void ConvBaseOperator::reshapeImageDescriptors() {
hl_tensor_reshape(imageDesc_,
1,
channels_,
imageH_,
imageW_,
channels_ * imageH_ * imageW_,
imageH_ * imageW_,
imageW_,
1);
hl_tensor_reshape(outputDesc_,
1,
numFilters_,
outputH_,
outputW_,
numFilters_ * outputH_ * outputW_,
outputH_ * outputW_,
outputW_,
1);
hl_reset_convolution_descriptor(convDesc_,
imageDesc_,
filterDesc_,
paddingY_,
padding_,
strideY_,
stride_);
}
void CudnnConvLayer::forward(PassType passType) {
Layer::forward(passType);
int batchSize = getInput(0).getBatchSize();
resetOutput(batchSize, calOutputSize());
for (size_t i = 0; i != inputLayers_.size(); ++i) {
projections_[i]->forward(&getInput(i), &getOutput(), passType);
}
if (biases_) {
REGISTER_TIMER_INFO("CudnnConvBiasTimer", getName().c_str());
int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
hl_tensor_reshape(outputDesc_, batchSize, numFilters_ / groups_[0],
outputH_[0], outputW_[0], numFilters_ * outputH_[0] * outputW_[0],
outputH_[0] * outputW_[0], outputW_[0], 1);
outputOffset_ = getOutputValue()->getWidth() / groups_[0];
for (int g = 0; g < groups_[0]; ++g) {
real *biasData = biases_->getW()->getData() + biasOffset_ * g;
real *outData = getOutputValue()->getData() + outputOffset_ * g;
hl_convolution_forward_add_bias(biasDesc_, biasData,
outputDesc_, outData);
}
}
forwardActivation();
}
bool CudnnConvLayer::init(const LayerMap &layerMap,
const ParameterMap ¶meterMap) {
if (!ConvBaseLayer::init(layerMap, parameterMap)) return false;
CHECK(useGpu_) << "CudnnConvLayer only support gpu";
CHECK_EQ(inputLayers_.size(), parameters_.size());
projections_.reserve(inputLayers_.size());
projConf_.reserve(inputLayers_.size());
numFilters_ = config_.num_filters();
CHECK(config_.shared_biases());
for (size_t i = 0; i < inputLayers_.size(); i++) {
ProjectionConfig* conf = new ProjectionConfig();
conf->set_type("conv");
conf->set_num_filters(numFilters_);
ConvConfig* convConf = conf->mutable_conv_conf();
*convConf = *(config_.mutable_inputs(i)->mutable_conv_conf());
conf->set_input_size(getPrev(i)->getSize());
conf->set_output_size(getSize());
projConf_.emplace_back(conf);
projections_.emplace_back(Projection::create(*projConf_[i],
parameters_[i], useGpu_));
}
if (biases_.get() && sharedBiases_) {
hl_create_tensor_descriptor(&biasDesc_);
hl_create_tensor_descriptor(&outputDesc_);
hl_tensor_reshape(biasDesc_, 1, numFilters_ / groups_[0], 1, 1);
biasOffset_ = numFilters_ / groups_[0];
}
return true;
}
void ConvBaseProjection::reshapeTensorDesc(int batchSize) {
// The stride between two consecutive samples in the output of ConvProjection
// may not be numFilters_ * outputH_ * outputW_ (conv) or
// channels_ * imageH_ * imageW_ (deconv)
// for example, in the case of layer ConcatenateLayer2 with two
// ConvProjection, the stride is the output_size of layer ConcatenateLayer2.
// So the calculation of nStride is different from CudnnConvLayer.
size_t nStrideImage, nStrideOutput;
if (isDeconv_) {
nStrideImage = out_->value->getStride();
nStrideOutput = numFilters_ * outputH_ * outputW_;
} else {
nStrideImage = channels_ * imageH_ * imageW_;
nStrideOutput = out_->value->getStride();
}
hl_tensor_reshape(imageDesc_,
batchSize,
channels_ / groups_,
imageH_,
imageW_,
nStrideImage,
imageH_ * imageW_,
imageW_,
1);
hl_tensor_reshape(outputDesc_,
batchSize,
numFilters_ / groups_,
outputH_,
outputW_,
nStrideOutput,
outputH_ * outputW_,
outputW_,
1);
hl_reset_convolution_descriptor(convDesc_,
imageDesc_,
filterDesc_,
paddingH_,
paddingW_,
strideH_,
strideW_,
dilationH_,
dilationW_);
}
