nn->delta_first_layer[neuron] +=
-	      nn->delta_hidden_layer[neuron]
-	      * (*arefWhid(nn, channel, neuron));
+                            nn->delta_hidden_layer[channel]  // [i_a] bug
+                            *  nn->Whid[channel *
nn->first_layer_size + neuron];
+	// (*arefWhid(nn, channel, neuron));
+                    }
+                }




-- 
Met vriendelijke groeten / Best regards,

Ger Hobbelt

--------------------------------------------------
web:    http://www.hobbelt.com/
        http://www.hebbut.net/
mail:   ge...@ho...
mobile: +31-6-11 120 978
--------------------------------------------------