THROW_IE_EXCEPTION << "Cannot cast layer to TensorIterator.";
}
- std::map<uint64_t, std::vector<std::pair<std::string, uint64_t>>> ngraph_parameter_id_to_ie_layer_port;
- std::map<std::pair<std::string, uint64_t>, uint64_t> ie_layer_port_to_tensor_iterator_input_id;
-
- // inputs/outputs of TensorIterator body (ie)
- std::map<std::string, DataPtr> in_info_map;
- std::map<std::string, DataPtr> out_info_map;
-
// inputs/outputs of TensorIterator (ngraph representation)
auto parameters = tensor_iterator->get_function()->get_parameters();
auto results = tensor_iterator->get_function()->get_results();
// IE TensorIterator doesn't include such nodes so we create CNNNetwork in a separate scope
// to call the destructor and delete these "Input"/data nodes.
- // These layers will hold the necessary subnet after destruction of CNNNetwork.
- std::set<InferenceEngine::CNNLayerPtr> body_input_layers;
- // This map will save information about data nodes
- std::map<std::string, std::vector<TensorDesc>> layer_name_to_tensor_desc;
+ TensorIterator::Body body;
{
CNNNetwork body_net(tensor_iterator->get_function());
CNNNetwork net(InferenceEngine::details::convertFunctionToICNNNetwork(body_net.getFunction(), body_net));
}
// Get inputs/outputs of cnn network
- InputsDataMap in_info_map_with_parameters;
- in_info_map_with_parameters = net.getInputsInfo();
- out_info_map = net.getOutputsInfo();
+ auto in_info_map_with_parameters = net.getInputsInfo();
+ auto out_info_map = net.getOutputsInfo();
+
+ IE_ASSERT(in_info_map_with_parameters.size() == parameters.size());
+ IE_ASSERT(out_info_map.size() == results.size());
+
+ InferenceEngine::TensorIterator::Body temp_body;
+ temp_body.inputs.resize(in_info_map_with_parameters.size());
+ temp_body.outputs.resize(out_info_map.size());
- // Fill the map to get layer and port of the body by the parameter index.
+ // Fill inputs/outs in order aligned with ng representation
uint64_t counter = 0;
for (const auto& param : parameters) {
auto info = in_info_map_with_parameters.at(param->get_friendly_name());
- auto data_ptr = info->getInputData();
- auto input_to = getInputTo(data_ptr);
- for (const auto& next_layer : input_to) {
- auto port_idx = find_input_idx(next_layer.second, data_ptr);
- ngraph_parameter_id_to_ie_layer_port[counter].push_back({next_layer.first, port_idx});
- }
- counter++;
+ temp_body.inputs[counter++] = info->getInputData();
}
- // Temporary body to call deep copy
- InferenceEngine::TensorIterator::Body temp_body;
- for (const auto& in : in_info_map_with_parameters) {
- temp_body.inputs.emplace_back(in.second->getInputData());
- }
+ auto map_ng_result_to_ie_name = [] (std::shared_ptr<ngraph::op::v0::Result> res_op) {
+ auto result = res_op->input(0).get_source_output();
- for (const auto& out : out_info_map) {
- temp_body.outputs.emplace_back(out.second);
+ std::string name = result.get_node()->get_friendly_name();
+ if (result.get_node()->get_output_size() > 1) {
+ name += "." + std::to_string(result.get_index());
+ }
+ return name;
+ };
+
+ counter = 0;
+ for (const auto& result : results) {
+ auto data = out_info_map.at(map_ng_result_to_ie_name(result));
+ temp_body.outputs[counter++] = data;
}
// This deep copy will hold all unreachable constants. See the comment in CopyTIBody function.
- auto deep_cp_body = InferenceEngine::NetPass::CopyTIBody(temp_body);
- for (const auto& data_ptr : deep_cp_body.inputs) {
- auto input_to = getInputTo(data_ptr);
- for (const auto& node : input_to) {
- // Make it compatible with ir v7: delete Input layers in body
- if (node.second->type != "Input") {
- body_input_layers.emplace(node.second);
- // Save information about data nodes to re-create them with correct names.
- for (const auto& data : node.second->insData) {
- layer_name_to_tensor_desc[node.second->name].emplace_back(data.lock()->getTensorDesc());
- }
- }
+ body = InferenceEngine::NetPass::CopyTIBody(temp_body);
+
+ // Check if data is really const layer holder
+ auto is_constant_holder = [] (const DataPtr data) {
+ return data->getPrecision() == Precision::UNSPECIFIED;
+ };
+
+ // Strip unreached node holder from Inputs node.
+ auto holder = body.inputs.back();
+ if (is_constant_holder(holder)) {
+ auto& holder_map = getInputTo(holder);
+ // remove_if
+ for( auto it = holder_map.begin(); it != holder_map.end(); ) {
+ if( it->second->type == "Input")
+ it = holder_map.erase(it);
+ else
+ ++it;
}
}
- for (const auto& data_ptr : deep_cp_body.outputs) {
- out_info_map[data_ptr->getName()] = data_ptr;
- }
- }
-
- auto holder = std::make_shared<Data>("const_holder", Precision::UNSPECIFIED);
- for (const auto& input_layer : body_input_layers) {
- // Save all constants to the holder so that they are not deleted.
- if (input_layer->insData.empty()) {
- getInputTo(holder)[input_layer->name] = input_layer;
- continue;
+ // TODO: Disable this WA after total switch onto Ngraph
+ // WA: Some plugins (like GPU) require matching of Data object name and producer Layer name.
+ // Data name is expected in format "[layer_name]" or "[layer_name].[port_idx]" in case
+ // of multiple inputs. We have to restore it if possible and ignore original names of
+ // Ngraph parameter and result ops.
+ // Will not change data name if:
+ // - data has several consumer layers
+ // - data has no consumer (example if data is straight used as output)
+ //
+ for (auto &in : body.inputs) {
+ if (is_constant_holder(in))
+ continue;
+
+ const auto input_to = getInputTo(in);
+ if (input_to.size() != 1)
+ continue;
+
+ const auto consumer_layer = input_to.begin()->second;
+ const auto consumer_in_port_set = consumer_layer->insData;
+ const auto found = std::find_if(consumer_in_port_set.begin(), consumer_in_port_set.end(),
+ [&in] (const DataWeakPtr &wptr) { return wptr.lock() == in; });
+ IE_ASSERT(found != consumer_in_port_set.end());
+ const auto consumer_port_idx = std::distance(consumer_in_port_set.begin(), found);
+
+ auto new_name = consumer_layer->name;
+ if (consumer_in_port_set.size() > 1) {
+ new_name += '.' + std::to_string(consumer_port_idx);
+ }
+ in->setName(new_name);
}
- // Re-create the data nodes with the correct names and fill inputs of TensorIterator (ie)
- for (size_t i = 0; i < input_layer->insData.size(); i++) {
- if (!input_layer->insData[i].lock()) {
- std::string data_name = (input_layer->insData.size() == 1)
- ? input_layer->name
- : input_layer->name + "." + std::to_string(i);
-
- DataPtr data(new Data(data_name, layer_name_to_tensor_desc[input_layer->name][i]));
- input_layer->insData[i] = data;
- getInputTo(data)[input_layer->name] = input_layer;
- in_info_map[data_name] = data;
+ // TODO: this WA restore original precisions of outputs.
+ // convertFunctionToICNNNetwork has internal fallback policy for unsupported
+ // precisions for inputs/outputs ports. Particular for U8 will be translated
+ // to FP32. However Loop body has strong requirements for continue_condition
+ // port, it should be BOOL(U8).
+ //
+ for (int i = 0; i < results.size(); i++) {
+ auto result = results[i];
+ auto output = body.outputs[i];
+ if (result->get_element_type() == ngraph::element::u8) {
+ output->setPrecision(InferenceEngine::Precision::U8);
}
}
}
LayerParams params = {layer->get_friendly_name(), "TensorIterator",
details::convertPrecision(layer->get_output_element_type(0))};
auto res = std::make_shared<InferenceEngine::TensorIterator>(params);
-
- // Body: inputs
- uint64_t counter = 0;
- for (const auto& in : in_info_map) {
- res->body.inputs.emplace_back(in.second);
-
- // Fill the map to get the input index by layer and port of the body.
- auto input_to = getInputTo(in.second);
- for (const auto& next_layer : input_to) {
- auto port_idx = find_input_idx(next_layer.second, in.second);
- ie_layer_port_to_tensor_iterator_input_id[{next_layer.first, port_idx}] = counter;
- }
- counter++;
- }
- // the holder should be the last input element.
- res->body.inputs.emplace_back(holder);
-
- // Body: outputs
- for (const auto& out : out_info_map) {
- res->body.outputs.emplace_back(out.second);
- }
+ res->body = body;
// Port map: outputs
for (const auto& desc : tensor_iterator->get_output_descriptions()) {
- auto result = results[desc->m_body_value_index]->input(0).get_source_output();
-
- std::string name = result.get_node()->get_friendly_name();
- if (result.get_node()->get_output_size() > 1) {
- name += "." + std::to_string(result.get_index());
- }
- auto output_layer = out_info_map.at(name);
-
- // Find index in outputs of the IE TensorIterator body
- auto it = std::find(res->body.outputs.begin(), res->body.outputs.end(), output_layer);
- if (it == res->body.outputs.end()) {
- THROW_IE_EXCEPTION << "Output layer not found.";
- }
- auto body_output_idx = it - res->body.outputs.begin();
+ auto body_output_idx = desc->m_body_value_index;
std::string type_name = desc->get_type_info().name;
if (type_name == "ConcatOutputDescription") {
// Port map : inputs and back edges
for (const auto& desc : tensor_iterator->get_input_descriptions()) {
- for (const auto& mapping : ngraph_parameter_id_to_ie_layer_port[desc->m_body_parameter_index]) {
- auto body_input_index = ie_layer_port_to_tensor_iterator_input_id.at(mapping);
- std::string type_name = desc->get_type_info().name;
-
- if (type_name == "SliceInputDescription") {
- auto input_desc = ::ngraph::as_type_ptr<ngraph::op::TensorIterator::SliceInputDescription>(desc);
- IE_ASSERT(input_desc != nullptr);
-
- res->input_port_map.emplace_back(InferenceEngine::TensorIterator::PortMap {
- static_cast<int>(input_desc->m_input_index), static_cast<int>(body_input_index),
- static_cast<int>(input_desc->m_axis), static_cast<int>(input_desc->m_stride),
- static_cast<int>(input_desc->m_start), static_cast<int>(input_desc->m_end),
- static_cast<int>(input_desc->m_part_size)});
- } else if (type_name == "MergedInputDescription") {
- auto input_desc = ::ngraph::as_type_ptr<ngraph::op::TensorIterator::MergedInputDescription>(desc);
- IE_ASSERT(input_desc != nullptr);
-
- res->input_port_map.emplace_back(InferenceEngine::TensorIterator::PortMap {
- static_cast<int>(input_desc->m_input_index), static_cast<int>(body_input_index), -1, 1, 0, -1, 1});
-
- auto result = results[input_desc->m_body_value_index]->inputs()[0].get_source_output();
-
- // Create correct name for output.
- std::string output_name = result.get_node()->get_friendly_name();
- if (result.get_node()->get_output_size() > 1) {
- output_name += "." + std::to_string(result.get_index());
- }
-
- auto output_layer = out_info_map.at(output_name);
- // Find index in outputs of the IE TensorIterator body
- auto it = std::find(res->body.outputs.begin(), res->body.outputs.end(), output_layer);
- if (it == res->body.outputs.end()) {
- THROW_IE_EXCEPTION << "Output layer not found.";
- }
- auto body_output_idx = it - res->body.outputs.begin();
-
- res->back_edges.emplace_back(InferenceEngine::TensorIterator::PortMap {
- static_cast<int>(body_output_idx), static_cast<int>(body_input_index), -1, 1, 0, -1, 1});
- } else if (type_name == "InvariantInputDescription") {
- auto input_desc = ::ngraph::as_type_ptr<ngraph::op::TensorIterator::InvariantInputDescription>(desc);
- IE_ASSERT(input_desc != nullptr);
-
- res->input_port_map.emplace_back(InferenceEngine::TensorIterator::PortMap {
- static_cast<int>(input_desc->m_input_index), static_cast<int>(body_input_index), -1, 1, 0, -1, 1});
- } else {
- THROW_IE_EXCEPTION << "Incorrect type of the input description.";
- }
+ auto body_input_index = desc->m_body_parameter_index;
+
+ if (const auto slice_desc = std::dynamic_pointer_cast<ngraph::op::TensorIterator::SliceInputDescription>(desc)) {
+ res->input_port_map.emplace_back(InferenceEngine::TensorIterator::PortMap {
+ static_cast<int>(slice_desc->m_input_index), static_cast<int>(body_input_index),
+ static_cast<int>(slice_desc->m_axis), static_cast<int>(slice_desc->m_stride),
+ static_cast<int>(slice_desc->m_start), static_cast<int>(slice_desc->m_end),
+ static_cast<int>(slice_desc->m_part_size)});
+ } else if (const auto merge_desc = std::dynamic_pointer_cast<ngraph::op::TensorIterator::MergedInputDescription>(desc)) {
+ res->input_port_map.emplace_back(InferenceEngine::TensorIterator::PortMap {
+ static_cast<int>(merge_desc->m_input_index), static_cast<int>(body_input_index), -1, 1, 0, -1, 1});
+
+ auto body_output_idx = merge_desc->m_body_value_index;
+
+ res->back_edges.emplace_back(InferenceEngine::TensorIterator::PortMap {
+ static_cast<int>(body_output_idx), static_cast<int>(body_input_index), -1, 1, 0, -1, 1});
+ } else if (const auto inv_desc = std::dynamic_pointer_cast<ngraph::op::TensorIterator::InvariantInputDescription>(desc)) {
+ res->input_port_map.emplace_back(InferenceEngine::TensorIterator::PortMap {
+ static_cast<int>(inv_desc->m_input_index), static_cast<int>(body_input_index), -1, 1, 0, -1, 1});
+ } else {
+ THROW_IE_EXCEPTION << "Incorrect type of the input description.";
+ }
+ }
+
+ if (const auto loop_op = std::dynamic_pointer_cast<const ngraph::opset5::Loop>(layer)) {
+ auto spec_port = loop_op->get_special_body_ports();
+ if (spec_port.current_iteration_input_idx != -1) {
+ auto ie_port_idx = spec_port.current_iteration_input_idx;
+ res->params["loop_body_current_iteration_idx"] = std::to_string(ie_port_idx);
+ }
+ if (spec_port.body_condition_output_idx != -1) {
+ auto body_output_idx = spec_port.body_condition_output_idx;
+ res->params["loop_body_condition_output_idx"] = std::to_string(body_output_idx);
}
+ res->params["loop_trip_count_idx"] = "0";
+ res->params["loop_execution_condition_idx"] = "1";
}
return res;
projects / platform / upstream / dldt.git / commitdiff