DoubleBuffer2D neo::createDoubleBuffer2D(sys::ComputeSystem &cs, cl_int2 size) {
DoubleBuffer2D db;
db[_front] = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), size.x, size.y);
db[_back] = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), size.x, size.y);
return db;
}
void Predictor::activate(sys::ComputeSystem &cs, const std::vector<cl::Image2D> &visibleStates, NonlinearityType nonlinearityType, bool bufferSwap) {
// Start by clearing summation buffer
{
cl_float4 zeroColor = { 0.0f, 0.0f, 0.0f, 0.0f };
cl::array<cl::size_type, 3> zeroOrigin = { 0, 0, 0 };
cl::array<cl::size_type, 3> hiddenRegion = { _hiddenSize.x, _hiddenSize.y, 1 };
cs.getQueue().enqueueFillImage(_hiddenSummationTemp[_back], zeroColor, zeroOrigin, hiddenRegion);
}
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
VisibleLayer &vl = _visibleLayers[vli];
VisibleLayerDesc &vld = _visibleLayerDescs[vli];
int argIndex = 0;
_activateKernel.setArg(argIndex++, visibleStates[vli]);
_activateKernel.setArg(argIndex++, _hiddenSummationTemp[_back]);
_activateKernel.setArg(argIndex++, _hiddenSummationTemp[_front]);
_activateKernel.setArg(argIndex++, vl._weights[_back]);
_activateKernel.setArg(argIndex++, vld._size);
_activateKernel.setArg(argIndex++, vl._hiddenToVisible);
_activateKernel.setArg(argIndex++, vld._radius);
cs.getQueue().enqueueNDRangeKernel(_activateKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
// Swap buffers
std::swap(_hiddenSummationTemp[_front], _hiddenSummationTemp[_back]);
}
if (nonlinearityType == _binary) {
int argIndex = 0;
_solveHiddenBinaryKernel.setArg(argIndex++, _hiddenSummationTemp[_back]);
_solveHiddenBinaryKernel.setArg(argIndex++, _hiddenStates[_front]);
cs.getQueue().enqueueNDRangeKernel(_solveHiddenBinaryKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
}
else if (nonlinearityType == _tanH) {
int argIndex = 0;
_solveHiddenTanHKernel.setArg(argIndex++, _hiddenSummationTemp[_back]);
_solveHiddenTanHKernel.setArg(argIndex++, _hiddenStates[_front]);
cs.getQueue().enqueueNDRangeKernel(_solveHiddenTanHKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
}
else
cs.getQueue().enqueueCopyImage(_hiddenSummationTemp[_back], _hiddenStates[_front], { 0, 0, 0 }, { 0, 0, 0 }, { static_cast<cl::size_type>(_hiddenSize.x), static_cast<cl::size_type>(_hiddenSize.y), 1 });
// Swap hidden state buffers
std::swap(_hiddenStates[_front], _hiddenStates[_back]);
}
void PredictorSwarm::activate(sys::ComputeSystem &cs, const cl::Image2D &targets, const std::vector<cl::Image2D> &visibleStates, const std::vector<cl::Image2D> &visibleStatesPrev, float activeRatio, int inhibitionRadius, float noise, std::mt19937 &rng) {
// Start by clearing summation buffer
{
cl_float4 zeroColor = { 0.0f, 0.0f, 0.0f, 0.0f };
cl::array<cl::size_type, 3> zeroOrigin = { 0, 0, 0 };
cl::array<cl::size_type, 3> hiddenRegion = { _hiddenSize.x, _hiddenSize.y, 1 };
//cs.getQueue().enqueueCopyImage(_hiddenBiases[_back], _hiddenSummationTemp[_back], zeroOrigin, zeroOrigin, hiddenRegion);
cs.getQueue().enqueueFillImage(_hiddenSummationTemp[_back], cl_float4{ 0.0f, 0.0f, 0.0f, 0.0f }, zeroOrigin, hiddenRegion);
}
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
VisibleLayer &vl = _visibleLayers[vli];
VisibleLayerDesc &vld = _visibleLayerDescs[vli];
int argIndex = 0;
_activateKernel.setArg(argIndex++, visibleStates[vli]);
_activateKernel.setArg(argIndex++, _hiddenSummationTemp[_back]);
_activateKernel.setArg(argIndex++, _hiddenSummationTemp[_front]);
_activateKernel.setArg(argIndex++, vl._weights[_back]);
_activateKernel.setArg(argIndex++, vld._size);
_activateKernel.setArg(argIndex++, vl._hiddenToVisible);
_activateKernel.setArg(argIndex++, vld._radius);
cs.getQueue().enqueueNDRangeKernel(_activateKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
// Swap buffers
std::swap(_hiddenSummationTemp[_front], _hiddenSummationTemp[_back]);
}
{
std::uniform_int_distribution<int> seedDist(0, 999);
cl_uint2 seed = { seedDist(rng), seedDist(rng) };
int argIndex = 0;
_solveHiddenKernel.setArg(argIndex++, _hiddenSummationTemp[_back]);
_solveHiddenKernel.setArg(argIndex++, _hiddenStates[_front]);
_solveHiddenKernel.setArg(argIndex++, _hiddenActivations[_front]);
_solveHiddenKernel.setArg(argIndex++, _hiddenSize);
_solveHiddenKernel.setArg(argIndex++, inhibitionRadius);
_solveHiddenKernel.setArg(argIndex++, activeRatio);
cs.getQueue().enqueueNDRangeKernel(_solveHiddenKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
}
// Swap hidden state buffers
std::swap(_hiddenStates[_front], _hiddenStates[_back]);
std::swap(_hiddenActivations[_front], _hiddenActivations[_back]);
}
void Predictor::writeToStream(sys::ComputeSystem &cs, std::ostream &os) const {
abort(); // Not yet working
os << _hiddenSize.x << " " << _hiddenSize.y << std::endl;
{
std::vector<cl_float> hiddenStates(_hiddenSize.x * _hiddenSize.y);
cs.getQueue().enqueueReadImage(_hiddenStates[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_hiddenSize.x), static_cast<cl::size_type>(_hiddenSize.y), 1 }, 0, 0, hiddenStates.data());
for (int si = 0; si < hiddenStates.size(); si++)
os << hiddenStates[si] << " ";
os << std::endl;
}
// Layer information
os << _visibleLayers.size() << std::endl;
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
const VisibleLayer &vl = _visibleLayers[vli];
const VisibleLayerDesc &vld = _visibleLayerDescs[vli];
// Desc
os << vld._size.x << " " << vld._size.y << " " << vld._radius << std::endl;
// Layer
int weightDiam = vld._radius * 2 + 1;
int numWeights = weightDiam * weightDiam;
cl_int3 weightsSize = cl_int3{ _hiddenSize.x, _hiddenSize.y, numWeights };
int totalNumWeights = weightsSize.x * weightsSize.y * weightsSize.z;
{
std::vector<cl_float> weights(totalNumWeights);
cs.getQueue().enqueueReadImage(vl._weights[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(weightsSize.x), static_cast<cl::size_type>(weightsSize.y), static_cast<cl::size_type>(weightsSize.z) }, 0, 0, weights.data());
for (int wi = 0; wi < weights.size(); wi++)
os << weights[wi] << " ";
}
os << std::endl;
os << vl._hiddenToVisible.x << " " << vl._hiddenToVisible.y << " " << vl._visibleToHidden.x << " " << vl._visibleToHidden.y << " " << vl._reverseRadii.x << " " << vl._reverseRadii.y << std::endl;
}
}
void Predictor::learn(sys::ComputeSystem &cs, float tdError, const cl::Image2D &targets, std::vector<cl::Image2D> &visibleStatesPrev, float weightAlpha, float weightLambda) {
// Learn weights
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
VisibleLayer &vl = _visibleLayers[vli];
VisibleLayerDesc &vld = _visibleLayerDescs[vli];
int argIndex = 0;
_learnWeightsTracesKernel.setArg(argIndex++, visibleStatesPrev[vli]);
_learnWeightsTracesKernel.setArg(argIndex++, targets);
_learnWeightsTracesKernel.setArg(argIndex++, _hiddenStates[_front]);
_learnWeightsTracesKernel.setArg(argIndex++, vl._weights[_back]);
_learnWeightsTracesKernel.setArg(argIndex++, vl._weights[_front]);
_learnWeightsTracesKernel.setArg(argIndex++, vld._size);
_learnWeightsTracesKernel.setArg(argIndex++, vl._hiddenToVisible);
_learnWeightsTracesKernel.setArg(argIndex++, vld._radius);
_learnWeightsTracesKernel.setArg(argIndex++, weightAlpha);
_learnWeightsTracesKernel.setArg(argIndex++, weightLambda);
_learnWeightsTracesKernel.setArg(argIndex++, tdError);
cs.getQueue().enqueueNDRangeKernel(_learnWeightsTracesKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
std::swap(vl._weights[_front], vl._weights[_back]);
}
}
void HEInet::learnPrediction(sys::ComputeSystem &cs, const cl::Image2D &inputImage, float alpha) {
cl_float2 eFeedForwardDimsToEDims = { static_cast<float>(_eiLayers.front().getConfig()._eWidth + 1) / static_cast<float>(_eiLayers.front().getConfig()._eFeedForwardWidth + 1), static_cast<float>(_eiLayers.front().getConfig()._eHeight + 1) / static_cast<float>(_eiLayers.front().getConfig()._eFeedForwardHeight + 1) };
cl_float2 eFeedForwardDimsToIDims = { static_cast<float>(_eiLayers.front().getConfig()._iWidth + 1) / static_cast<float>(_eiLayers.front().getConfig()._eFeedForwardWidth + 1), static_cast<float>(_eiLayers.front().getConfig()._iHeight + 1) / static_cast<float>(_eiLayers.front().getConfig()._eFeedForwardHeight + 1) };
cl_int2 eDims = { _eiLayers.front().getConfig()._eWidth, _eiLayers.front().getConfig()._eHeight };
cl_int2 iDims = { _eiLayers.front().getConfig()._iWidth, _eiLayers.front().getConfig()._iHeight };
int index = 0;
_kernels->_predictionLearnKernel.setArg(index++, _eSpikeSumsIterPrev);
_kernels->_predictionLearnKernel.setArg(index++, _iSpikeSumsIterPrev);
_kernels->_predictionLearnKernel.setArg(index++, inputImage);
_kernels->_predictionLearnKernel.setArg(index++, _predictionPrev);
_kernels->_predictionLearnKernel.setArg(index++, _predictionFromEWeights._weightsPrev);
_kernels->_predictionLearnKernel.setArg(index++, _predictionFromIWeights._weightsPrev);
_kernels->_predictionLearnKernel.setArg(index++, _predictionFromEWeights._weights);
_kernels->_predictionLearnKernel.setArg(index++, _predictionFromIWeights._weights);
_kernels->_predictionLearnKernel.setArg(index++, eFeedForwardDimsToEDims);
_kernels->_predictionLearnKernel.setArg(index++, eFeedForwardDimsToIDims);
_kernels->_predictionLearnKernel.setArg(index++, eDims);
_kernels->_predictionLearnKernel.setArg(index++, iDims);
_kernels->_predictionLearnKernel.setArg(index++, _predictionRadiusFromE);
_kernels->_predictionLearnKernel.setArg(index++, _predictionRadiusFromI);
_kernels->_predictionLearnKernel.setArg(index++, alpha);
cs.getQueue().enqueueNDRangeKernel(_kernels->_predictionLearnKernel, cl::NullRange, cl::NDRange(_eiLayers.front().getConfig()._eFeedForwardWidth, _eiLayers.front().getConfig()._eFeedForwardHeight));
}
void HEInet::update(sys::ComputeSystem &cs, const cl::Image2D &inputFrequencyImage, const cl::Image2D &zeroImage, float eta, float shDecay, float saDecay) {
// Update input spikes
int index = 0;
_kernels->_updateInputSpikesKernel.setArg(index++, inputFrequencyImage);
_kernels->_updateInputSpikesKernel.setArg(index++, _inputSpikeTimersPrev);
_kernels->_updateInputSpikesKernel.setArg(index++, _inputSpikesHistoryPrev);
_kernels->_updateInputSpikesKernel.setArg(index++, _inputSpikeTimers);
_kernels->_updateInputSpikesKernel.setArg(index++, _inputSpikes);
_kernels->_updateInputSpikesKernel.setArg(index++, _inputSpikesHistory);
_kernels->_updateInputSpikesKernel.setArg(index++, shDecay);
cs.getQueue().enqueueNDRangeKernel(_kernels->_updateInputSpikesKernel, cl::NullRange, cl::NDRange(_eiLayers.front().getConfig()._eFeedForwardWidth, _eiLayers.front().getConfig()._eFeedForwardHeight));
const cl::Image2D* pLayerInput = &_inputSpikesPrev;
// Feed forward
for (int li = 0; li < _eiLayers.size(); li++) {
_eiLayers[li].eActivate(cs, *pLayerInput, eta, shDecay, saDecay);
pLayerInput = &_eiLayers[li]._eLayer._statesPrev;
}
pLayerInput = &zeroImage;
// Feed back
for (int li = _eiLayers.size() - 1; li >= 0; li--) {
_eiLayers[li].iActivate(cs, *pLayerInput, eta, shDecay, saDecay);
pLayerInput = &_eiLayers[li]._iLayer._statesPrev;
}
}
void PredictorSwarm::learn(sys::ComputeSystem &cs, float reward, float gamma, const cl::Image2D &targets, std::vector<cl::Image2D> &visibleStatesPrev, cl_float2 weightAlpha, cl_float2 weightLambda, cl_float biasAlpha, cl_float activeRatio, float noise) {
// Learn weights
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
VisibleLayer &vl = _visibleLayers[vli];
VisibleLayerDesc &vld = _visibleLayerDescs[vli];
int argIndex = 0;
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, visibleStatesPrev[vli]);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, targets);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, _hiddenStates[_back]);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, _hiddenActivations[_front]);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, _hiddenStates[_front]);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, vl._weights[_back]);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, vl._weights[_front]);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, vl._qTraces[_back]);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, vl._qTraces[_front]);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, vld._size);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, vl._hiddenToVisible);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, vld._radius);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, weightAlpha);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, weightLambda);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, reward);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, gamma);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, activeRatio);
_learnWeightsTracesInhibitedKernel.setArg(argIndex++, noise);
cs.getQueue().enqueueNDRangeKernel(_learnWeightsTracesInhibitedKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
std::swap(vl._weights[_front], vl._weights[_back]);
std::swap(vl._qTraces[_front], vl._qTraces[_back]);
}
}
void HEInet::sumSpikes(sys::ComputeSystem &cs, float scalar) {
int index = 0;
_kernels->_sumSpikesKernel.setArg(index++, _eiLayers.front()._eLayer._states);
_kernels->_sumSpikesKernel.setArg(index++, _eSpikeSumsPrev);
_kernels->_sumSpikesKernel.setArg(index++, _eSpikeSums);
_kernels->_sumSpikesKernel.setArg(index++, scalar);
cs.getQueue().enqueueNDRangeKernel(_kernels->_sumSpikesKernel, cl::NullRange, cl::NDRange(_eiLayers.front().getConfig()._eWidth, _eiLayers.front().getConfig()._eHeight));
index = 0;
_kernels->_sumSpikesKernel.setArg(index++, _eiLayers.front()._iLayer._states);
_kernels->_sumSpikesKernel.setArg(index++, _iSpikeSumsPrev);
_kernels->_sumSpikesKernel.setArg(index++, _iSpikeSums);
_kernels->_sumSpikesKernel.setArg(index++, scalar);
cs.getQueue().enqueueNDRangeKernel(_kernels->_sumSpikesKernel, cl::NullRange, cl::NDRange(_eiLayers.front().getConfig()._iWidth, _eiLayers.front().getConfig()._iHeight));
}
void AgentSwarm::clearMemory(sys::ComputeSystem &cs) {
cl_float4 zeroColor = { 0.0f, 0.0f, 0.0f, 0.0f };
cl::array<cl::size_type, 3> zeroOrigin = { 0, 0, 0 };
for (int l = 0; l < _layers.size(); l++) {
cl::array<cl::size_type, 3> layerRegion = { _layerDescs[l]._hiddenSize.x, _layerDescs[l]._hiddenSize.y, 1 };
cs.getQueue().enqueueFillImage(_layers[l]._scHiddenStatesPrev, zeroColor, zeroOrigin, layerRegion);
}
}
void HEInet::setInputPhase(sys::ComputeSystem &cs, cl_uint4 color) {
cl::size_t<3> zeroCoord;
zeroCoord[0] = zeroCoord[1] = zeroCoord[2] = 0;
cl::size_t<3> eFeedForwardDimsCoord;
eFeedForwardDimsCoord[0] = _eiLayers.front().getConfig()._eFeedForwardWidth;
eFeedForwardDimsCoord[1] = _eiLayers.front().getConfig()._eFeedForwardHeight;
eFeedForwardDimsCoord[2] = 1;
cs.getQueue().enqueueFillImage(_inputSpikeTimersPrev, color, zeroCoord, eFeedForwardDimsCoord);
}
void HEInet::setInputPhase(sys::ComputeSystem &cs, const cl::Image2D &inputPhaseImage) {
cl::size_t<3> zeroCoord;
zeroCoord[0] = zeroCoord[1] = zeroCoord[2] = 0;
cl::size_t<3> eFeedForwardDimsCoord;
eFeedForwardDimsCoord[0] = _eiLayers.front().getConfig()._eFeedForwardWidth;
eFeedForwardDimsCoord[1] = _eiLayers.front().getConfig()._eFeedForwardHeight;
eFeedForwardDimsCoord[2] = 1;
cs.getQueue().enqueueCopyImage(inputPhaseImage, _inputSpikeTimersPrev, zeroCoord, zeroCoord, eFeedForwardDimsCoord);
}
void HEInet::spikeSumBegin(sys::ComputeSystem &cs) {
cl_float4 zeroColor = { 0.0f, 0.0f, 0.0f, 0.0f };
cl::size_t<3> zeroCoord;
zeroCoord[0] = zeroCoord[1] = zeroCoord[2] = 0;
cl::size_t<3> eDims;
eDims[0] = _eiLayers.front().getConfig()._eWidth;
eDims[1] = _eiLayers.front().getConfig()._eHeight;
eDims[2] = 1;
cl::size_t<3> iDims;
iDims[0] = _eiLayers.front().getConfig()._iWidth;
iDims[1] = _eiLayers.front().getConfig()._iHeight;
iDims[2] = 1;
cs.getQueue().enqueueFillImage(_eSpikeSums, zeroColor, zeroCoord, eDims);
cs.getQueue().enqueueFillImage(_eSpikeSumsPrev, zeroColor, zeroCoord, eDims);
cs.getQueue().enqueueFillImage(_iSpikeSums, zeroColor, zeroCoord, iDims);
cs.getQueue().enqueueFillImage(_iSpikeSumsPrev, zeroColor, zeroCoord, iDims);
}
void neo::randomUniform(cl::Image3D &image3D, sys::ComputeSystem &cs, cl::Kernel &randomUniform3DKernel, cl_int3 size, cl_float2 range, std::mt19937 &rng) {
int argIndex = 0;
std::uniform_int_distribution<int> seedDist;
cl_uint2 seed = { seedDist(rng), seedDist(rng) };
randomUniform3DKernel.setArg(argIndex++, image3D);
randomUniform3DKernel.setArg(argIndex++, seed);
randomUniform3DKernel.setArg(argIndex++, range);
cs.getQueue().enqueueNDRangeKernel(randomUniform3DKernel, cl::NullRange, cl::NDRange(size.x, size.y, size.z));
}
void HTFE::clearMemory(sys::ComputeSystem &cs) {
// ------------------------------------------------------------------------------
// -------------------------------- Clear Memory --------------------------------
// ------------------------------------------------------------------------------
cl_uint4 clear = { 0, 0, 0, 0 };
for (int l = 0; l < _layers.size(); l++) {
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = _layerDescs[l]._width;
region[1] = _layerDescs[l]._height;
region[2] = 1;
cs.getQueue().enqueueFillImage(_layers[l]._hiddenStatesFeedBackPrevPrev, clear, origin, region);
cs.getQueue().enqueueFillImage(_layers[l]._hiddenStatesFeedBackPrev, clear, origin, region);
cs.getQueue().enqueueFillImage(_layers[l]._hiddenStatesFeedBack, clear, origin, region);
}
}
void ComparisonSparseCoder::reconstruct(sys::ComputeSystem &cs, const cl::Image2D &hiddenStates, int visibleLayerIndex, cl::Image2D &visibleStates) {
VisibleLayer &vl = _visibleLayers[visibleLayerIndex];
VisibleLayerDesc &vld = _visibleLayerDescs[visibleLayerIndex];
int argIndex = 0;
_forwardKernel.setArg(argIndex++, hiddenStates);
_forwardKernel.setArg(argIndex++, visibleStates);
_forwardKernel.setArg(argIndex++, vl._weights[_back]);
_forwardKernel.setArg(argIndex++, vld._size);
_forwardKernel.setArg(argIndex++, _hiddenSize);
_forwardKernel.setArg(argIndex++, vl._visibleToHidden);
_forwardKernel.setArg(argIndex++, vl._hiddenToVisible);
_forwardKernel.setArg(argIndex++, vld._radius);
_forwardKernel.setArg(argIndex++, vl._reverseRadii);
cs.getQueue().enqueueNDRangeKernel(_forwardKernel, cl::NullRange, cl::NDRange(vld._size.x, vld._size.y));
}
void HTFE::activate(sys::ComputeSystem &cs) {
{
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = _inputWidth;
region[1] = _inputHeight;
region[2] = 1;
cs.getQueue().enqueueWriteImage(_inputImage, CL_TRUE, origin, region, 0, 0, _input.data());
}
std::uniform_int_distribution<int> seedDist(0, 99999);
// ------------------------------------------------------------------------------
// ------------------------------------ Go up -----------------------------------
// ------------------------------------------------------------------------------
cl::Image2D* pPrevLayer = &_inputImage;
int prevWidth = _inputWidth;
int prevHeight = _inputHeight;
for (int l = 0; l < _layers.size(); l++) {
float localActivity = std::round(_layerDescs[l]._sparsity * std::pow(2 * _layerDescs[l]._inhibitionRadius + 1, 2));
Int2 layerSize;
layerSize._x = _layerDescs[l]._width;
layerSize._y = _layerDescs[l]._height;
Int2 layerSizeMinusOne;
layerSizeMinusOne._x = _layerDescs[l]._width - 1;
layerSizeMinusOne._y = _layerDescs[l]._height - 1;
Float2 layerSizeMinusOneInv;
layerSizeMinusOneInv._x = 1.0f / (_layerDescs[l]._width - 1);
layerSizeMinusOneInv._y = 1.0f / (_layerDescs[l]._height - 1);
Int2 inputSize;
inputSize._x = prevWidth;
inputSize._y = prevHeight;
Int2 inputSizeMinusOne;
inputSizeMinusOne._x = prevWidth - 1;
inputSizeMinusOne._y = prevHeight - 1;
Float2 inputSizeMinusOneInv;
inputSizeMinusOneInv._x = 1.0f / (prevWidth - 1);
inputSizeMinusOneInv._y = 1.0f / (prevHeight - 1);
// -------------------------------- Activate --------------------------------
int index = 0;
_layerHiddenFeedForwardActivateKernel.setArg(index++, *pPrevLayer);
_layerHiddenFeedForwardActivateKernel.setArg(index++, _layers[l]._hiddenStatesFeedBackPrev);
_layerHiddenFeedForwardActivateKernel.setArg(index++, _layers[l]._feedForwardWeightsPrev);
_layerHiddenFeedForwardActivateKernel.setArg(index++, _layers[l]._lateralWeightsPrev);
_layerHiddenFeedForwardActivateKernel.setArg(index++, _layers[l]._hiddenBiasesPrev);
_layerHiddenFeedForwardActivateKernel.setArg(index++, _layers[l]._hiddenFeedForwardActivations);
_layerHiddenFeedForwardActivateKernel.setArg(index++, layerSize);
_layerHiddenFeedForwardActivateKernel.setArg(index++, layerSizeMinusOneInv);
_layerHiddenFeedForwardActivateKernel.setArg(index++, inputSize);
_layerHiddenFeedForwardActivateKernel.setArg(index++, inputSizeMinusOne);
_layerHiddenFeedForwardActivateKernel.setArg(index++, _layerDescs[l]._receptiveFieldRadius);
_layerHiddenFeedForwardActivateKernel.setArg(index++, _layerDescs[l]._lateralConnectionRadius);
cs.getQueue().enqueueNDRangeKernel(_layerHiddenFeedForwardActivateKernel, cl::NullRange, cl::NDRange(_layerDescs[l]._width, _layerDescs[l]._height));
// ---------------------------------- Inhibit ---------------------------------
index = 0;
_layerHiddenInhibitKernel.setArg(index++, _layers[l]._hiddenFeedForwardActivations);
_layerHiddenInhibitKernel.setArg(index++, _layers[l]._hiddenStatesFeedForwardPrev);
_layerHiddenInhibitKernel.setArg(index++, _layers[l]._hiddenStatesFeedForward);
_layerHiddenInhibitKernel.setArg(index++, layerSize);
_layerHiddenInhibitKernel.setArg(index++, _layerDescs[l]._inhibitionRadius);
_layerHiddenInhibitKernel.setArg(index++, localActivity);
cs.getQueue().enqueueNDRangeKernel(_layerHiddenInhibitKernel, cl::NullRange, cl::NDRange(_layerDescs[l]._width, _layerDescs[l]._height));
pPrevLayer = &_layers[l]._hiddenStatesFeedForward;
prevWidth = _layerDescs[l]._width;
prevHeight = _layerDescs[l]._height;
}
// ------------------------------------------------------------------------------
// -------------------------------- Go back down --------------------------------
// ------------------------------------------------------------------------------
for (int l = _layers.size() - 1; l >= 0; l--) {
if (l > 0) {
pPrevLayer = &_layers[l - 1]._hiddenStatesFeedForward;
prevWidth = _layerDescs[l - 1]._width;
prevHeight = _layerDescs[l - 1]._height;
}
else {
pPrevLayer = &_inputImage;
prevWidth = _inputWidth;
prevHeight = _inputHeight;
}
float localActivity = std::round(_layerDescs[l]._sparsity * std::pow(2 * _layerDescs[l]._inhibitionRadius + 1, 2));
Int2 layerSize;
layerSize._x = _layerDescs[l]._width;
layerSize._y = _layerDescs[l]._height;
Int2 layerSizeMinusOne;
layerSizeMinusOne._x = _layerDescs[l]._width - 1;
layerSizeMinusOne._y = _layerDescs[l]._height - 1;
Float2 layerSizeMinusOneInv;
layerSizeMinusOneInv._x = 1.0f / (_layerDescs[l]._width - 1);
layerSizeMinusOneInv._y = 1.0f / (_layerDescs[l]._height - 1);
Int2 inputSize;
inputSize._x = prevWidth;
inputSize._y = prevHeight;
Int2 inputSizeMinusOne;
inputSizeMinusOne._x = prevWidth - 1;
inputSizeMinusOne._y = prevHeight - 1;
Float2 inputSizeMinusOneInv;
inputSizeMinusOneInv._x = 1.0f / (prevWidth - 1);
inputSizeMinusOneInv._y = 1.0f / (prevHeight - 1);
Int2 nextSize;
Int2 nextSizeMinusOne;
if (l == _layers.size() - 1) {
nextSize._x = nextSize._y = 1;
nextSizeMinusOne._x = nextSizeMinusOne._y = 0;
}
else {
nextSize._x = _layerDescs[l + 1]._width;
nextSize._y = _layerDescs[l + 1]._height;
nextSizeMinusOne._x = _layerDescs[l + 1]._width - 1;
nextSizeMinusOne._y = _layerDescs[l + 1]._height - 1;
}
// -------------------------------- Activate --------------------------------
int index = 0;
if (l == _layers.size() - 1) {
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = _layerDescs[l]._width;
region[1] = _layerDescs[l]._height;
region[2] = 1;
cs.getQueue().enqueueCopyImage(_layers[l]._hiddenFeedForwardActivations, _layers[l]._hiddenFeedBackActivations, origin, origin, region);
}
else {
_layerHiddenFeedBackActivateKernel.setArg(index++, _layers[l]._hiddenFeedForwardActivations);
_layerHiddenFeedBackActivateKernel.setArg(index++, _layers[l + 1]._hiddenFeedBackActivations);
_layerHiddenFeedBackActivateKernel.setArg(index++, _layers[l]._feedBackWeightsPrev);
_layerHiddenFeedBackActivateKernel.setArg(index++, _layers[l]._hiddenFeedBackActivations);
_layerHiddenFeedBackActivateKernel.setArg(index++, layerSize);
_layerHiddenFeedBackActivateKernel.setArg(index++, layerSizeMinusOneInv);
_layerHiddenFeedBackActivateKernel.setArg(index++, nextSize);
_layerHiddenFeedBackActivateKernel.setArg(index++, nextSizeMinusOne);
_layerHiddenFeedBackActivateKernel.setArg(index++, _layerDescs[l]._feedBackConnectionRadius);
cs.getQueue().enqueueNDRangeKernel(_layerHiddenFeedBackActivateKernel, cl::NullRange, cl::NDRange(_layerDescs[l]._width, _layerDescs[l]._height));
}
// ---------------------------------- Inhibit ---------------------------------
index = 0;
_layerHiddenInhibitKernel.setArg(index++, _layers[l]._hiddenFeedBackActivations);
_layerHiddenInhibitKernel.setArg(index++, _layers[l]._hiddenStatesFeedBackPrev);
_layerHiddenInhibitKernel.setArg(index++, _layers[l]._hiddenStatesFeedBack);
_layerHiddenInhibitKernel.setArg(index++, layerSize);
_layerHiddenInhibitKernel.setArg(index++, _layerDescs[l]._inhibitionRadius);
_layerHiddenInhibitKernel.setArg(index++, localActivity);
cs.getQueue().enqueueNDRangeKernel(_layerHiddenInhibitKernel, cl::NullRange, cl::NDRange(_layerDescs[l]._width, _layerDescs[l]._height));
// --------------------- Make Predictions (Reconstruction) ---------------------
index = 0;
_layerVisibleReconstructKernel.setArg(index++, _layers[l]._hiddenStatesFeedBack);
_layerVisibleReconstructKernel.setArg(index++, _layers[l]._reconstructionWeightsPrev);
_layerVisibleReconstructKernel.setArg(index++, _layers[l]._visibleBiasesPrev);
_layerVisibleReconstructKernel.setArg(index++, _layers[l]._visibleReconstruction);
_layerVisibleReconstructKernel.setArg(index++, _layerDescs[l]._reconstructionRadius);
_layerVisibleReconstructKernel.setArg(index++, inputSizeMinusOne);
_layerVisibleReconstructKernel.setArg(index++, inputSizeMinusOneInv);
_layerVisibleReconstructKernel.setArg(index++, layerSize);
_layerVisibleReconstructKernel.setArg(index++, layerSizeMinusOne);
_layerVisibleReconstructKernel.setArg(index++, layerSizeMinusOneInv);
cs.getQueue().enqueueNDRangeKernel(_layerVisibleReconstructKernel, cl::NullRange, cl::NDRange(prevWidth, prevHeight));
}
{
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = _inputWidth;
region[1] = _inputHeight;
region[2] = 1;
cs.getQueue().enqueueReadImage(_layers.front()._visibleReconstruction, CL_TRUE, origin, region, 0, 0, _prediction.data());
}
}
void HTFE::createRandom(sys::ComputeSystem &cs, sys::ComputeProgram &program, int inputWidth, int inputHeight, const std::vector<LayerDesc> &layerDescs, float minInitWeight, float maxInitWeight) {
std::mt19937 generator(time(nullptr));
std::uniform_int_distribution<int> seedDist(0, 99999);
_inputWidth = inputWidth;
_inputHeight = inputHeight;
_layerDescs = layerDescs;
_layers.resize(_layerDescs.size());
cl::Kernel initializeLayerHiddenKernel = cl::Kernel(program.getProgram(), "initializeLayerHidden");
cl::Kernel initializeLayerVisibleKernel = cl::Kernel(program.getProgram(), "initializeLayerVisible");
_input.clear();
_input.resize(_inputWidth * _inputHeight, 0.0f);
_prediction.clear();
_prediction.resize(_inputWidth * _inputHeight, 0.0f);
_inputImage = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _inputWidth, _inputHeight);
_inputImagePrev = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _inputWidth, _inputHeight);
{
cl_uint4 clear = { 0, 0, 0, 0 };
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = _inputWidth;
region[1] = _inputHeight;
region[2] = 1;
cs.getQueue().enqueueFillImage(_inputImage, clear, origin, region);
cs.getQueue().enqueueFillImage(_inputImagePrev, clear, origin, region);
}
int prevWidth = _inputWidth;
int prevHeight = _inputHeight;
for (int l = 0; l < _layers.size(); l++) {
int numFeedForwardWeights = std::pow(_layerDescs[l]._receptiveFieldRadius * 2 + 1, 2);
int numReconstructionWeights = std::pow(_layerDescs[l]._reconstructionRadius * 2 + 1, 2);
int numLateralWeights = std::pow(_layerDescs[l]._lateralConnectionRadius * 2 + 1, 2);
int numFeedBackWeights = std::pow(_layerDescs[l]._feedBackConnectionRadius * 2 + 1, 2);
_layers[l]._hiddenFeedForwardActivations = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_RG, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height);
_layers[l]._hiddenFeedBackActivations = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_RG, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height);
_layers[l]._hiddenFeedBackActivationsPrev = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_RG, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height);
_layers[l]._hiddenStatesFeedForward = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height);
_layers[l]._hiddenStatesFeedForwardPrev = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height);
_layers[l]._hiddenStatesFeedBack = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height);
_layers[l]._hiddenStatesFeedBackPrev = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height);
_layers[l]._hiddenStatesFeedBackPrevPrev = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height);
_layers[l]._feedForwardWeights = cl::Image3D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height, numFeedForwardWeights);
_layers[l]._feedForwardWeightsPrev = cl::Image3D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height, numFeedForwardWeights);
_layers[l]._reconstructionWeights = cl::Image3D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), prevWidth, prevHeight, numReconstructionWeights);
_layers[l]._reconstructionWeightsPrev = cl::Image3D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), prevWidth, prevHeight, numReconstructionWeights);
_layers[l]._visibleBiases = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), prevWidth, prevHeight);
_layers[l]._visibleBiasesPrev = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), prevWidth, prevHeight);
_layers[l]._hiddenBiases = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height);
_layers[l]._hiddenBiasesPrev = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height);
_layers[l]._lateralWeights = cl::Image3D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height, numLateralWeights);
_layers[l]._lateralWeightsPrev = cl::Image3D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height, numLateralWeights);
_layers[l]._feedBackWeights = cl::Image3D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height, numFeedBackWeights);
_layers[l]._feedBackWeightsPrev = cl::Image3D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._width, _layerDescs[l]._height, numFeedBackWeights);
_layers[l]._visibleReconstruction = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), prevWidth, prevHeight);
_layers[l]._visibleReconstructionPrev = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), prevWidth, prevHeight);
// Initialize
Uint2 initSeedHidden;
initSeedHidden._x = seedDist(generator);
initSeedHidden._y = seedDist(generator);
int index = 0;
initializeLayerHiddenKernel.setArg(index++, _layers[l]._hiddenFeedForwardActivations);
initializeLayerHiddenKernel.setArg(index++, _layers[l]._hiddenFeedBackActivations);
initializeLayerHiddenKernel.setArg(index++, _layers[l]._hiddenStatesFeedForward);
initializeLayerHiddenKernel.setArg(index++, _layers[l]._feedForwardWeights);
initializeLayerHiddenKernel.setArg(index++, _layers[l]._hiddenBiases);
initializeLayerHiddenKernel.setArg(index++, _layers[l]._lateralWeights);
initializeLayerHiddenKernel.setArg(index++, _layers[l]._feedBackWeights);
initializeLayerHiddenKernel.setArg(index++, numFeedForwardWeights);
initializeLayerHiddenKernel.setArg(index++, numLateralWeights);
initializeLayerHiddenKernel.setArg(index++, numFeedBackWeights);
initializeLayerHiddenKernel.setArg(index++, initSeedHidden);
initializeLayerHiddenKernel.setArg(index++, _layerDescs[l]._sparsity);
initializeLayerHiddenKernel.setArg(index++, _layerDescs[l]._lateralScalar);
initializeLayerHiddenKernel.setArg(index++, _layerDescs[l]._feedBackScalar);
initializeLayerHiddenKernel.setArg(index++, minInitWeight);
initializeLayerHiddenKernel.setArg(index++, maxInitWeight);
cs.getQueue().enqueueNDRangeKernel(initializeLayerHiddenKernel, cl::NullRange, cl::NDRange(_layerDescs[l]._width, _layerDescs[l]._height));
Uint2 initSeedVisible;
initSeedVisible._x = seedDist(generator);
initSeedVisible._y = seedDist(generator);
index = 0;
initializeLayerVisibleKernel.setArg(index++, _layers[l]._visibleBiases);
initializeLayerVisibleKernel.setArg(index++, _layers[l]._visibleReconstruction);
initializeLayerVisibleKernel.setArg(index++, _layers[l]._reconstructionWeights);
initializeLayerVisibleKernel.setArg(index++, numReconstructionWeights);
initializeLayerVisibleKernel.setArg(index++, initSeedVisible);
initializeLayerVisibleKernel.setArg(index++, minInitWeight);
initializeLayerVisibleKernel.setArg(index++, maxInitWeight);
cs.getQueue().enqueueNDRangeKernel(initializeLayerVisibleKernel, cl::NullRange, cl::NDRange(prevWidth, prevHeight));
{
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = _layerDescs[l]._width;
region[1] = _layerDescs[l]._height;
region[2] = 1;
cs.getQueue().enqueueCopyImage(_layers[l]._hiddenFeedBackActivations, _layers[l]._hiddenFeedBackActivationsPrev, origin, origin, region);
}
{
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = prevWidth;
region[1] = prevHeight;
region[2] = 1;
cs.getQueue().enqueueCopyImage(_layers[l]._visibleReconstruction, _layers[l]._visibleReconstructionPrev, origin, origin, region);
}
{
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = _layerDescs[l]._width;
region[1] = _layerDescs[l]._height;
region[2] = 1;
cs.getQueue().enqueueCopyImage(_layers[l]._hiddenStatesFeedForward, _layers[l]._hiddenStatesFeedForwardPrev, origin, origin, region);
cs.getQueue().enqueueCopyImage(_layers[l]._hiddenStatesFeedForward, _layers[l]._hiddenStatesFeedBack, origin, origin, region);
cs.getQueue().enqueueCopyImage(_layers[l]._hiddenStatesFeedForward, _layers[l]._hiddenStatesFeedBackPrev, origin, origin, region);
cs.getQueue().enqueueCopyImage(_layers[l]._hiddenStatesFeedForward, _layers[l]._hiddenStatesFeedBackPrevPrev, origin, origin, region);
}
{
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = _layerDescs[l]._width;
region[1] = _layerDescs[l]._height;
region[2] = numFeedForwardWeights;
cs.getQueue().enqueueCopyImage(_layers[l]._feedForwardWeights, _layers[l]._feedForwardWeightsPrev, origin, origin, region);
}
{
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = prevWidth;
region[1] = prevHeight;
region[2] = 1;
cs.getQueue().enqueueCopyImage(_layers[l]._visibleBiases, _layers[l]._visibleBiasesPrev, origin, origin, region);
}
{
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = _layerDescs[l]._width;
region[1] = _layerDescs[l]._height;
region[2] = 1;
cs.getQueue().enqueueCopyImage(_layers[l]._hiddenBiases, _layers[l]._hiddenBiasesPrev, origin, origin, region);
}
{
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = _layerDescs[l]._width;
region[1] = _layerDescs[l]._height;
region[2] = numLateralWeights;
cs.getQueue().enqueueCopyImage(_layers[l]._lateralWeights, _layers[l]._lateralWeightsPrev, origin, origin, region);
}
{
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = _layerDescs[l]._width;
region[1] = _layerDescs[l]._height;
region[2] = numFeedBackWeights;
cs.getQueue().enqueueCopyImage(_layers[l]._feedBackWeights, _layers[l]._feedBackWeightsPrev, origin, origin, region);
}
{
cl::size_t<3> origin;
origin[0] = 0;
origin[1] = 0;
origin[2] = 0;
cl::size_t<3> region;
region[0] = prevWidth;
region[1] = prevHeight;
region[2] = numReconstructionWeights;
cs.getQueue().enqueueCopyImage(_layers[l]._reconstructionWeights, _layers[l]._reconstructionWeightsPrev, origin, origin, region);
}
prevWidth = _layerDescs[l]._width;
prevHeight = _layerDescs[l]._height;
}
_layerHiddenFeedForwardActivateKernel = cl::Kernel(program.getProgram(), "layerHiddenFeedForwardActivate");
_layerHiddenFeedBackActivateKernel = cl::Kernel(program.getProgram(), "layerHiddenFeedBackActivate");
_layerHiddenInhibitKernel = cl::Kernel(program.getProgram(), "layerHiddenInhibit");
_layerVisibleReconstructKernel = cl::Kernel(program.getProgram(), "layerVisibleReconstruct");
_layerHiddenWeightUpdateKernel = cl::Kernel(program.getProgram(), "layerHiddenWeightUpdate");
_layerHiddenWeightUpdateLastKernel = cl::Kernel(program.getProgram(), "layerHiddenWeightUpdateLast");
_layerVisibleWeightUpdateKernel = cl::Kernel(program.getProgram(), "layerVisibleWeightUpdate");
}
void AgentER::simStep(sys::ComputeSystem &cs, const cl::Image2D &input, const cl::Image2D &actionTaken, float reward, std::mt19937 &rng, bool learn, bool whiten) {
// Keep previous best action for later
std::vector<float> prevBestAction(_actionSize.x * _actionSize.y);
std::vector<float> prevTakenAction(_actionSize.x * _actionSize.y);
cs.getQueue().enqueueReadImage(getAction(), CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_actionSize.x), static_cast<cl::size_type>(_actionSize.y), 1 }, 0, 0, prevBestAction.data());
cs.getQueue().enqueueReadImage(actionTaken, CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_actionSize.x), static_cast<cl::size_type>(_actionSize.y), 1 }, 0, 0, prevTakenAction.data());
// Place previous Q into Q buffer
{
int argIndex = 0;
_setQKernel.setArg(argIndex++, _qTransform);
_setQKernel.setArg(argIndex++, _qInput);
_setQKernel.setArg(argIndex++, _prevQ);
cs.getQueue().enqueueNDRangeKernel(_setQKernel, cl::NullRange, cl::NDRange(_qSize.x, _qSize.y));
}
// Whiten input
if (whiten)
_inputWhitener.filter(cs, input, _whiteningKernelRadius, _whiteningIntensity);
_actionWhitener.filter(cs, actionTaken, _whiteningKernelRadius, _whiteningIntensity);
_qWhitener.filter(cs, _qInput, _whiteningKernelRadius, _whiteningIntensity);
// Feed forward
for (int l = 0; l < _layers.size(); l++) {
{
std::vector<cl::Image2D> visibleStates;
if (l == 0) {
visibleStates.resize(3);
visibleStates[0] = whiten ? _inputWhitener.getResult() : input;
visibleStates[1] = _actionWhitener.getResult();
visibleStates[2] = _qWhitener.getResult();
}
else {
visibleStates.resize(2);
visibleStates[0] = _layers[l - 1]._sc.getHiddenStates()[_back];
visibleStates[1] = _layers[l]._sc.getHiddenStates()[_back];
}
_layers[l]._sc.activate(cs, visibleStates, _layerDescs[l]._scActiveRatio);
}
}
for (int l = _layers.size() - 1; l >= 0; l--) {
std::vector<cl::Image2D> visibleStates;
if (l < _layers.size() - 1) {
visibleStates.resize(2);
visibleStates[0] = _layers[l]._sc.getHiddenStates()[_back];
visibleStates[1] = _layers[l + 1]._pred.getHiddenStates()[_back];
}
else {
visibleStates.resize(1);
visibleStates[0] = _layers[l]._sc.getHiddenStates()[_back];
}
//_layers[l]._pred.activate(cs, visibleStates, l != 0);
}
// Q predictor
{
std::vector<cl::Image2D> visibleStates;
if (0 < _layers.size() - 1) {
visibleStates.resize(2);
visibleStates[0] = _layers[0]._sc.getHiddenStates()[_back];
visibleStates[1] = _layers[0 + 1]._pred.getHiddenStates()[_back];
}
else {
visibleStates.resize(1);
visibleStates[0] = _layers[0]._sc.getHiddenStates()[_back];
}
//_qPred.activate(cs, visibleStates, false);
}
// Recover Q
std::vector<float> qValues(_qSize.x * _qSize.y);
cs.getQueue().enqueueReadImage(_qPred.getHiddenStates()[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_qSize.x), static_cast<cl::size_type>(_qSize.y), 1 }, 0, 0, qValues.data());
// Average all Q values
float q = 0.0f;
for (int i = 0; i < qValues.size(); i++)
q += qValues[i];
q /= qValues.size();
// Bellman equation
float tdError = reward + _qGamma * q - _prevValue;
float newQ = _prevValue + _qAlpha * tdError;
// Update older samples
float g = _qGamma;
for (std::list<ReplayFrame>::iterator it = _frames.begin(); it != _frames.end(); it++) {
it->_q += g * tdError;
g *= _qGamma;
}
// Add replay sample
ReplayFrame frame;
frame._q = frame._originalQ = newQ;
frame._layerStateBitIndices.resize(_layers.size());
frame._layerPredBitIndices.resize(_layers.size());
for (int l = 0; l < _layers.size(); l++) {
std::vector<float> state(_layerDescs[l]._size.x * _layerDescs[l]._size.y);
cs.getQueue().enqueueReadImage(_layers[l]._sc.getHiddenStates()[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_layerDescs[l]._size.x), static_cast<cl::size_type>(_layerDescs[l]._size.y), 1 }, 0, 0, state.data());
std::vector<float> pred;
if (l == 0) {
pred.resize(_actionSize.x * _actionSize.y);
cs.getQueue().enqueueReadImage(_layers[l]._sc.getHiddenStates()[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_actionSize.x), static_cast<cl::size_type>(_actionSize.y), 1 }, 0, 0, state.data());
}
else {
pred.resize(_layerDescs[l - 1]._size.x * _layerDescs[l - 1]._size.y);
cs.getQueue().enqueueReadImage(_layers[l]._sc.getHiddenStates()[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_layerDescs[l - 1]._size.x), static_cast<cl::size_type>(_layerDescs[l - 1]._size.y), 1 }, 0, 0, pred.data());
}
for (int i = 0; i < state.size(); i++)
if (state[i] > 0.0f)
frame._layerStateBitIndices[l].push_back(i);
for (int i = 0; i < pred.size(); i++)
if (pred[i] > 0.0f)
frame._layerPredBitIndices[l].push_back(i);
}
// Add last action taken and last "thought best" action
frame._prevExploratoryAction = prevTakenAction;
frame._prevBestAction = prevBestAction;
for (int i = 0; i < prevBestAction.size(); i++)
frame._prevBestAction[i] = std::min(1.0f, std::max(-1.0f, prevBestAction[i]));
_frames.push_front(frame);
while (_frames.size() > _maxReplayFrames)
_frames.pop_back();
if (learn && _frames.size() > 1) {
// Convert list to vector
std::vector<ReplayFrame*> pFrames(_frames.size());
int index = 0;
for (std::list<ReplayFrame>::iterator it = _frames.begin(); it != _frames.end(); it++)
pFrames[index++] = &(*it);
std::uniform_int_distribution<int> replayDist(0, _frames.size() - 2);
for (int iter = 0; iter < _replayIterations; iter++) {
int randIndex = replayDist(rng);
ReplayFrame* pFrame = pFrames[randIndex];
ReplayFrame* pFramePrev = pFrames[randIndex + 1];
// Load data
cl_int2 prevLayerSize = _actionSize;
for (int l = 0; l < _layers.size(); l++) {
std::vector<float> state(_layerDescs[l]._size.x * _layerDescs[l]._size.y, 0.0f);
std::vector<float> statePrev(_layerDescs[l]._size.x * _layerDescs[l]._size.y, 0.0f);
std::vector<float> pred(prevLayerSize.x * prevLayerSize.y, 0.0f);
std::vector<float> predPrev(prevLayerSize.x * prevLayerSize.y, 0.0f);
for (int i = 0; i < pFrame->_layerStateBitIndices[l].size(); i++)
state[pFrame->_layerStateBitIndices[l][i]] = 1.0f;
for (int i = 0; i < pFramePrev->_layerStateBitIndices[l].size(); i++)
statePrev[pFramePrev->_layerStateBitIndices[l][i]] = 1.0f;
for (int i = 0; i < pFrame->_layerPredBitIndices[l].size(); i++)
pred[pFrame->_layerPredBitIndices[l][i]] = 1.0f;
for (int i = 0; i < pFramePrev->_layerPredBitIndices[l].size(); i++)
predPrev[pFramePrev->_layerPredBitIndices[l][i]] = 1.0f;
cs.getQueue().enqueueWriteImage(_layers[l]._scStatesTemp[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_layerDescs[l]._size.x), static_cast<cl::size_type>(_layerDescs[l]._size.y), 1 }, 0, 0, state.data());
cs.getQueue().enqueueWriteImage(_layers[l]._scStatesTemp[_front], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_layerDescs[l]._size.x), static_cast<cl::size_type>(_layerDescs[l]._size.y), 1 }, 0, 0, statePrev.data());
cs.getQueue().enqueueWriteImage(_layers[l]._predStatesTemp[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(prevLayerSize.x), static_cast<cl::size_type>(prevLayerSize.y), 1 }, 0, 0, pred.data());
cs.getQueue().enqueueWriteImage(_layers[l]._predStatesTemp[_front], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(prevLayerSize.x), static_cast<cl::size_type>(prevLayerSize.y), 1 }, 0, 0, predPrev.data());
prevLayerSize = _layerDescs[l]._size;
}
cs.getQueue().enqueueFillImage(_qTarget, cl_float4{ pFrame->_q, pFrame->_q, pFrame->_q, pFrame->_q }, { 0, 0, 0 }, { static_cast<cl::size_type>(_qSize.x), static_cast<cl::size_type>(_qSize.y), 1 });
// Choose better action to learn
cs.getQueue().enqueueWriteImage(_actionTarget, CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_actionSize.x), static_cast<cl::size_type>(_actionSize.y), 1 }, 0, 0,
(pFrame->_q > pFrame->_originalQ ? pFrame->_prevExploratoryAction.data() : pFrame->_prevBestAction.data()));
for (int l = 0; l < _layers.size(); l++) {
std::vector<cl::Image2D> visibleStates;
if (l != 0) {
visibleStates.resize(2);
visibleStates[0] = _layers[l - 1]._sc.getHiddenStates()[_back];
visibleStates[1] = _layers[l]._sc.getHiddenStates()[_back];
_layers[l]._sc.activate(cs, visibleStates, _layerDescs[l]._scActiveRatio, false);
_layers[l]._sc.learn(cs, visibleStates, _layerDescs[l]._scBoostAlpha, _layerDescs[l]._scActiveRatio);
}
std::vector<cl::Image2D> visibleStatesPrev;
if (l < _layers.size() - 1) {
visibleStatesPrev.resize(2);
visibleStatesPrev[0] = _layers[l]._scStatesTemp[_front];
visibleStatesPrev[1] = _layers[l + 1]._predStatesTemp[_front];
}
else {
visibleStatesPrev.resize(1);
visibleStatesPrev[0] = _layers[l]._scStatesTemp[_front];
}
//_layers[l]._pred.activate(cs, visibleStatesPrev, l != 0, false);
if (l == 0)
_layers[l]._pred.learnCurrent(cs, _actionTarget, visibleStatesPrev, _layerDescs[l]._predWeightAlpha);
else
_layers[l]._pred.learnCurrent(cs, _layers[l - 1]._sc.getHiddenStates()[_back], visibleStatesPrev, _layerDescs[l]._predWeightAlpha);
}
// Q Pred
{
std::vector<cl::Image2D> visibleStatesPrev;
if (0 < _layers.size() - 1) {
visibleStatesPrev.resize(2);
visibleStatesPrev[0] = _layers[0]._sc.getHiddenStates()[_front];
visibleStatesPrev[1] = _layers[0 + 1]._pred.getHiddenStates()[_front];
}
else {
visibleStatesPrev.resize(1);
visibleStatesPrev[0] = _layers[0]._sc.getHiddenStates()[_front];
}
//_qPred.activate(cs, visibleStatesPrev, false, false);
_qPred.learnCurrent(cs, _qTarget, visibleStatesPrev, _qWeightAlpha);
}
}
}
std::cout << "Q: " << newQ << std::endl;
_prevQ = newQ;
_prevTDError = tdError;
_prevValue = q;
}
void ComparisonSparseCoder::activate(sys::ComputeSystem &cs, const std::vector<cl::Image2D> &visibleStates, float activeRatio, bool bufferSwap) {
// Start by clearing summation buffer to biases
{
cl::array<cl::size_type, 3> zeroOrigin = { 0, 0, 0 };
cl::array<cl::size_type, 3> hiddenRegion = { _hiddenSize.x, _hiddenSize.y, 1 };
cs.getQueue().enqueueCopyImage(_hiddenBiases[_back], _hiddenActivationSummationTemp[_back], zeroOrigin, zeroOrigin, hiddenRegion);
}
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
VisibleLayer &vl = _visibleLayers[vli];
VisibleLayerDesc &vld = _visibleLayerDescs[vli];
if (!vld._isPredictiveCoding) {
if (vld._ignoreMiddle) {
int argIndex = 0;
_activateIgnoreMiddleKernel.setArg(argIndex++, visibleStates[vli]);
_activateIgnoreMiddleKernel.setArg(argIndex++, _hiddenActivationSummationTemp[_back]);
_activateIgnoreMiddleKernel.setArg(argIndex++, _hiddenActivationSummationTemp[_front]);
_activateIgnoreMiddleKernel.setArg(argIndex++, vl._weights[_back]);
_activateIgnoreMiddleKernel.setArg(argIndex++, vld._size);
_activateIgnoreMiddleKernel.setArg(argIndex++, vl._hiddenToVisible);
_activateIgnoreMiddleKernel.setArg(argIndex++, vld._radius);
cs.getQueue().enqueueNDRangeKernel(_activateIgnoreMiddleKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
}
else {
int argIndex = 0;
_activateKernel.setArg(argIndex++, visibleStates[vli]);
_activateKernel.setArg(argIndex++, _hiddenActivationSummationTemp[_back]);
_activateKernel.setArg(argIndex++, _hiddenActivationSummationTemp[_front]);
_activateKernel.setArg(argIndex++, vl._weights[_back]);
_activateKernel.setArg(argIndex++, vld._size);
_activateKernel.setArg(argIndex++, vl._hiddenToVisible);
_activateKernel.setArg(argIndex++, vld._radius);
cs.getQueue().enqueueNDRangeKernel(_activateKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
}
// Swap buffers
std::swap(_hiddenActivationSummationTemp[_front], _hiddenActivationSummationTemp[_back]);
}
}
// Start by clearing summation buffer to biases
{
cl::array<cl::size_type, 3> zeroOrigin = { 0, 0, 0 };
cl::array<cl::size_type, 3> hiddenRegion = { _hiddenSize.x, _hiddenSize.y, 1 };
cs.getQueue().enqueueFillImage(_hiddenPredictionSummationTemp[_back], cl_float4{ 0.0f, 0.0f, 0.0f, 0.0f }, zeroOrigin, hiddenRegion);
}
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
VisibleLayer &vl = _visibleLayers[vli];
VisibleLayerDesc &vld = _visibleLayerDescs[vli];
if (vld._isPredictiveCoding) {
if (vld._ignoreMiddle) {
int argIndex = 0;
_activateIgnoreMiddleKernel.setArg(argIndex++, visibleStates[vli]);
_activateIgnoreMiddleKernel.setArg(argIndex++, _hiddenPredictionSummationTemp[_back]);
_activateIgnoreMiddleKernel.setArg(argIndex++, _hiddenPredictionSummationTemp[_front]);
_activateIgnoreMiddleKernel.setArg(argIndex++, vl._weights[_back]);
_activateIgnoreMiddleKernel.setArg(argIndex++, vld._size);
_activateIgnoreMiddleKernel.setArg(argIndex++, vl._hiddenToVisible);
_activateIgnoreMiddleKernel.setArg(argIndex++, vld._radius);
cs.getQueue().enqueueNDRangeKernel(_activateIgnoreMiddleKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
}
else {
int argIndex = 0;
_activateKernel.setArg(argIndex++, visibleStates[vli]);
_activateKernel.setArg(argIndex++, _hiddenPredictionSummationTemp[_back]);
_activateKernel.setArg(argIndex++, _hiddenPredictionSummationTemp[_front]);
_activateKernel.setArg(argIndex++, vl._weights[_back]);
_activateKernel.setArg(argIndex++, vld._size);
_activateKernel.setArg(argIndex++, vl._hiddenToVisible);
_activateKernel.setArg(argIndex++, vld._radius);
cs.getQueue().enqueueNDRangeKernel(_activateKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
}
// Swap buffers
std::swap(_hiddenPredictionSummationTemp[_front], _hiddenPredictionSummationTemp[_back]);
}
}
// Back now contains the sums. Solve sparse codes from this
{
int argIndex = 0;
_solveHiddenKernel.setArg(argIndex++, _hiddenActivationSummationTemp[_back]);
_solveHiddenKernel.setArg(argIndex++, _hiddenPredictionSummationTemp[_back]);
_solveHiddenKernel.setArg(argIndex++, _hiddenStates[_front]);
_solveHiddenKernel.setArg(argIndex++, _hiddenSize);
_solveHiddenKernel.setArg(argIndex++, _lateralRadius);
_solveHiddenKernel.setArg(argIndex++, activeRatio);
cs.getQueue().enqueueNDRangeKernel(_solveHiddenKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
}
// Swap hidden state buffers
//if (bufferSwap)
std::swap(_hiddenStates[_front], _hiddenStates[_back]);
}
void ComparisonSparseCoder::createRandom(sys::ComputeSystem &cs, sys::ComputeProgram &program,
const std::vector<VisibleLayerDesc> &visibleLayerDescs,
cl_int2 hiddenSize, cl_int lateralRadius, cl_float2 initWeightRange,
std::mt19937 &rng)
{
_visibleLayerDescs = visibleLayerDescs;
_lateralRadius = lateralRadius;
_hiddenSize = hiddenSize;
cl_float4 zeroColor = { 0.0f, 0.0f, 0.0f, 0.0f };
cl::array<cl::size_type, 3> zeroOrigin = { 0, 0, 0 };
cl::array<cl::size_type, 3> hiddenRegion = { _hiddenSize.x, _hiddenSize.y, 1 };
_visibleLayers.resize(_visibleLayerDescs.size());
cl::Kernel randomUniform2DKernel = cl::Kernel(program.getProgram(), "randomUniform2D");
cl::Kernel randomUniform3DKernel = cl::Kernel(program.getProgram(), "randomUniform3D");
// Create layers
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
VisibleLayer &vl = _visibleLayers[vli];
VisibleLayerDesc &vld = _visibleLayerDescs[vli];
const cl_channel_order weightChannels = vld._useTraces ? CL_RG : CL_R;
vl._hiddenToVisible = cl_float2{ static_cast<float>(vld._size.x) / static_cast<float>(_hiddenSize.x),
static_cast<float>(vld._size.y) / static_cast<float>(_hiddenSize.y)
};
vl._visibleToHidden = cl_float2{ static_cast<float>(_hiddenSize.x) / static_cast<float>(vld._size.x),
static_cast<float>(_hiddenSize.y) / static_cast<float>(vld._size.y)
};
vl._reverseRadii = { static_cast<int>(std::ceil(vl._visibleToHidden.x * (vld._radius + 0.5f))), static_cast<int>(std::ceil(vl._visibleToHidden.y * (vld._radius + 0.5f))) };
// Create images
{
int weightDiam = vld._radius * 2 + 1;
int numWeights = weightDiam * weightDiam;
cl_int3 weightsSize = cl_int3{ _hiddenSize.x, _hiddenSize.y, numWeights };
vl._weights = createDoubleBuffer3D(cs, weightsSize, weightChannels, CL_FLOAT);
randomUniform(vl._weights[_back], cs, randomUniform3DKernel, weightsSize, initWeightRange, rng);
}
}
// Hidden state data
_hiddenStates = createDoubleBuffer2D(cs, _hiddenSize, CL_R, CL_FLOAT);
_hiddenBiases = createDoubleBuffer2D(cs, _hiddenSize, CL_R, CL_FLOAT);
//randomUniform(_hiddenBiases[_back], cs, randomUniform2DKernel, _hiddenSize, initWeightRange, rng);
cs.getQueue().enqueueFillImage(_hiddenBiases[_back], zeroColor, zeroOrigin, hiddenRegion);
_hiddenActivationSummationTemp = createDoubleBuffer2D(cs, _hiddenSize, CL_R, CL_FLOAT);
_hiddenPredictionSummationTemp = createDoubleBuffer2D(cs, _hiddenSize, CL_R, CL_FLOAT);
cs.getQueue().enqueueFillImage(_hiddenStates[_back], zeroColor, zeroOrigin, hiddenRegion);
// Create kernels
_activateKernel = cl::Kernel(program.getProgram(), "cscActivate");
_activateIgnoreMiddleKernel = cl::Kernel(program.getProgram(), "cscActivateIgnoreMiddle");
_solveHiddenKernel = cl::Kernel(program.getProgram(), "cscSolveHidden");
_learnHiddenBiasesKernel = cl::Kernel(program.getProgram(), "cscLearnHiddenBiases");
_learnHiddenWeightsActivationKernel = cl::Kernel(program.getProgram(), "cscLearnHiddenWeightsActivation");
_learnHiddenWeightsTracesActivationKernel = cl::Kernel(program.getProgram(), "cscLearnHiddenWeightsTracesActivation");
_learnHiddenWeightsPredictionKernel = cl::Kernel(program.getProgram(), "cscLearnHiddenWeightsPrediction");
_learnHiddenWeightsTracesPredictionKernel = cl::Kernel(program.getProgram(), "cscLearnHiddenWeightsTracesPrediction");
_forwardKernel = cl::Kernel(program.getProgram(), "cscForward");
}
void AgentSwarm::createRandom(sys::ComputeSystem &cs, sys::ComputeProgram &program,
cl_int2 inputSize, cl_int2 actionSize, cl_int firstLayerPredictorRadius, const std::vector<LayerDesc> &layerDescs,
cl_float2 initWeightRange,
std::mt19937 &rng)
{
_layerDescs = layerDescs;
_layers.resize(_layerDescs.size());
cl_int2 prevLayerSize = inputSize;
for (int l = 0; l < _layers.size(); l++) {
std::vector<ComparisonSparseCoder::VisibleLayerDesc> scDescs(2);
scDescs[0]._size = prevLayerSize;
scDescs[0]._radius = _layerDescs[l]._feedForwardRadius;
scDescs[0]._ignoreMiddle = false;
scDescs[0]._weightAlpha = _layerDescs[l]._scWeightAlpha;
scDescs[0]._weightLambda = _layerDescs[l]._scWeightLambda;
scDescs[0]._useTraces = false;
scDescs[1]._size = _layerDescs[l]._hiddenSize;
scDescs[1]._radius = _layerDescs[l]._recurrentRadius;
scDescs[1]._ignoreMiddle = true;
scDescs[1]._weightAlpha = _layerDescs[l]._scWeightRecurrentAlpha;
scDescs[1]._weightLambda = _layerDescs[l]._scWeightLambda;
scDescs[1]._useTraces = false;
_layers[l]._sc.createRandom(cs, program, scDescs, _layerDescs[l]._hiddenSize, _layerDescs[l]._lateralRadius, initWeightRange, rng);
_layers[l]._modulatedFeedForwardInput = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), prevLayerSize.x, prevLayerSize.y);
_layers[l]._modulatedRecurrentInput = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._hiddenSize.x, _layerDescs[l]._hiddenSize.y);
std::vector<Predictor::VisibleLayerDesc> predDescs;
if (l < _layers.size() - 1) {
predDescs.resize(2);
predDescs[0]._size = _layerDescs[l]._hiddenSize;
predDescs[0]._radius = _layerDescs[l]._predictiveRadius;
predDescs[1]._size = _layerDescs[l + 1]._hiddenSize;
predDescs[1]._radius = _layerDescs[l]._feedBackRadius;
}
else {
predDescs.resize(1);
predDescs[0]._size = _layerDescs[l]._hiddenSize;
predDescs[0]._radius = _layerDescs[l]._predictiveRadius;
}
_layers[l]._pred.createRandom(cs, program, predDescs, prevLayerSize, initWeightRange, false, rng);
std::vector<Swarm::VisibleLayerDesc> swarmDescs;
if (l == 0) {
swarmDescs.resize(3);
swarmDescs[0]._size = inputSize;
swarmDescs[0]._qRadius = _layerDescs[l]._qRadiusHiddenFeedForwardAttention;
swarmDescs[0]._startRadius = _layerDescs[l]._startRadiusHiddenFeedForwardAttention;
swarmDescs[1]._size = _layerDescs[l]._hiddenSize;
swarmDescs[1]._qRadius = _layerDescs[l]._qRadiusHiddenRecurrentAttention;
swarmDescs[1]._startRadius = _layerDescs[l]._startRadiusHiddenRecurrentAttention;
swarmDescs[2]._size = actionSize;
swarmDescs[2]._qRadius = _layerDescs[l]._qRadiusHiddenAction;
swarmDescs[2]._startRadius = _layerDescs[l]._startRadiusHiddenAction;
}
else {
swarmDescs.resize(3);
swarmDescs[0]._size = _layerDescs[l - 1]._hiddenSize;
swarmDescs[0]._qRadius = _layerDescs[l]._qRadiusHiddenFeedForwardAttention;
swarmDescs[0]._startRadius = _layerDescs[l]._startRadiusHiddenFeedForwardAttention;
swarmDescs[1]._size = _layerDescs[l]._hiddenSize;
swarmDescs[1]._qRadius = _layerDescs[l]._qRadiusHiddenRecurrentAttention;
swarmDescs[1]._startRadius = _layerDescs[l]._startRadiusHiddenRecurrentAttention;
swarmDescs[2]._size = _layerDescs[l - 1]._hiddenSize;
swarmDescs[2]._qRadius = _layerDescs[l]._qRadiusHiddenAction;
swarmDescs[2]._startRadius = _layerDescs[l]._startRadiusHiddenAction;
}
_layers[l]._swarm.createRandom(cs, program, swarmDescs, _layerDescs[l]._qSize, _layerDescs[l]._hiddenSize, _layerDescs[l]._qRadius, initWeightRange, rng);
// Create baselines
_layers[l]._baseLines = createDoubleBuffer2D(cs, _layerDescs[l]._hiddenSize, CL_R, CL_FLOAT);
_layers[l]._reward = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._hiddenSize.x, _layerDescs[l]._hiddenSize.y);
_layers[l]._scHiddenStatesPrev = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._hiddenSize.x, _layerDescs[l]._hiddenSize.y);
cl_float4 zeroColor = { 0.0f, 0.0f, 0.0f, 0.0f };
cl::array<cl::size_type, 3> zeroOrigin = { 0, 0, 0 };
if (l != 0) {
cl::array<cl::size_type, 3> actionRegion = { _layers[l]._swarm.getVisibleLayerDesc(2)._size.x, _layers[l]._swarm.getVisibleLayerDesc(2)._size.y, 1 };
_layers[l]._inhibitedAction = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), swarmDescs[1]._size.x, swarmDescs[1]._size.y);
cs.getQueue().enqueueFillImage(_layers[l]._inhibitedAction, zeroColor, zeroOrigin, actionRegion);
}
cl::array<cl::size_type, 3> layerRegion = { _layerDescs[l]._hiddenSize.x, _layerDescs[l]._hiddenSize.y, 1 };
cs.getQueue().enqueueFillImage(_layers[l]._baseLines[_back], zeroColor, zeroOrigin, layerRegion);
cs.getQueue().enqueueFillImage(_layers[l]._reward, zeroColor, zeroOrigin, layerRegion);
cs.getQueue().enqueueFillImage(_layers[l]._scHiddenStatesPrev, zeroColor, zeroOrigin, layerRegion);
prevLayerSize = _layerDescs[l]._hiddenSize;
}
{
cl_float4 zeroColor = { 0.0f, 0.0f, 0.0f, 0.0f };
cl::array<cl::size_type, 3> zeroOrigin = { 0, 0, 0 };
cl::array<cl::size_type, 3> layerRegion = { _layerDescs.back()._hiddenSize.x, _layerDescs.back()._hiddenSize.y, 1 };
_lastLayerAction = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs.back()._hiddenSize.x, _layerDescs.back()._hiddenSize.y);
cs.getQueue().enqueueFillImage(_lastLayerAction, zeroColor, zeroOrigin, layerRegion);
}
_baseLineUpdateKernel = cl::Kernel(program.getProgram(), "phBaseLineUpdate");
_baseLineUpdateSumErrorKernel = cl::Kernel(program.getProgram(), "phBaseLineUpdateSumError");
_inhibitKernel = cl::Kernel(program.getProgram(), "phInhibit");
_modulateKernel = cl::Kernel(program.getProgram(), "phModulate");
}
void ComparisonSparseCoder::writeToStream(sys::ComputeSystem &cs, std::ostream &os) const {
abort(); // Fix me
os << _hiddenSize.x << " " << _hiddenSize.y << " " << _lateralRadius << std::endl;
{
std::vector<cl_float> hiddenStates(_hiddenSize.x * _hiddenSize.y);
cs.getQueue().enqueueReadImage(_hiddenStates[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_hiddenSize.x), static_cast<cl::size_type>(_hiddenSize.y), 1 }, 0, 0, hiddenStates.data());
for (int si = 0; si < hiddenStates.size(); si++)
os << hiddenStates[si] << " ";
os << std::endl;
}
{
std::vector<cl_float> hiddenBiases(_hiddenSize.x * _hiddenSize.y);
cs.getQueue().enqueueReadImage(_hiddenBiases[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_hiddenSize.x), static_cast<cl::size_type>(_hiddenSize.y), 1 }, 0, 0, hiddenBiases.data());
for (int bi = 0; bi < hiddenBiases.size(); bi++)
os << hiddenBiases[bi] << " ";
os << std::endl;
}
// Layer information
os << _visibleLayers.size() << std::endl;
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
const VisibleLayer &vl = _visibleLayers[vli];
const VisibleLayerDesc &vld = _visibleLayerDescs[vli];
// Desc
os << vld._size.x << " " << vld._size.y << " " << vld._radius << " " << vld._weightAlpha << " " << vld._weightLambda << " " << vld._ignoreMiddle << " " << vld._useTraces << std::endl;
// Layer
int weightDiam = vld._radius * 2 + 1;
int numWeights = weightDiam * weightDiam;
cl_int3 weightsSize = cl_int3{ _hiddenSize.x, _hiddenSize.y, numWeights };
int totalNumWeights = weightsSize.x * weightsSize.y * weightsSize.z;
if (vld._useTraces) {
std::vector<cl_float2> weights(totalNumWeights);
//cs.getQueue().enqueueReadImage(vl._weights[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(weightsSize.x), static_cast<cl::size_type>(weightsSize.y), static_cast<cl::size_type>(weightsSize.z) }, 0, 0, weights.data());
for (int wi = 0; wi < weights.size(); wi++)
os << weights[wi].x << " " << weights[wi].y << " ";
}
else {
std::vector<cl_float> weights(totalNumWeights);
//cs.getQueue().enqueueReadImage(vl._weights[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(weightsSize.x), static_cast<cl::size_type>(weightsSize.y), static_cast<cl::size_type>(weightsSize.z) }, 0, 0, weights.data());
for (int wi = 0; wi < weights.size(); wi++)
os << weights[wi] << " ";
}
os << std::endl;
os << vl._hiddenToVisible.x << " " << vl._hiddenToVisible.y << " " << vl._visibleToHidden.x << " " << vl._visibleToHidden.y << " " << vl._reverseRadii.x << " " << vl._reverseRadii.y << std::endl;
}
}
void ComparisonSparseCoder::learn(sys::ComputeSystem &cs, const cl::Image2D &rewards, std::vector<cl::Image2D> &visibleStates, float boostAlpha, float activeRatio) {
// Learn biases
{
int argIndex = 0;
_learnHiddenBiasesKernel.setArg(argIndex++, _hiddenBiases[_back]);
_learnHiddenBiasesKernel.setArg(argIndex++, _hiddenBiases[_front]);
_learnHiddenBiasesKernel.setArg(argIndex++, _hiddenStates[_back]);
_learnHiddenBiasesKernel.setArg(argIndex++, boostAlpha);
_learnHiddenBiasesKernel.setArg(argIndex++, activeRatio);
cs.getQueue().enqueueNDRangeKernel(_learnHiddenBiasesKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
std::swap(_hiddenBiases[_front], _hiddenBiases[_back]);
}
// Learn weights
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
VisibleLayer &vl = _visibleLayers[vli];
VisibleLayerDesc &vld = _visibleLayerDescs[vli];
if (!vld._isPredictiveCoding) {
if (vld._useTraces) {
int argIndex = 0;
_learnHiddenWeightsTracesActivationKernel.setArg(argIndex++, rewards);
_learnHiddenWeightsTracesActivationKernel.setArg(argIndex++, visibleStates[vli]);
_learnHiddenWeightsTracesActivationKernel.setArg(argIndex++, _hiddenStates[_back]);
_learnHiddenWeightsTracesActivationKernel.setArg(argIndex++, _hiddenActivationSummationTemp[_back]);
_learnHiddenWeightsTracesActivationKernel.setArg(argIndex++, vl._weights[_back]);
_learnHiddenWeightsTracesActivationKernel.setArg(argIndex++, vl._weights[_front]);
_learnHiddenWeightsTracesActivationKernel.setArg(argIndex++, vld._size);
_learnHiddenWeightsTracesActivationKernel.setArg(argIndex++, vl._hiddenToVisible);
_learnHiddenWeightsTracesActivationKernel.setArg(argIndex++, vld._radius);
_learnHiddenWeightsTracesActivationKernel.setArg(argIndex++, vld._weightAlpha);
_learnHiddenWeightsTracesActivationKernel.setArg(argIndex++, vld._weightLambda);
cs.getQueue().enqueueNDRangeKernel(_learnHiddenWeightsTracesActivationKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
}
else {
int argIndex = 0;
_learnHiddenWeightsActivationKernel.setArg(argIndex++, visibleStates[vli]);
_learnHiddenWeightsActivationKernel.setArg(argIndex++, _hiddenStates[_back]);
_learnHiddenWeightsActivationKernel.setArg(argIndex++, _hiddenActivationSummationTemp[_back]);
_learnHiddenWeightsActivationKernel.setArg(argIndex++, vl._weights[_back]);
_learnHiddenWeightsActivationKernel.setArg(argIndex++, vl._weights[_front]);
_learnHiddenWeightsActivationKernel.setArg(argIndex++, vld._size);
_learnHiddenWeightsActivationKernel.setArg(argIndex++, vl._hiddenToVisible);
_learnHiddenWeightsActivationKernel.setArg(argIndex++, vld._radius);
_learnHiddenWeightsActivationKernel.setArg(argIndex++, vld._weightAlpha);
cs.getQueue().enqueueNDRangeKernel(_learnHiddenWeightsActivationKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
}
std::swap(vl._weights[_front], vl._weights[_back]);
}
}
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
VisibleLayer &vl = _visibleLayers[vli];
VisibleLayerDesc &vld = _visibleLayerDescs[vli];
if (vld._isPredictiveCoding) {
if (vld._useTraces) {
int argIndex = 0;
_learnHiddenWeightsTracesPredictionKernel.setArg(argIndex++, rewards);
_learnHiddenWeightsTracesPredictionKernel.setArg(argIndex++, visibleStates[vli]);
_learnHiddenWeightsTracesPredictionKernel.setArg(argIndex++, _hiddenStates[_back]);
_learnHiddenWeightsTracesPredictionKernel.setArg(argIndex++, _hiddenPredictionSummationTemp[_back]);
_learnHiddenWeightsTracesPredictionKernel.setArg(argIndex++, vl._weights[_back]);
_learnHiddenWeightsTracesPredictionKernel.setArg(argIndex++, vl._weights[_front]);
_learnHiddenWeightsTracesPredictionKernel.setArg(argIndex++, vld._size);
_learnHiddenWeightsTracesPredictionKernel.setArg(argIndex++, vl._hiddenToVisible);
_learnHiddenWeightsTracesPredictionKernel.setArg(argIndex++, vld._radius);
_learnHiddenWeightsTracesPredictionKernel.setArg(argIndex++, vld._weightAlpha);
_learnHiddenWeightsTracesPredictionKernel.setArg(argIndex++, vld._weightLambda);
cs.getQueue().enqueueNDRangeKernel(_learnHiddenWeightsTracesPredictionKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
}
else {
int argIndex = 0;
_learnHiddenWeightsPredictionKernel.setArg(argIndex++, visibleStates[vli]);
_learnHiddenWeightsPredictionKernel.setArg(argIndex++, _hiddenStates[_back]);
_learnHiddenWeightsPredictionKernel.setArg(argIndex++, _hiddenPredictionSummationTemp[_back]);
_learnHiddenWeightsPredictionKernel.setArg(argIndex++, vl._weights[_back]);
_learnHiddenWeightsPredictionKernel.setArg(argIndex++, vl._weights[_front]);
_learnHiddenWeightsPredictionKernel.setArg(argIndex++, vld._size);
_learnHiddenWeightsPredictionKernel.setArg(argIndex++, vl._hiddenToVisible);
_learnHiddenWeightsPredictionKernel.setArg(argIndex++, vld._radius);
_learnHiddenWeightsPredictionKernel.setArg(argIndex++, vld._weightAlpha);
cs.getQueue().enqueueNDRangeKernel(_learnHiddenWeightsPredictionKernel, cl::NullRange, cl::NDRange(_hiddenSize.x, _hiddenSize.y));
}
std::swap(vl._weights[_front], vl._weights[_back]);
}
}
}
void AgentER::createRandom(sys::ComputeSystem &cs, sys::ComputeProgram &program,
cl_int2 inputSize, cl_int2 actionSize, cl_int2 qSize,
const std::vector<LayerDesc> &layerDescs,
cl_float2 initWeightRange,
std::mt19937 &rng)
{
_inputSize = inputSize;
_actionSize = actionSize;
_qSize = qSize;
_layerDescs = layerDescs;
_layers.resize(_layerDescs.size());
cl::Kernel randomUniform2DXYKernel = cl::Kernel(program.getProgram(), "randomUniform2DXY");
cl_int2 prevLayerSize = inputSize;
for (int l = 0; l < _layers.size(); l++) {
std::vector<ComparisonSparseCoder::VisibleLayerDesc> scDescs;
if (l == 0) {
scDescs.resize(3);
scDescs[0]._size = prevLayerSize;
scDescs[0]._radius = _layerDescs[l]._feedForwardRadius;
scDescs[0]._ignoreMiddle = false;
scDescs[0]._weightAlpha = _layerDescs[l]._scWeightAlpha;
scDescs[0]._useTraces = false;
scDescs[1]._size = _actionSize;
scDescs[1]._radius = _layerDescs[l]._feedForwardRadius;
scDescs[1]._ignoreMiddle = false;
scDescs[1]._weightAlpha = _layerDescs[l]._scWeightAlpha;
scDescs[1]._useTraces = false;
scDescs[2]._size = _qSize;
scDescs[2]._radius = _layerDescs[l]._feedForwardRadius;
scDescs[2]._ignoreMiddle = false;
scDescs[2]._weightAlpha = _layerDescs[l]._scWeightAlpha;
scDescs[2]._useTraces = false;
}
else {
scDescs.resize(2);
scDescs[0]._size = prevLayerSize;
scDescs[0]._radius = _layerDescs[l]._feedForwardRadius;
scDescs[0]._ignoreMiddle = false;
scDescs[0]._weightAlpha = _layerDescs[l]._scWeightAlpha;
scDescs[0]._useTraces = false;
scDescs[1]._size = _layerDescs[l]._size;
scDescs[1]._radius = _layerDescs[l]._recurrentRadius;
scDescs[1]._ignoreMiddle = true;
scDescs[1]._weightAlpha = _layerDescs[l]._scWeightRecurrentAlpha;
scDescs[1]._useTraces = false;
}
_layers[l]._sc.createRandom(cs, program, scDescs, _layerDescs[l]._size, _layerDescs[l]._lateralRadius, initWeightRange, rng);
std::vector<Predictor::VisibleLayerDesc> predDescs;
if (l < _layers.size() - 1) {
predDescs.resize(2);
predDescs[0]._size = _layerDescs[l]._size;
predDescs[0]._radius = _layerDescs[l]._predictiveRadius;
predDescs[1]._size = _layerDescs[l + 1]._size;
predDescs[1]._radius = _layerDescs[l]._feedBackRadius;
}
else {
predDescs.resize(1);
predDescs[0]._size = _layerDescs[l]._size;
predDescs[0]._radius = _layerDescs[l]._predictiveRadius;
}
if (l == 0)
_layers[l]._pred.createRandom(cs, program, predDescs, _actionSize, initWeightRange, true, rng);
else
_layers[l]._pred.createRandom(cs, program, predDescs, _layerDescs[l - 1]._size, initWeightRange, true, rng);
// Create baselines
_layers[l]._predReward = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._size.x, _layerDescs[l]._size.y);
_layers[l]._propagatedPredReward = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _layerDescs[l]._size.x, _layerDescs[l]._size.y);
cl_float4 zeroColor = { 0.0f, 0.0f, 0.0f, 0.0f };
cl::array<cl::size_type, 3> zeroOrigin = { 0, 0, 0 };
cl::array<cl::size_type, 3> layerRegion = { _layerDescs[l]._size.x, _layerDescs[l]._size.y, 1 };
cs.getQueue().enqueueFillImage(_layers[l]._predReward, zeroColor, zeroOrigin, layerRegion);
cs.getQueue().enqueueFillImage(_layers[l]._propagatedPredReward, zeroColor, zeroOrigin, layerRegion);
_layers[l]._scStatesTemp = createDoubleBuffer2D(cs, _layerDescs[l]._size, CL_R, CL_FLOAT);
_layers[l]._predStatesTemp = createDoubleBuffer2D(cs, prevLayerSize, CL_R, CL_FLOAT);
prevLayerSize = _layerDescs[l]._size;
}
_qInput = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _qSize.x, _qSize.y);
_qTarget = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _qSize.x, _qSize.y);
_actionTarget = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), _actionSize.x, _actionSize.y);
_qTransform = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_RG, CL_FLOAT), _qSize.x, _qSize.y);
// Q Predictor
{
std::vector<Predictor::VisibleLayerDesc> predDescs;
if (0 < _layers.size() - 1) {
predDescs.resize(2);
predDescs[0]._size = _layerDescs[0]._size;
predDescs[0]._radius = _layerDescs[0]._predictiveRadius;
predDescs[1]._size = _layerDescs[0 + 1]._size;
predDescs[1]._radius = _layerDescs[0]._feedBackRadius;
}
else {
predDescs.resize(1);
predDescs[0]._size = _layerDescs[0]._size;
predDescs[0]._radius = _layerDescs[0]._predictiveRadius;
}
_qPred.createRandom(cs, program, predDescs, _qSize, initWeightRange, true, rng);
}
// Random Q transform
randomUniformXY(_qTransform, cs, randomUniform2DXYKernel, _qSize, { -1.0f, 1.0f }, rng);
_inputWhitener.create(cs, program, _inputSize, CL_R, CL_FLOAT);
_actionWhitener.create(cs, program, _actionSize, CL_R, CL_FLOAT);
_qWhitener.create(cs, program, _qSize, CL_R, CL_FLOAT);
_predictionRewardKernel = cl::Kernel(program.getProgram(), "phPredictionReward");
_predictionRewardPropagationKernel = cl::Kernel(program.getProgram(), "phPredictionRewardPropagation");
_setQKernel = cl::Kernel(program.getProgram(), "phSetQ");
}
void PredictorSwarm::createRandom(sys::ComputeSystem &cs, sys::ComputeProgram &program,
const std::vector<VisibleLayerDesc> &visibleLayerDescs, cl_int2 hiddenSize, cl_float2 initWeightRange,
std::mt19937 &rng)
{
_visibleLayerDescs = visibleLayerDescs;
_hiddenSize = hiddenSize;
_visibleLayers.resize(_visibleLayerDescs.size());
cl::Kernel randomUniform2DKernel = cl::Kernel(program.getProgram(), "randomUniform2D");
cl::Kernel randomUniform3DKernel = cl::Kernel(program.getProgram(), "randomUniform3D");
cl::Kernel randomUniform3DXZKernel = cl::Kernel(program.getProgram(), "randomUniform3DXZ");
cl_float4 zeroColor = { 0.0f, 0.0f, 0.0f, 0.0f };
cl::array<cl::size_type, 3> zeroOrigin = { 0, 0, 0 };
cl::array<cl::size_type, 3> hiddenRegion = { _hiddenSize.x, _hiddenSize.y, 1 };
// Create layers
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
VisibleLayer &vl = _visibleLayers[vli];
VisibleLayerDesc &vld = _visibleLayerDescs[vli];
vl._hiddenToVisible = cl_float2{ static_cast<float>(vld._size.x) / static_cast<float>(_hiddenSize.x),
static_cast<float>(vld._size.y) / static_cast<float>(_hiddenSize.y)
};
vl._visibleToHidden = cl_float2{ static_cast<float>(_hiddenSize.x) / static_cast<float>(vld._size.x),
static_cast<float>(_hiddenSize.y) / static_cast<float>(vld._size.y)
};
vl._reverseRadii = cl_int2{ static_cast<int>(std::ceil(vl._visibleToHidden.x * (vld._radius + 0.5f))), static_cast<int>(std::ceil(vl._visibleToHidden.y * (vld._radius + 0.5f))) };
int weightDiam = vld._radius * 2 + 1;
int numWeights = weightDiam * weightDiam;
cl_int3 weightsSize = { _hiddenSize.x, _hiddenSize.y, numWeights };
vl._weights = createDoubleBuffer3D(cs, weightsSize, CL_RGBA, CL_FLOAT);
randomUniformXZ(vl._weights[_back], cs, randomUniform3DXZKernel, weightsSize, initWeightRange, rng);
vl._qTraces = createDoubleBuffer3D(cs, weightsSize, CL_R, CL_FLOAT);
cs.getQueue().enqueueFillImage(vl._qTraces[_back], zeroColor, zeroOrigin, { static_cast<cl::size_type>(weightsSize.x), static_cast<cl::size_type>(weightsSize.y), static_cast<cl::size_type>(weightsSize.z) });
}
// Hidden state data
_hiddenStates = createDoubleBuffer2D(cs, _hiddenSize, CL_RG, CL_FLOAT);
_hiddenActivations = createDoubleBuffer2D(cs, _hiddenSize, CL_RG, CL_FLOAT);
_hiddenSummationTemp = createDoubleBuffer2D(cs, _hiddenSize, CL_RG, CL_FLOAT);
cs.getQueue().enqueueFillImage(_hiddenStates[_back], zeroColor, zeroOrigin, hiddenRegion);
cs.getQueue().enqueueFillImage(_hiddenActivations[_back], zeroColor, zeroOrigin, hiddenRegion);
// Create kernels
_activateKernel = cl::Kernel(program.getProgram(), "predActivateSwarm");
_solveHiddenKernel = cl::Kernel(program.getProgram(), "predSolveHiddenSwarm");
_solveHiddenNoInhibitionKernel = cl::Kernel(program.getProgram(), "predSolveHiddenNoInhibitionSwarm");
_learnWeightsTracesInhibitedKernel = cl::Kernel(program.getProgram(), "predLearnWeightsTracesSwarm");
_reconstructionErrorKernel = cl::Kernel(program.getProgram(), "predReconstructionErrorSwarm");
}
void Predictor::readFromStream(sys::ComputeSystem &cs, sys::ComputeProgram &program, std::istream &is) {
abort(); // Not yet working
is >> _hiddenSize.x >> _hiddenSize.y;
_hiddenStates = createDoubleBuffer2D(cs, _hiddenSize, CL_R, CL_FLOAT);
_hiddenSummationTemp = createDoubleBuffer2D(cs, _hiddenSize, CL_R, CL_FLOAT);
{
std::vector<cl_float> hiddenStates(_hiddenSize.x * _hiddenSize.y);
for (int si = 0; si < hiddenStates.size(); si++)
is >> hiddenStates[si];
cs.getQueue().enqueueWriteImage(_hiddenStates[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_hiddenSize.x), static_cast<cl::size_type>(_hiddenSize.y), 1 }, 0, 0, hiddenStates.data());
}
// Layer information
int numLayers;
is >> numLayers;
_visibleLayerDescs.resize(numLayers);
_visibleLayers.resize(numLayers);
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
VisibleLayer &vl = _visibleLayers[vli];
VisibleLayerDesc &vld = _visibleLayerDescs[vli];
// Desc
is >> vld._size.x >> vld._size.y >> vld._radius;
// Layer
int weightDiam = vld._radius * 2 + 1;
int numWeights = weightDiam * weightDiam;
cl_int3 weightsSize = cl_int3{ _hiddenSize.x, _hiddenSize.y, numWeights };
int totalNumWeights = weightsSize.x * weightsSize.y * weightsSize.z;
{
vl._weights = createDoubleBuffer3D(cs, weightsSize, CL_R, CL_FLOAT);
std::vector<cl_float> weights(totalNumWeights);
for (int wi = 0; wi < weights.size(); wi++)
is >> weights[wi];
cs.getQueue().enqueueWriteImage(vl._weights[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(weightsSize.x), static_cast<cl::size_type>(weightsSize.y), static_cast<cl::size_type>(weightsSize.z) }, 0, 0, weights.data());
}
is >> vl._hiddenToVisible.x >> vl._hiddenToVisible.y >> vl._visibleToHidden.x >> vl._visibleToHidden.y >> vl._reverseRadii.x >> vl._reverseRadii.y;
}
// Create kernels
_activateKernel = cl::Kernel(program.getProgram(), "predActivate");
//_solveHiddenKernel = cl::Kernel(program.getProgram(), "predSolveHidden");
_learnWeightsKernel = cl::Kernel(program.getProgram(), "predLearnWeights");
}
void HTFE::learn(sys::ComputeSystem &cs) {
// ------------------------------------------------------------------------------
// ---------------------- Weight Update and Predictions ------------------------
// ------------------------------------------------------------------------------
cl::Image2D* pPrevLayer = &_inputImage;
int prevWidth = _inputWidth;
int prevHeight = _inputHeight;
cl::Image2D* pPrevLayerFeedForwardPrev = &_inputImagePrev;
cl::Image2D* pPrevLayerFeedBackPrev = &_inputImagePrev;
for (int l = 0; l < _layers.size(); l++) {
float localActivity = std::round(_layerDescs[l]._sparsity * std::pow(2 * _layerDescs[l]._inhibitionRadius + 1, 2));
Int2 layerSize;
layerSize._x = _layerDescs[l]._width;
layerSize._y = _layerDescs[l]._height;
Int2 layerSizeMinusOne;
layerSizeMinusOne._x = _layerDescs[l]._width - 1;
layerSizeMinusOne._y = _layerDescs[l]._height - 1;
Float2 layerSizeMinusOneInv;
layerSizeMinusOneInv._x = 1.0f / (_layerDescs[l]._width - 1);
layerSizeMinusOneInv._y = 1.0f / (_layerDescs[l]._height - 1);
Int2 inputSize;
inputSize._x = prevWidth;
inputSize._y = prevHeight;
Int2 inputSizeMinusOne;
inputSizeMinusOne._x = prevWidth - 1;
inputSizeMinusOne._y = prevHeight - 1;
Float2 inputSizeMinusOneInv;
inputSizeMinusOneInv._x = 1.0f / (prevWidth - 1);
inputSizeMinusOneInv._y = 1.0f / (prevHeight - 1);
Int2 nextSize;
Int2 nextSizeMinusOne;
if (l == _layers.size() - 1) {
nextSize._x = nextSize._y = 1;
nextSizeMinusOne._x = nextSizeMinusOne._y = 0;
}
else {
nextSize._x = _layerDescs[l + 1]._width;
nextSize._y = _layerDescs[l + 1]._height;
nextSizeMinusOne._x = _layerDescs[l + 1]._width - 1;
nextSizeMinusOne._y = _layerDescs[l + 1]._height - 1;
}
// ------------------------------- Weight Updates -------------------------------
Float4 alphas;
alphas._x = _layerDescs[l]._feedForwardAlpha;
alphas._y = _layerDescs[l]._lateralAlpha;
alphas._z = _layerDescs[l]._feedBackAlpha;
alphas._w = _layerDescs[l]._hiddenBiasAlpha;
int index = 0;
if (l == _layers.size() - 1) {
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layers[l]._visibleReconstructionPrev);
_layerHiddenWeightUpdateLastKernel.setArg(index++, *pPrevLayer);
_layerHiddenWeightUpdateLastKernel.setArg(index++, *pPrevLayerFeedForwardPrev);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layers[l]._hiddenFeedBackActivationsPrev);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layers[l]._hiddenStatesFeedBackPrev);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layers[l]._hiddenStatesFeedBackPrevPrev);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layers[l]._reconstructionWeightsPrev);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layers[l]._feedForwardWeightsPrev);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layers[l]._lateralWeightsPrev);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layers[l]._hiddenBiasesPrev);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layers[l]._feedForwardWeights);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layers[l]._lateralWeights);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layers[l]._hiddenBiases);
_layerHiddenWeightUpdateLastKernel.setArg(index++, layerSize);
_layerHiddenWeightUpdateLastKernel.setArg(index++, layerSizeMinusOne);
_layerHiddenWeightUpdateLastKernel.setArg(index++, layerSizeMinusOneInv);
_layerHiddenWeightUpdateLastKernel.setArg(index++, inputSize);
_layerHiddenWeightUpdateLastKernel.setArg(index++, inputSizeMinusOne);
_layerHiddenWeightUpdateLastKernel.setArg(index++, inputSizeMinusOneInv);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layerDescs[l]._receptiveFieldRadius);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layerDescs[l]._lateralConnectionRadius);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layerDescs[l]._reconstructionRadius);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layerDescs[l]._sparsity);
_layerHiddenWeightUpdateLastKernel.setArg(index++, alphas);
_layerHiddenWeightUpdateLastKernel.setArg(index++, _layerDescs[l]._weightDecay);
cs.getQueue().enqueueNDRangeKernel(_layerHiddenWeightUpdateLastKernel, cl::NullRange, cl::NDRange(_layerDescs[l]._width, _layerDescs[l]._height));
}
else {
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l]._visibleReconstructionPrev);
_layerHiddenWeightUpdateKernel.setArg(index++, *pPrevLayer);
_layerHiddenWeightUpdateKernel.setArg(index++, *pPrevLayerFeedForwardPrev);
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l]._hiddenFeedBackActivationsPrev);
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l]._hiddenStatesFeedBackPrev);
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l]._hiddenStatesFeedBackPrevPrev);
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l + 1]._hiddenStatesFeedBackPrev);
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l]._reconstructionWeightsPrev);
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l]._feedForwardWeightsPrev);
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l]._lateralWeightsPrev);
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l]._hiddenBiasesPrev);
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l]._feedBackWeightsPrev);
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l]._feedForwardWeights);
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l]._lateralWeights);
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l]._hiddenBiases);
_layerHiddenWeightUpdateKernel.setArg(index++, _layers[l]._feedBackWeights);
_layerHiddenWeightUpdateKernel.setArg(index++, layerSize);
_layerHiddenWeightUpdateKernel.setArg(index++, layerSizeMinusOne);
_layerHiddenWeightUpdateKernel.setArg(index++, layerSizeMinusOneInv);
_layerHiddenWeightUpdateKernel.setArg(index++, inputSize);
_layerHiddenWeightUpdateKernel.setArg(index++, inputSizeMinusOne);
_layerHiddenWeightUpdateKernel.setArg(index++, inputSizeMinusOneInv);
_layerHiddenWeightUpdateKernel.setArg(index++, nextSize);
_layerHiddenWeightUpdateKernel.setArg(index++, nextSizeMinusOne);
_layerHiddenWeightUpdateKernel.setArg(index++, _layerDescs[l]._receptiveFieldRadius);
_layerHiddenWeightUpdateKernel.setArg(index++, _layerDescs[l]._lateralConnectionRadius);
_layerHiddenWeightUpdateKernel.setArg(index++, _layerDescs[l]._feedBackConnectionRadius);
_layerHiddenWeightUpdateKernel.setArg(index++, _layerDescs[l]._reconstructionRadius);
_layerHiddenWeightUpdateKernel.setArg(index++, _layerDescs[l]._sparsity);
_layerHiddenWeightUpdateKernel.setArg(index++, alphas);
_layerHiddenWeightUpdateKernel.setArg(index++, _layerDescs[l]._weightDecay);
cs.getQueue().enqueueNDRangeKernel(_layerHiddenWeightUpdateKernel, cl::NullRange, cl::NDRange(_layerDescs[l]._width, _layerDescs[l]._height));
}
index = 0;
_layerVisibleWeightUpdateKernel.setArg(index++, _layers[l]._visibleReconstructionPrev);
_layerVisibleWeightUpdateKernel.setArg(index++, *pPrevLayer);
_layerVisibleWeightUpdateKernel.setArg(index++, _layers[l]._hiddenStatesFeedBackPrev);
_layerVisibleWeightUpdateKernel.setArg(index++, _layers[l]._reconstructionWeightsPrev);
_layerVisibleWeightUpdateKernel.setArg(index++, _layers[l]._visibleBiasesPrev);
_layerVisibleWeightUpdateKernel.setArg(index++, _layers[l]._reconstructionWeights);
_layerVisibleWeightUpdateKernel.setArg(index++, _layers[l]._visibleBiases);
_layerVisibleWeightUpdateKernel.setArg(index++, _layerDescs[l]._reconstructionRadius);
_layerVisibleWeightUpdateKernel.setArg(index++, inputSizeMinusOne);
_layerVisibleWeightUpdateKernel.setArg(index++, inputSizeMinusOneInv);
_layerVisibleWeightUpdateKernel.setArg(index++, layerSize);
_layerVisibleWeightUpdateKernel.setArg(index++, layerSizeMinusOne);
_layerVisibleWeightUpdateKernel.setArg(index++, layerSizeMinusOneInv);
_layerVisibleWeightUpdateKernel.setArg(index++, _layerDescs[l]._reconstructionAlpha);
cs.getQueue().enqueueNDRangeKernel(_layerVisibleWeightUpdateKernel, cl::NullRange, cl::NDRange(prevWidth, prevHeight));
pPrevLayer = &_layers[l]._hiddenStatesFeedForward; // Or _hiddenStatesFeedBack ?
prevWidth = _layerDescs[l]._width;
prevHeight = _layerDescs[l]._height;
pPrevLayerFeedForwardPrev = &_layers[l]._hiddenStatesFeedForwardPrev;
pPrevLayerFeedBackPrev = &_layers[l]._hiddenStatesFeedBackPrev;
}
}
void ComparisonSparseCoder::readFromStream(sys::ComputeSystem &cs, sys::ComputeProgram &program, std::istream &is) {
abort(); // Fix me
is >> _hiddenSize.x >> _hiddenSize.y >> _lateralRadius;
_hiddenStates = createDoubleBuffer2D(cs, _hiddenSize, CL_R, CL_FLOAT);
_hiddenBiases = createDoubleBuffer2D(cs, _hiddenSize, CL_R, CL_FLOAT);
_hiddenActivationSummationTemp = createDoubleBuffer2D(cs, _hiddenSize, CL_R, CL_FLOAT);
//_hiddenReconstructionSummationTemp = createDoubleBuffer2D(cs, _hiddenSize, CL_R, CL_FLOAT);
{
std::vector<cl_float> hiddenStates(_hiddenSize.x * _hiddenSize.y);
for (int si = 0; si < hiddenStates.size(); si++)
is >> hiddenStates[si];
cs.getQueue().enqueueWriteImage(_hiddenStates[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_hiddenSize.x), static_cast<cl::size_type>(_hiddenSize.y), 1 }, 0, 0, hiddenStates.data());
}
{
std::vector<cl_float> hiddenBiases(_hiddenSize.x * _hiddenSize.y);
for (int bi = 0; bi < hiddenBiases.size(); bi++)
is >> hiddenBiases[bi];
cs.getQueue().enqueueWriteImage(_hiddenBiases[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(_hiddenSize.x), static_cast<cl::size_type>(_hiddenSize.y), 1 }, 0, 0, hiddenBiases.data());
}
// Layer information
int numLayers;
is >> numLayers;
_visibleLayerDescs.resize(numLayers);
_visibleLayers.resize(numLayers);
for (int vli = 0; vli < _visibleLayers.size(); vli++) {
VisibleLayer &vl = _visibleLayers[vli];
VisibleLayerDesc &vld = _visibleLayerDescs[vli];
// Desc
is >> vld._size.x >> vld._size.y >> vld._radius >> vld._weightAlpha >> vld._weightLambda >> vld._ignoreMiddle >> vld._useTraces;
// Layer
//vl._reconstructionError = cl::Image2D(cs.getContext(), CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), vld._size.x, vld._size.y);
int weightDiam = vld._radius * 2 + 1;
int numWeights = weightDiam * weightDiam;
cl_int3 weightsSize = cl_int3{ _hiddenSize.x, _hiddenSize.y, numWeights };
int totalNumWeights = weightsSize.x * weightsSize.y * weightsSize.z;
if (vld._useTraces) {
//vl._weights = createDoubleBuffer3D(cs, weightsSize, CL_RG, CL_FLOAT);
std::vector<cl_float2> weights(totalNumWeights);
for (int wi = 0; wi < weights.size(); wi++)
is >> weights[wi].x >> weights[wi].y;
//cs.getQueue().enqueueWriteImage(vl._weights[_back], CL_TRUE, { 0, 0, 0 }, { static_cast<cl::size_type>(weightsSize.x), static_cast<cl::size_type>(weightsSize.y), static_cast<cl::size_type>(weightsSize.z) }, 0, 0, weights.data());
}
else {
void AgentSwarm::simStep(sys::ComputeSystem &cs, float reward, const cl::Image2D &input, std::mt19937 &rng) {
// Feed forward
cl_int2 prevLayerSize = _layers.front()._sc.getVisibleLayerDesc(0)._size;
cl::Image2D prevLayerState = input;
for (int l = 0; l < _layers.size(); l++) {
{
std::vector<cl::Image2D> visibleStates(2);
// Modulate
{
int argIndex = 0;
_modulateKernel.setArg(argIndex++, prevLayerState);
_modulateKernel.setArg(argIndex++, _layers[l]._swarm.getVisibleLayer(0)._actionsExploratory);
_modulateKernel.setArg(argIndex++, _layers[l]._modulatedFeedForwardInput);
_modulateKernel.setArg(argIndex++, _layerDescs[l]._minAttention);
cs.getQueue().enqueueNDRangeKernel(_modulateKernel, cl::NullRange, cl::NDRange(prevLayerSize.x, prevLayerSize.y));
}
// Modulate
{
int argIndex = 0;
_modulateKernel.setArg(argIndex++, _layers[l]._scHiddenStatesPrev);
_modulateKernel.setArg(argIndex++, _layers[l]._swarm.getVisibleLayer(1)._actionsExploratory);
_modulateKernel.setArg(argIndex++, _layers[l]._modulatedRecurrentInput);
_modulateKernel.setArg(argIndex++, _layerDescs[l]._minAttention);
cs.getQueue().enqueueNDRangeKernel(_modulateKernel, cl::NullRange, cl::NDRange(_layerDescs[l]._hiddenSize.x, _layerDescs[l]._hiddenSize.y));
}
visibleStates[0] = _layers[l]._modulatedFeedForwardInput;
visibleStates[1] = _layers[l]._modulatedRecurrentInput;
//_layers[l]._sc.activate(cs, visibleStates, _layerDescs[l]._scActiveRatio);
_layers[l]._sc.learn(cs, _layers[l]._reward, visibleStates, _layerDescs[l]._scBoostAlpha, _layerDescs[l]._scActiveRatio);
}
// Get reward
/*if (l == 0) {
int argIndex = 0;
_baseLineUpdateKernel.setArg(argIndex++, _layers[l]._pred.getVisibleLayer(0)._errors);
_baseLineUpdateKernel.setArg(argIndex++, _layers[l]._sc.getHiddenStates()[_back]);
_baseLineUpdateKernel.setArg(argIndex++, _layers[l]._baseLines[_back]);
_baseLineUpdateKernel.setArg(argIndex++, _layers[l]._baseLines[_front]);
_baseLineUpdateKernel.setArg(argIndex++, _layers[l]._reward);
_baseLineUpdateKernel.setArg(argIndex++, _layerDescs[l]._baseLineDecay);
_baseLineUpdateKernel.setArg(argIndex++, _layerDescs[l]._baseLineSensitivity);
cs.getQueue().enqueueNDRangeKernel(_baseLineUpdateKernel, cl::NullRange, cl::NDRange(_layerDescs[l]._hiddenSize.x, _layerDescs[l]._hiddenSize.y));
}
else {
int argIndex = 0;
_baseLineUpdateSumErrorKernel.setArg(argIndex++, _layers[l - 1]._pred.getVisibleLayer(1)._errors);
_baseLineUpdateSumErrorKernel.setArg(argIndex++, _layers[l]._pred.getVisibleLayer(0)._errors);
_baseLineUpdateSumErrorKernel.setArg(argIndex++, _layers[l]._sc.getHiddenStates()[_back]);
_baseLineUpdateSumErrorKernel.setArg(argIndex++, _layers[l]._baseLines[_back]);
_baseLineUpdateSumErrorKernel.setArg(argIndex++, _layers[l]._baseLines[_front]);
_baseLineUpdateSumErrorKernel.setArg(argIndex++, _layers[l]._reward);
_baseLineUpdateSumErrorKernel.setArg(argIndex++, _layerDescs[l]._baseLineDecay);
_baseLineUpdateSumErrorKernel.setArg(argIndex++, _layerDescs[l]._baseLineSensitivity);
cs.getQueue().enqueueNDRangeKernel(_baseLineUpdateSumErrorKernel, cl::NullRange, cl::NDRange(_layerDescs[l]._hiddenSize.x, _layerDescs[l]._hiddenSize.y));
}*/
prevLayerState = _layers[l]._sc.getHiddenStates()[_back];
prevLayerSize = _layerDescs[l]._hiddenSize;
}
for (int l = _layers.size() - 1; l >= 0; l--) {
std::vector<cl::Image2D> visibleStates;
if (l < _layers.size() - 1) {
visibleStates.resize(2);
visibleStates[0] = _layers[l]._sc.getHiddenStates()[_back];
visibleStates[1] = _layers[l + 1]._pred.getHiddenStates()[_back];
}
else {
visibleStates.resize(1);
visibleStates[0] = _layers[l]._sc.getHiddenStates()[_back];
}
//_layers[l]._pred.activate(cs, visibleStates);
//if (l == 0)
// _layers[l]._pred.propagateError(cs, input);
//else
// _layers[l]._pred.propagateError(cs, _layers[l - 1]._sc.getHiddenStates()[_back]);
}
for (int l = _layers.size() - 1; l >= 0; l--) {
std::vector<cl::Image2D> visibleStatesPrev;
if (l < _layers.size() - 1) {
visibleStatesPrev.resize(2);
visibleStatesPrev[0] = _layers[l]._scHiddenStatesPrev;
visibleStatesPrev[1] = _layers[l + 1]._pred.getHiddenStates()[_front];
}
else {
visibleStatesPrev.resize(1);
visibleStatesPrev[0] = _layers[l]._scHiddenStatesPrev;
}
if (l == 0)
_layers[l]._pred.learn(cs, input, visibleStatesPrev, _layerDescs[l]._predWeightAlpha);
else
_layers[l]._pred.learn(cs, _layers[l - 1]._sc.getHiddenStates()[_back], visibleStatesPrev, _layerDescs[l]._predWeightAlpha);
}
// Swarm
for (int l = _layers.size() - 1; l >= 0; l--) {
std::vector<cl::Image2D> visibleStatesPrev;
if (l < _layers.size() - 1) {
_layers[l]._swarm.simStep(cs, reward, _layers[l]._sc.getHiddenStates()[_back], _layers[l + 1]._inhibitedAction,
_layerDescs[l]._swarmExpPert, _layerDescs[l]._swarmExpBreak,
_layerDescs[l]._swarmAnnealingIterations, _layerDescs[l]._swarmActionDeriveAlpha,
_layerDescs[l]._swarmQHiddenAlpha, _layerDescs[l]._swarmQAlpha, _layerDescs[l]._swarmPredAlpha,
_layerDescs[l]._swarmLambda, _layerDescs[l]._swarmGamma, rng);
}
else {
_layers[l]._swarm.simStep(cs, reward, _layers[l]._sc.getHiddenStates()[_back], _lastLayerAction,
_layerDescs[l]._swarmExpPert, _layerDescs[l]._swarmExpBreak,
_layerDescs[l]._swarmAnnealingIterations, _layerDescs[l]._swarmActionDeriveAlpha,
_layerDescs[l]._swarmQHiddenAlpha, _layerDescs[l]._swarmQAlpha, _layerDescs[l]._swarmPredAlpha,
_layerDescs[l]._swarmLambda, _layerDescs[l]._swarmGamma, rng);
}
// If not first layer, inhibit the action
if (l != 0) {
int argIndex = 0;
_inhibitKernel.setArg(argIndex++, _layers[l]._swarm.getVisibleLayer(2)._actionsExploratory);
_inhibitKernel.setArg(argIndex++, _layers[l]._inhibitedAction);
_inhibitKernel.setArg(argIndex++, _layerDescs[l - 1]._hiddenSize);
_inhibitKernel.setArg(argIndex++, _layerDescs[l - 1]._lateralRadius);
_inhibitKernel.setArg(argIndex++, _layerDescs[l - 1]._scActiveRatio);
cs.getQueue().enqueueNDRangeKernel(_inhibitKernel, cl::NullRange, cl::NDRange(_layerDescs[l - 1]._hiddenSize.x, _layerDescs[l - 1]._hiddenSize.y));
}
}
// Buffer updates
for (int l = 0; l < _layers.size(); l++) {
cl::array<cl::size_type, 3> zeroOrigin = { 0, 0, 0 };
cl::array<cl::size_type, 3> layerRegion = { _layerDescs[l]._hiddenSize.x, _layerDescs[l]._hiddenSize.y, 1 };
cs.getQueue().enqueueCopyImage(_layers[l]._sc.getHiddenStates()[_back], _layers[l]._scHiddenStatesPrev, zeroOrigin, zeroOrigin, layerRegion);
std::swap(_layers[l]._baseLines[_front], _layers[l]._baseLines[_back]);
}
}