////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/// getErrorDerivative
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
RealVector MultiLayerPerceptron::calcErrorAndGradient( const RealVector& input, const RealVector& target, double& error )
{
/// The current state on the nodes will be used for backpropagation.
error = calcError( input, target );
if ( m_deltaE.size() == 0 )
{
/// Calculate the gradient components with a single backward propagation if there are no hidden layers.
propagateBackward( m_deltaEOutput, m_deltaEInput, m_weights[ 0 ] );
for ( size_t iInput = 0; iInput < m_weights[ 0 ].size(); ++iInput )
{
for ( size_t iOutput = 0; iOutput < m_weights[ 0 ][ iInput ].size(); ++iOutput )
{
m_weightDerivatives[ 0 ][ iInput ][ iOutput ] = m_input[ iInput ] * m_deltaEOutput[ iOutput ];
}
}
}
else
{
/// Calculate the derivatives of the activation function on all the hidden nodes. This is independent of
/// backpropagation.
calcDerivativesActivationFunc();
/// Propagate output deltas to first hidden node.
propagateBackward( m_deltaEOutput, m_deltaE.back(), m_weights.back() );
for ( size_t i = 0; i < m_deltaE.back().size(); ++i )
{
m_deltaE.back()[ i ] *= m_dfdy.back()[ i ];
}
/// Calculate deltas for all hidden nodes.
/// delta_nk = f'(y_nk) * sum_i w_nki delta_(n+1)i
/// propageBackward calculates the sum.
for ( size_t iHiddenLayer = m_deltaE.size() - 1; iHiddenLayer > 0; --iHiddenLayer )
{
propagateBackward( m_deltaE[ iHiddenLayer ], m_deltaE[ iHiddenLayer - 1 ], m_weights[ iHiddenLayer ] );
for ( size_t iNeuron = 0; iNeuron < m_deltaE[ iHiddenLayer - 1 ].size(); ++iNeuron )
{
m_deltaE[ iHiddenLayer - 1 ][ iNeuron ] *= m_dfdy[ iHiddenLayer - 1 ][ iNeuron ];
}
}
/// Calculate the derivatives. dE/dw_nij = x_(n-1)j * delta_i
for ( size_t iLayer = 0; iLayer < m_weights.size(); ++iLayer )
{
const RealVector& sourceNeuronActivations = iLayer != 0 ? m_x[ iLayer - 1 ] : m_input;
const RealVector& deltaVec = iLayer < m_deltaE.size() ? m_deltaE[ iLayer ] : m_deltaEOutput;
for ( size_t iSourceNeuron = 0; iSourceNeuron < m_weights[ iLayer ].size(); ++iSourceNeuron )
{
for ( size_t iTargetNeuron = 0; iTargetNeuron < m_weights[ iLayer ][ iSourceNeuron ].size(); ++iTargetNeuron )
{
m_weightDerivatives[ iLayer ][ iSourceNeuron ][ iTargetNeuron ] = sourceNeuronActivations[ iSourceNeuron ] * deltaVec[ iTargetNeuron ];
}
}
}
}
/// @see composeGradient
return composeGradient();
}
bool run()
{
#if DFG_ENABLE(DEBUG_PROPAGATION_VERBOSE)
m_count = 0;
#endif
// 1) propagate predictions
do {
m_changed = false;
// Forward propagation is near-optimal for both topologically-sorted and
// DFS-sorted code.
propagateForward();
if (!m_changed)
break;
// Backward propagation reduces the likelihood that pathological code will
// cause slowness. Loops (especially nested ones) resemble backward flow.
// This pass captures two cases: (1) it detects if the forward fixpoint
// found a sound solution and (2) short-circuits backward flow.
m_changed = false;
propagateBackward();
} while (m_changed);
// 2) repropagate predictions while doing double voting.
do {
m_changed = false;
doRoundOfDoubleVoting();
propagateForward();
if (!m_changed)
break;
m_changed = false;
doRoundOfDoubleVoting();
propagateBackward();
} while (m_changed);
return true;
}
bool FirmwareMaster::handleMessage(Message &message) {
int remaining;
byte slavesCount;
uint16_t temp;
DateTime now;
bool s;
uint32_t pcId;
switch (message.m_type) {
case SCAN_MESSAGE_RESPONSE:
d.println(F("SCAN_MESSAGE_RESPONSE:"));
s = m_bank_data.registerId(message.m_fromId);
break;
case SCAN_MESSAGE_START:
d.println(F("SCAN_MESSAGE_START:"));
pcId = message.m_fromId;
message.clearData();
set(message, m_id, 0, SCAN_MESSAGE);
propagateForward(message);
message.m_targetId = pcId;
message.m_type = SCAN_MESSAGE_START_RESPONSE;
return true;
case MASTER_DATA_READ:
d.println(F("MASTER_DATA_READ:"));
Utils::toByte(temp = (uint16_t)m_temp_sensor.readDigital(), message.m_data);
d.print(F("t=")).println(temp);
Utils::toByte(temp = (uint16_t)m_current_sensor.read(), message.m_data+2);
d.print(F("i=")).println(temp);
Utils::toByte(temp = (uint16_t)m_voltage_sensor.read(), message.m_data+4);
d.print(F("v=")).println(temp);
set(message, m_id, message.m_fromId, MASTER_DATA_READ_RESPONSE);
return true;
case MASTER_DATA_GATHER:
d.println(F("MASTER_DATA_GATHER:"));
now = m_rtc_clock.now();
m_bank_data.setTime(now.year(), now.month(), now.day(), now.hour(), now.minute(), now.second());
remaining = m_bank_data.addData(
m_id,
m_temp_sensor.readDigital(),
m_current_sensor.read(),
m_voltage_sensor.read()
);
// Do not emit a response until all data is gathered first!
if (remaining > 0) { // this handles the case when no slave are connected to master
d.print(F("Emitting SLAVE_DATA_READ command to"));
// Emit one SLAVE_DATA_READ message to every connected slave
set(message, m_id, 0, SLAVE_DATA_READ);
message.clearData();
slavesCount = m_bank_data.registeredIdCount();
d.print((uint32_t)slavesCount).println(" slaves");
for (byte i = 0; i < slavesCount; i++) {
message.m_targetId = m_bank_data.registeredIdAt(i);
propagateForward(message);
}
} else {
set(message, m_id, message.m_fromId, MASTER_DATA_GATHER_RESPONSE);
message.clearData();
message.m_data[CUSTOM_MESSAGE_DATA_LENGTH-1] = 1;
return true;
}
break;
case SLAVE_DATA_READ_RESPONSE:
d.println(F("SLAVE_DATA_READ_RESPONSE:"));
// When the last requested SLAVE_DATA_READ response is processed, the file will be
// automatically flushed to a new csv row.
remaining = m_bank_data.addData(
message.m_fromId,
message.m_data
);
if (remaining == 0) {
set(message, m_id, 1234, MASTER_DATA_GATHER_COMPLETE);
message.clearData();
message.m_data[CUSTOM_MESSAGE_DATA_LENGTH-1] = 1;
propagateBackward(message);
return false;
}
break;
case MASTER_ID_WRITE:
d.println(F("MASTER_ID_WRITE"));
m_eeprom_writer.writeId(Utils::toInt32(message.m_data));
message.clearData();
message.m_data[CUSTOM_MESSAGE_DATA_LENGTH-1] = 1;
message.m_type = MASTER_ID_WRITE_RESPONSE;
message.m_targetId = message.m_fromId;
message.m_fromId = m_id;
return true;
case MASTER_GPS_TIME_GET:
case MASTER_GPS_DATE_GET:
case MASTER_GPS_LAT_GET:
case MASTER_GPS_LON_GET:
case MASTER_GPS_ALTITUDE_GET:
case MASTER_GPS_FIX_TYPE_GET:
case MASTER_GPS_SPEED_GET:
case MASTER_GPS_TRACK_GET:
case MASTER_GPS_QUALITY_GET:
//case MASTER_GPS_ENABLED_GET: // TODO: finish this command
return processGpsCommand(message);
}
return false;
}
