Skip to content

O
opencv
Commit cfd845ac authored by Vadim Pisarevsky's avatar Vadim Pisarevsky
Browse files
Options

Merge pull request #10141 from LaurentBerger:MLP_ReLU

parents 2a8344f7 a44573c4
Hide whitespace changes
Inline Side-by-side
Showing with 220 additions and 74 deletions
modules/ml/include/opencv2/ml.hpp
View file @ cfd845ac
...@@ -1503,14 +1503,18 @@ public: ...@@ -1503,14 +1503,18 @@ public:
enum ActivationFunctions { enum ActivationFunctions {
/** Identity function: \f$f(x)=x\f$ */ /** Identity function: \f$f(x)=x\f$ */
IDENTITY = 0, IDENTITY = 0,
/** Symmetrical sigmoid: \f$f(x)=\beta*(1-e^{-\alpha x})/(1+e^{-\alpha x}\f$ /** Symmetrical sigmoid: \f$f(x)=\beta*(1-e^{-\alpha x})/(1+e^{-\alpha x})\f$
@note @note
If you are using the default sigmoid activation function with the default parameter values If you are using the default sigmoid activation function with the default parameter values
fparam1=0 and fparam2=0 then the function used is y = 1.7159\*tanh(2/3 \* x), so the output fparam1=0 and fparam2=0 then the function used is y = 1.7159\*tanh(2/3 \* x), so the output
will range from [-1.7159, 1.7159], instead of [0,1].*/ will range from [-1.7159, 1.7159], instead of [0,1].*/
SIGMOID_SYM = 1, SIGMOID_SYM = 1,
/** Gaussian function: \f$f(x)=\beta e^{-\alpha x*x}\f$ */ /** Gaussian function: \f$f(x)=\beta e^{-\alpha x*x}\f$ */
GAUSSIAN = 2 GAUSSIAN = 2,
/** ReLU function: \f$f(x)=max(0,x)\f$ */
RELU = 3,
/** Leaky ReLU function: for x>0 \f$f(x)=x \f$ and x<=0 \f$f(x)=\alpha x \f$*/
LEAKYRELU= 4
}; };
/** Train options */ /** Train options */
......
modules/ml/src/ann_mlp.cpp
View file @ cfd845ac
...@@ -135,7 +135,7 @@ public: ...@@ -135,7 +135,7 @@ public:
void setActivationFunction(int _activ_func, double _f_param1, double _f_param2 ) void setActivationFunction(int _activ_func, double _f_param1, double _f_param2 )
{ {
if( _activ_func < 0 || _activ_func > GAUSSIAN ) if( _activ_func < 0 || _activ_func > LEAKYRELU)
CV_Error( CV_StsOutOfRange, "Unknown activation function" ); CV_Error( CV_StsOutOfRange, "Unknown activation function" );
activ_func = _activ_func; activ_func = _activ_func;
...@@ -153,11 +153,23 @@ public: ...@@ -153,11 +153,23 @@ public:
case GAUSSIAN: case GAUSSIAN:
max_val = 1.; min_val = 0.05; max_val = 1.; min_val = 0.05;
max_val1 = 1.; min_val1 = 0.02; max_val1 = 1.; min_val1 = 0.02;
if( fabs(_f_param1) < FLT_EPSILON ) if (fabs(_f_param1) < FLT_EPSILON)
_f_param1 = 1.; _f_param1 = 1.;
if( fabs(_f_param2) < FLT_EPSILON ) if (fabs(_f_param2) < FLT_EPSILON)
_f_param2 = 1.; _f_param2 = 1.;
break; break;
case RELU:
if (fabs(_f_param1) < FLT_EPSILON)
_f_param1 = 1;
min_val = max_val = min_val1 = max_val1 = 0.;
_f_param2 = 0.;
break;
case LEAKYRELU:
if (fabs(_f_param1) < FLT_EPSILON)
_f_param1 = 0.01;
min_val = max_val = min_val1 = max_val1 = 0.;
_f_param2 = 0.;
break;
default: default:
min_val = max_val = min_val1 = max_val1 = 0.; min_val = max_val = min_val1 = max_val1 = 0.;
_f_param1 = 1.; _f_param1 = 1.;
...@@ -368,47 +380,61 @@ public: ...@@ -368,47 +380,61 @@ public:
} }
} }
void calc_activ_func( Mat& sums, const Mat& w ) const void calc_activ_func(Mat& sums, const Mat& w) const
{ {
const double* bias = w.ptr<double>(w.rows-1); const double* bias = w.ptr<double>(w.rows - 1);
int i, j, n = sums.rows, cols = sums.cols; int i, j, n = sums.rows, cols = sums.cols;
double scale = 0, scale2 = f_param2; double scale = 0, scale2 = f_param2;
switch( activ_func ) switch (activ_func)
{ {
case IDENTITY: case IDENTITY:
scale = 1.; scale = 1.;
break; break;
case SIGMOID_SYM: case SIGMOID_SYM:
scale = -f_param1; scale = -f_param1;
break; break;
case GAUSSIAN: case GAUSSIAN:
scale = -f_param1*f_param1; scale = -f_param1*f_param1;
break; break;
default: case RELU:
; scale = 1;
break;
case LEAKYRELU:
scale = 1;
break;
default:
;
} }
CV_Assert( sums.isContinuous() ); CV_Assert(sums.isContinuous());
if( activ_func != GAUSSIAN ) if (activ_func != GAUSSIAN)
{ {
for( i = 0; i < n; i++ ) for (i = 0; i < n; i++)
{ {
double* data = sums.ptr<double>(i); double* data = sums.ptr<double>(i);
for( j = 0; j < cols; j++ ) for (j = 0; j < cols; j++)
{
data[j] = (data[j] + bias[j])*scale; data[j] = (data[j] + bias[j])*scale;
if (activ_func == RELU)
if (data[j] < 0)
data[j] = 0;
if (activ_func == LEAKYRELU)
if (data[j] < 0)
data[j] *= f_param1;
}
} }
if( activ_func == IDENTITY ) if (activ_func == IDENTITY || activ_func == RELU || activ_func == LEAKYRELU)
return; return;
} }
else else
{ {
for( i = 0; i < n; i++ ) for (i = 0; i < n; i++)
{ {
double* data = sums.ptr<double>(i); double* data = sums.ptr<double>(i);
for( j = 0; j < cols; j++ ) for (j = 0; j < cols; j++)
{ {
double t = data[j] + bias[j]; double t = data[j] + bias[j];
data[j] = t*t*scale; data[j] = t*t*scale;
...@@ -416,92 +442,132 @@ public: ...@@ -416,92 +442,132 @@ public:
} }
} }
exp( sums, sums ); exp(sums, sums);
if( sums.isContinuous() ) if (sums.isContinuous())
{ {
cols *= n; cols *= n;
n = 1; n = 1;
} }
switch( activ_func ) switch (activ_func)
{ {
case SIGMOID_SYM: case SIGMOID_SYM:
for( i = 0; i < n; i++ ) for (i = 0; i < n; i++)
{
double* data = sums.ptr<double>(i);
for (j = 0; j < cols; j++)
{ {
double* data = sums.ptr<double>(i); if (!cvIsInf(data[j]))
for( j = 0; j < cols; j++ )
{ {
if(!cvIsInf(data[j])) double t = scale2*(1. - data[j]) / (1. + data[j]);
{ data[j] = t;
double t = scale2*(1. - data[j])/(1. + data[j]); }
data[j] = t; else
} {
else data[j] = -scale2;
{
data[j] = -scale2;
}
} }
} }
break; }
break;
case GAUSSIAN: case GAUSSIAN:
for( i = 0; i < n; i++ ) for (i = 0; i < n; i++)
{ {
double* data = sums.ptr<double>(i); double* data = sums.ptr<double>(i);
for( j = 0; j < cols; j++ ) for (j = 0; j < cols; j++)
data[j] = scale2*data[j]; data[j] = scale2*data[j];
} }
break; break;
default: default:
; ;
} }
} }
void calc_activ_func_deriv( Mat& _xf, Mat& _df, const Mat& w ) const void calc_activ_func_deriv(Mat& _xf, Mat& _df, const Mat& w) const
{ {
const double* bias = w.ptr<double>(w.rows-1); const double* bias = w.ptr<double>(w.rows - 1);
int i, j, n = _xf.rows, cols = _xf.cols; int i, j, n = _xf.rows, cols = _xf.cols;
if( activ_func == IDENTITY ) if (activ_func == IDENTITY)
{ {
for( i = 0; i < n; i++ ) for (i = 0; i < n; i++)
{ {
double* xf = _xf.ptr<double>(i); double* xf = _xf.ptr<double>(i);
double* df = _df.ptr<double>(i); double* df = _df.ptr<double>(i);
for( j = 0; j < cols; j++ ) for (j = 0; j < cols; j++)
{ {
xf[j] += bias[j]; xf[j] += bias[j];
df[j] = 1; df[j] = 1;
} }
} }
} }
else if( activ_func == GAUSSIAN ) else if (activ_func == RELU)
{
for (i = 0; i < n; i++)
{
double* xf = _xf.ptr<double>(i);
double* df = _df.ptr<double>(i);
for (j = 0; j < cols; j++)
{
xf[j] += bias[j];
if (xf[j] < 0)
{
xf[j] = 0;
df[j] = 0;
}
else
df[j] = 1;
}
}
}
else if (activ_func == LEAKYRELU)
{
for (i = 0; i < n; i++)
{
double* xf = _xf.ptr<double>(i);
double* df = _df.ptr<double>(i);
for (j = 0; j < cols; j++)
{
xf[j] += bias[j];
if (xf[j] < 0)
{
xf[j] = f_param1*xf[j];
df[j] = f_param1;
}
else
df[j] = 1;
}
}
}
else if (activ_func == GAUSSIAN)
{ {
double scale = -f_param1*f_param1; double scale = -f_param1*f_param1;
double scale2 = scale*f_param2; double scale2 = scale*f_param2;
for( i = 0; i < n; i++ ) for (i = 0; i < n; i++)
{ {
double* xf = _xf.ptr<double>(i); double* xf = _xf.ptr<double>(i);
double* df = _df.ptr<double>(i); double* df = _df.ptr<double>(i);
for( j = 0; j < cols; j++ ) for (j = 0; j < cols; j++)
{ {
double t = xf[j] + bias[j]; double t = xf[j] + bias[j];
df[j] = t*2*scale2; df[j] = t * 2 * scale2;
xf[j] = t*t*scale; xf[j] = t*t*scale;
} }
} }
exp( _xf, _xf ); exp(_xf, _xf);
for( i = 0; i < n; i++ ) for (i = 0; i < n; i++)
{ {
double* xf = _xf.ptr<double>(i); double* xf = _xf.ptr<double>(i);
double* df = _df.ptr<double>(i); double* df = _df.ptr<double>(i);
for( j = 0; j < cols; j++ ) for (j = 0; j < cols; j++)
df[j] *= xf[j]; df[j] *= xf[j];
} }
} }
...@@ -510,34 +576,34 @@ public: ...@@ -510,34 +576,34 @@ public:
double scale = f_param1; double scale = f_param1;
double scale2 = f_param2; double scale2 = f_param2;
for( i = 0; i < n; i++ ) for (i = 0; i < n; i++)
{ {
double* xf = _xf.ptr<double>(i); double* xf = _xf.ptr<double>(i);
double* df = _df.ptr<double>(i); double* df = _df.ptr<double>(i);
for( j = 0; j < cols; j++ ) for (j = 0; j < cols; j++)
{ {
xf[j] = (xf[j] + bias[j])*scale; xf[j] = (xf[j] + bias[j])*scale;
df[j] = -fabs(xf[j]); df[j] = -fabs(xf[j]);
} }
} }
exp( _df, _df ); exp(_df, _df);
// ((1+exp(-ax))^-1)'=a*((1+exp(-ax))^-2)*exp(-ax); // ((1+exp(-ax))^-1)'=a*((1+exp(-ax))^-2)*exp(-ax);
// ((1-exp(-ax))/(1+exp(-ax)))'=(a*exp(-ax)*(1+exp(-ax)) + a*exp(-ax)*(1-exp(-ax)))/(1+exp(-ax))^2= // ((1-exp(-ax))/(1+exp(-ax)))'=(a*exp(-ax)*(1+exp(-ax)) + a*exp(-ax)*(1-exp(-ax)))/(1+exp(-ax))^2=
// 2*a*exp(-ax)/(1+exp(-ax))^2 // 2*a*exp(-ax)/(1+exp(-ax))^2
scale *= 2*f_param2; scale *= 2 * f_param2;
for( i = 0; i < n; i++ ) for (i = 0; i < n; i++)
{ {
double* xf = _xf.ptr<double>(i); double* xf = _xf.ptr<double>(i);
double* df = _df.ptr<double>(i); double* df = _df.ptr<double>(i);
for( j = 0; j < cols; j++ ) for (j = 0; j < cols; j++)
{ {
int s0 = xf[j] > 0 ? 1 : -1; int s0 = xf[j] > 0 ? 1 : -1;
double t0 = 1./(1. + df[j]); double t0 = 1. / (1. + df[j]);
double t1 = scale*df[j]*t0*t0; double t1 = scale*df[j] * t0*t0;
t0 *= scale2*(1. - df[j])*s0; t0 *= scale2*(1. - df[j])*s0;
df[j] = t1; df[j] = t1;
xf[j] = t0; xf[j] = t0;
...@@ -1110,7 +1176,9 @@ public: ...@@ -1110,7 +1176,9 @@ public:
{ {
const char* activ_func_name = activ_func == IDENTITY ? "IDENTITY" : const char* activ_func_name = activ_func == IDENTITY ? "IDENTITY" :
activ_func == SIGMOID_SYM ? "SIGMOID_SYM" : activ_func == SIGMOID_SYM ? "SIGMOID_SYM" :
activ_func == GAUSSIAN ? "GAUSSIAN" : 0; activ_func == GAUSSIAN ? "GAUSSIAN" :
activ_func == RELU ? "RELU" :
activ_func == LEAKYRELU ? "LEAKYRELU" : 0;
if( activ_func_name ) if( activ_func_name )
fs << "activation_function" << activ_func_name; fs << "activation_function" << activ_func_name;
...@@ -1191,6 +1259,8 @@ public: ...@@ -1191,6 +1259,8 @@ public:
{ {
activ_func = activ_func_name == "SIGMOID_SYM" ? SIGMOID_SYM : activ_func = activ_func_name == "SIGMOID_SYM" ? SIGMOID_SYM :
activ_func_name == "IDENTITY" ? IDENTITY : activ_func_name == "IDENTITY" ? IDENTITY :
activ_func_name == "RELU" ? RELU :
activ_func_name == "LEAKYRELU" ? LEAKYRELU :
activ_func_name == "GAUSSIAN" ? GAUSSIAN : -1; activ_func_name == "GAUSSIAN" ? GAUSSIAN : -1;
CV_Assert( activ_func >= 0 ); CV_Assert( activ_func >= 0 );
} }
......
modules/ml/test/test_mltests2.cpp
View file @ cfd845ac
...@@ -85,6 +85,22 @@ int str_to_ann_train_method( String& str ) ...@@ -85,6 +85,22 @@ int str_to_ann_train_method( String& str )
return -1; return -1;
} }
int str_to_ann_activation_function(String& str)
{
if (!str.compare("IDENTITY"))
return ANN_MLP::IDENTITY;
if (!str.compare("SIGMOID_SYM"))
return ANN_MLP::SIGMOID_SYM;
if (!str.compare("GAUSSIAN"))
return ANN_MLP::GAUSSIAN;
if (!str.compare("RELU"))
return ANN_MLP::RELU;
if (!str.compare("LEAKYRELU"))
return ANN_MLP::LEAKYRELU;
CV_Error(CV_StsBadArg, "incorrect ann activation function string");
return -1;
}
void ann_check_data( Ptr<TrainData> _data ) void ann_check_data( Ptr<TrainData> _data )
{ {
CV_TRACE_FUNCTION(); CV_TRACE_FUNCTION();
...@@ -177,6 +193,62 @@ float ann_calc_error( Ptr<StatModel> ann, Ptr<TrainData> _data, map<int, int>& c ...@@ -177,6 +193,62 @@ float ann_calc_error( Ptr<StatModel> ann, Ptr<TrainData> _data, map<int, int>& c
return err; return err;
} }
TEST(ML_ANN, ActivationFunction)
{
String folder = string(cvtest::TS::ptr()->get_data_path());
String original_path = folder + "waveform.data";
String dataname = folder + "waveform";
Ptr<TrainData> tdata = TrainData::loadFromCSV(original_path, 0);
ASSERT_FALSE(tdata.empty()) << "Could not find test data file : " << original_path;
RNG& rng = theRNG();
rng.state = 1027401484159173092;
tdata->setTrainTestSplit(500);
vector<int> activationType;
activationType.push_back(ml::ANN_MLP::IDENTITY);
activationType.push_back(ml::ANN_MLP::SIGMOID_SYM);
activationType.push_back(ml::ANN_MLP::GAUSSIAN);
activationType.push_back(ml::ANN_MLP::RELU);
activationType.push_back(ml::ANN_MLP::LEAKYRELU);
vector<String> activationName;
activationName.push_back("_identity");
activationName.push_back("_sigmoid_sym");
activationName.push_back("_gaussian");
activationName.push_back("_relu");
activationName.push_back("_leakyrelu");
for (size_t i = 0; i < activationType.size(); i++)
{
Ptr<ml::ANN_MLP> x = ml::ANN_MLP::create();
Mat_<int> layerSizes(1, 4);
layerSizes(0, 0) = tdata->getNVars();
layerSizes(0, 1) = 100;
layerSizes(0, 2) = 100;
layerSizes(0, 3) = tdata->getResponses().cols;
x->setLayerSizes(layerSizes);
x->setActivationFunction(activationType[i]);
x->setTrainMethod(ml::ANN_MLP::RPROP, 0.01, 0.1);
x->setTermCriteria(TermCriteria(TermCriteria::COUNT, 300, 0.01));
x->train(tdata, ml::ANN_MLP::NO_OUTPUT_SCALE);
ASSERT_TRUE(x->isTrained()) << "Could not train networks with " << activationName[i];
#ifdef GENERATE_TESTDATA
x->save(dataname + activationName[i] + ".yml");
#else
Ptr<ml::ANN_MLP> y = Algorithm::load<ANN_MLP>(dataname + activationName[i] + ".yml");
ASSERT_TRUE(y != NULL) << "Could not load " << dataname + activationName[i] + ".yml";
Mat testSamples = tdata->getTestSamples();
Mat rx, ry, dst;
x->predict(testSamples, rx);
y->predict(testSamples, ry);
absdiff(rx, ry, dst);
double minVal, maxVal;
minMaxLoc(dst, &minVal, &maxVal);
ASSERT_TRUE(maxVal<FLT_EPSILON) << "Predict are not equal for " << dataname + activationName[i] + ".yml and " << activationName[i];
#endif
}
}
// 6. dtree // 6. dtree
// 7. boost // 7. boost
int str_to_boost_type( String& str ) int str_to_boost_type( String& str )
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment