提交 962be45a 编写于 作者: G guru4elephant

fix general model server

上级 f9d9463f
[{
"dict_name": "test_dict",
"shard": 2,
"dup": 1,
"timeout": 200,
"retry": 3,
"backup_request": 100,
"type": "ipport_list",
"load_balancer": "rr",
"nodes": [{
"ipport_list": "list://xxx.xxx.xxx.xxx:8000"
},{
"ipport_list": "list://xxx.xxx.xxx.xxx:8000"
}]
}]
is_lod_feed: true
is_lod_feed: false
is_lod_feed: true
feed_type: 1
feed_type: 0
feed_type: 1
feed_shape {
shape: -1
}
feed_shape {
shape: 1
shape: 2
shape: 3
}
feed_shape {
shape: -1
}
--enable_model_toolkit
--enable_cube=false
--enable_general_model=true
--general_model_path=./conf
--general_model_file=general_model.prototxt
engines {
name: "general_model"
type: "FLUID_CPU_ANALYSIS_DIR"
reloadable_meta: "./data/model/paddle/fluid_time_file"
reloadable_type: "timestamp_ne"
model_data_path: "./data/model/paddle/fluid/text_classification"
runtime_thread_num: 0
batch_infer_size: 0
enable_batch_align: 0
}
model_toolkit_path: "./conf/"
model_toolkit_file: "model_toolkit.prototxt"
port: 9292
services {
name: "GeneralModelService"
workflows: "workflow1"
}
\ No newline at end of file
workflows {
name: "workflow1"
workflow_type: "Sequence"
nodes {
name: "general_model_op"
type: "GeneralModelOp"
}
}
\ No newline at end of file
......@@ -43,14 +43,17 @@ int GeneralModelOp::inference() {
std::vector<int> elem_type;
std::vector<int> elem_size;
std::vector<int> capacity;
if (batch_size > 0) {
int var_num = req->insts(0).tensor_array_size();
VLOG(3) << "var num: " << var_num;
elem_type.resize(var_num);
elem_size.resize(var_num);
capacity.resize(var_num);
paddle::PaddleTensor lod_tensor;
for (int i = 0; i < var_num; ++i) {
elem_type[i] = req->insts(0).tensor_array(i).elem_type();
VLOG(3) << "var[" << i << "] has elem type: " << elem_type[i];
if (elem_type[i] == 0) { // int64
elem_size[i] = sizeof(int64_t);
lod_tensor.dtype = paddle::PaddleDType::INT64;
......@@ -58,9 +61,11 @@ int GeneralModelOp::inference() {
elem_size[i] = sizeof(float);
lod_tensor.dtype = paddle::PaddleDType::FLOAT32;
}
if (req->insts(0).tensor_array(i).shape(0) == -1) {
lod_tensor.lod.resize(1);
lod_tensor.lod[0].push_back(0);
VLOG(3) << "var[" << i << "] is lod_tensor";
} else {
lod_tensor.shape.push_back(batch_size);
capacity[i] = 1;
......@@ -68,42 +73,76 @@ int GeneralModelOp::inference() {
k < req->insts(0).tensor_array(i).shape_size();
++k) {
int dim = req->insts(0).tensor_array(i).shape(k);
VLOG(3) << "shape for var[" << i << "]: " << dim;
capacity[i] *= dim;
lod_tensor.shape.push_back(dim);
}
VLOG(3) << "var[" << i << "] is tensor, capacity: " << capacity[i];
}
if (i == 0) {
lod_tensor.name = "words";
} else {
lod_tensor.name = "label";
}
in->push_back(lod_tensor);
}
for (int i = 0; i < var_num; ++i) {
if ((*in)[i].lod.size() > 0) {
if (in->at(i).lod.size() == 1) {
for (int j = 0; j < batch_size; ++j) {
const Tensor & tensor = req->insts(j).tensor_array(i);
int data_len = tensor.data_size() / elem_size[i];
int cur_len = (*in)[i].lod[0].back();
(*in)[i].lod[0].push_back(cur_len + data_len);
int data_len = tensor.data_size();
VLOG(3) << "tensor size for var[" << i << "]: "
<< tensor.data_size();
int cur_len = in->at(i).lod[0].back();
VLOG(3) << "current len: " << cur_len;
in->at(i).lod[0].push_back(cur_len + data_len);
VLOG(3) << "new len: " << cur_len + data_len;
}
(*in)[i].data.Resize((*in)[i].lod[0].back());
in->at(i).data.Resize(in->at(i).lod[0].back() * elem_size[i]);
in->at(i).shape = {in->at(i).lod[0].back(), 1};
VLOG(3) << "var[" << i << "] is lod_tensor and len="
<< in->at(i).lod[0].back();
} else {
(*in)[i].data.Resize(batch_size * capacity[i]);
in->at(i).data.Resize(batch_size * capacity[i] * elem_size[i]);
VLOG(3) << "var[" << i << "] is tensor and capacity="
<< batch_size * capacity[i];
}
}
for (int i = 0; i < var_num; ++i) {
void * dst_ptr = (*in)[i].data.data();
int offset = 0;
for (int j = 0; j < batch_size; ++j) {
memcpy(dst_ptr + offset,
(void *)(req->insts(j).tensor_array(i).data().data()),
req->insts(j).tensor_array(i).data_size() * elem_size[i]);
if ((*in)[i].lod.size() > 0) {
offset += (*in)[i].lod[0][j + 1] * elem_size[i];
} else {
offset += capacity[i] * elem_size[i];
if (elem_type[i] == 0) {
int64_t * dst_ptr = static_cast<int64_t *>(in->at(i).data.data());
int offset = 0;
for (int j = 0; j < batch_size; ++j) {
for (int k = 0; k < req->insts(j).tensor_array(i).data_size(); ++k) {
dst_ptr[offset + k] =
*(const int64_t *)req->insts(j).tensor_array(i).data(k).c_str();
}
if (in->at(i).lod.size() == 1) {
offset = in->at(i).lod[0][j + 1];
} else {
offset += capacity[i];
}
}
} else {
float * dst_ptr = static_cast<float *>(in->at(i).data.data());
int offset = 0;
for (int j = 0; j < batch_size; ++j) {
for (int k = 0; k < req->insts(j).tensor_array(i).data_size(); ++k) {
dst_ptr[offset + k] =
*(const float *)req->insts(j).tensor_array(i).data(k).c_str();
}
if (in->at(i).lod.size() == 1) {
offset = in->at(i).lod[0][j + 1];
} else {
offset += capacity[i];
}
}
}
}
VLOG(3) << "going to infer";
TensorVector *out = butil::get_object<TensorVector>();
if (!out) {
LOG(ERROR) << "Failed get tls output object";
......@@ -119,44 +158,42 @@ int GeneralModelOp::inference() {
Response * res = mutable_data<Response>();
// we suppose the dtype of all fetch variables is float
for (int i = 0; i < batch_size; ++i) {
FetchInst * fetch_inst = res->add_insts();
for (int j = 0; j < out->size(); ++j) {
Tensor * tensor = fetch_inst->add_tensor_array();
tensor->set_elem_type(1);
if ((*out)[j].lod.size() > 0) {
if (out->at(j).lod.size() == 1) {
tensor->add_shape(-1);
tensor->mutable_data()->Reserve(
(*out)[j].lod[0].back() * sizeof(float));
} else {
int cap = 1;
for (int k = 1; k < (*out)[j].shape.size(); ++k) {
cap *= (*out)[j].shape[k];
tensor->add_shape((*out)[j].shape[k]);
for (int k = 1; k < out->at(j).shape.size(); ++k) {
tensor->add_shape(out->at(j).shape[k]);
}
tensor->mutable_data()->Reserve(cap * sizeof(float));
}
}
}
for (int i = 0; i < out->size(); ++i) {
if ((*out)[i].lod.size() > 0) {
float * data_ptr = static_cast<float *>(out->at(i).data.data());
int cap = 1;
for (int j = 0; j < out->at(i).shape.size(); ++j) {
cap *= out->at(i).shape[j];
}
if (out->at(i).lod.size() == 1) {
for (int j = 0; j < batch_size; ++j) {
Tensor * tensor = res->mutable_insts(j)->mutable_tensor_array(i);
void * dst_ptr = tensor->mutable_data()->mutable_data();
memcpy(dst_ptr,
(*out)[i].data.data() + (*out)[i].lod[0][j] * elem_size[i],
((*out)[i].lod[0][j + 1] - (*out)[i].lod[0][j])
* elem_size[i]);
for (int k = out->at(i).lod[0][j];
k < out->at(i).lod[0][j + 1];
k++) {
res->mutable_insts(j)->mutable_tensor_array(i)->add_data(
(char *)(&(data_ptr[k])), sizeof(float));
}
}
} else {
for (int j = 0; j < batch_size; ++j) {
Tensor * tensor = res->mutable_insts(j)->mutable_tensor_array(i);
void * dst_ptr = tensor->mutable_data()->mutable_data();
memcpy(dst_ptr,
(*out)[i].data.data() + j * capacity[i] * elem_size[i],
capacity[i] * elem_size[i]);
for (int k = j * cap; k < (j + 1) * cap; ++k) {
res->mutable_insts(j)->mutable_tensor_array(i)->add_data(
(char *)(&(data_ptr[k])), sizeof(float));
}
}
}
}
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册