CMSIS-NN: Tuned the softmax test criterion based upon analysis done with

scipy.

Testing/PatternGeneration/Softmax.py

@@ -5,7 +5,7 @@ import random
 import numpy as np
 import scipy.special as sp
 
-NBTESTSAMPLES = 1000
+NBTESTSAMPLES = 100
 
 def softmax(v):
   m = sp.softmax(v)
@@ -23,7 +23,7 @@ def writeTest(config,nb,vecDim):
 
 
     for _ in range(0,NBTESTSAMPLES):
-      va = np.random.randn(vecDim)
+      va = np.abs(np.random.randn(vecDim))
       va = va / np.sum(va)
 
       r = sp.softmax(va)

Testing/Patterns/NN/SoftmaxQ7/Dims1_s16.txt

@@ -1,6 +1,6 @@
 H
 2
-// 1000
-0x03E8
+// 100
+0x0064
 // 15
 0x000F

Testing/Source/Softmax.cpp

@@ -59,9 +59,10 @@ int16_t differences(int16_t *pa,int16_t *pb, int length)
           vec_in += this->vecDim;
        }
 
-       printf("Nb diffs : %d\n",differences(ref.ptr(),output.ptr(),this->nbSamples));
-
-       ASSERT_EQ(output,ref);
+       int diff = differences(ref.ptr(),output.ptr(),this->nbSamples);
+       // 5% of errors are accepted
+       ASSERT_TRUE(100.0*diff/this->nbSamples <= 5);
+       
     } 
 
   
@@ -91,6 +92,6 @@ int16_t differences(int16_t *pa,int16_t *pb, int length)
 
     void Softmax::tearDown(Testing::testID_t id,Client::PatternMgr *mgr)
     {
-        //output.dump(mgr);
+        output.dump(mgr);
         //temp.dump(mgr);
     }