EngineX-Iluvatar/enginex_bi_series-sherpa-onnx
Archived
16
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Adding warm up for Zipformer2 (#766)

Browse Source 
Signed-off-by: manickavela1998@gmail.com <manickavela1998@gmail.com>
This commit is contained in:
Manix
2024-04-16 06:46:55 +05:30
committed by GitHub
parent 5981adf454
commit fb4aee83ac
9 changed files with 99 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
sherpa-onnx/csrc/online-recognizer-transducer-impl.h
View File

@@ -32,6 +32,7 @@
#include "sherpa-onnx/csrc/online-transducer-modified-beam-search-decoder.h"
#include "sherpa-onnx/csrc/symbol-table.h"
#include "sherpa-onnx/csrc/utils.h"
#include "sherpa-onnx/csrc/onnx-utils.h"
namespace sherpa_onnx {
@@ -183,6 +184,41 @@ class OnlineRecognizerTransducerImpl : public OnlineRecognizerImpl {
s->NumFramesReady();
}
// Warmping up engine with wp: warm_up count and max-batch-size
void WarmpUpRecognizer(int32_t warmup, int32_t mbs) const {
auto max_batch_size = mbs;
if (warmup <= 0 || warmup > 100) {
return;
}
int32_t chunk_size = model_->ChunkSize();
int32_t chunk_shift = model_->ChunkShift();
int32_t feature_dim = 80;
std::vector<OnlineTransducerDecoderResult> results(max_batch_size);
std::vector<float> features_vec(max_batch_size * chunk_size * feature_dim);
std::vector<std::vector<Ort::Value>> states_vec(max_batch_size);
auto memory_info =
Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
std::array<int64_t, 3> x_shape{max_batch_size, chunk_size, feature_dim};
for (int32_t i = 0; i != max_batch_size; ++i) {
states_vec[i] = model_->GetEncoderInitStates();
results[i] = decoder_->GetEmptyResult();
}
for (int32_t i = 0; i != warmup; ++i) {
auto states = model_->StackStates(states_vec);
Ort::Value x = Ort::Value::CreateTensor(memory_info, features_vec.data(),
features_vec.size(), x_shape.data(),
x_shape.size());
auto x_copy = Clone(model_->Allocator(), &x);
auto pair = model_->RunEncoder(std::move(x), std::move(states),
std::move(x_copy));
decoder_->Decode(std::move(pair.first), &results);
}
}
void DecodeStreams(OnlineStream **ss, int32_t n) const override {
int32_t chunk_size = model_->ChunkSize();
int32_t chunk_shift = model_->ChunkShift();