m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge pull request #10531 from typhoonzero/refine_grpc_serde_code

Browse Source 
Refine serde code
testDrivenImageClassification
Wu Yi 7 years ago committed by GitHub
parent 6d371e452e 796a448ce4
commit 61343fbf53
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 172 additions and 178 deletions
Show Stats Download Patch File Download Diff File

218
paddle/fluid/operators/detail/sendrecvop_utils.cc
View File

File diff suppressed because it is too large Load Diff

6
paddle/fluid/operators/detail/serde_test.cc
Unescape View File

@ -117,11 +117,11 @@ void RunTestLodTensor(platform::Place place, int from_type = 0) {
// serialize var to ByteBuffer
framework::Variable var;
auto* tensor = var.GetMutable<framework::LoDTensor>();
tensor->Resize(framework::make_ddim({4, 8, 4, 2}));
tensor->Resize(framework::make_ddim({512, 8, 4, 2}));
framework::LoD lod;
lod.push_back(framework::Vector<size_t>({1, 3, 8}));
tensor->set_lod(lod);
int tensor_numel = 4 * 8 * 4 * 2;
int tensor_numel = 512 * 8 * 4 * 2;
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
auto& ctx = *pool.Get(place);
tensor->mutable_data<float>(place);
@ -142,7 +142,7 @@ void RunTestLodTensor(platform::Place place, int from_type = 0) {
EXPECT_TRUE(varmsg.ParseFromString(tmp));
EXPECT_EQ(varmsg.varname(), "myvar");
EXPECT_EQ(varmsg.type(), 0);
EXPECT_EQ(varmsg.dims()[0], 4);
EXPECT_EQ(varmsg.dims()[0], 512);
EXPECT_EQ(varmsg.dims()[1], 8);
EXPECT_EQ(varmsg.dims()[2], 4);
EXPECT_EQ(varmsg.dims()[3], 2);

40
paddle/fluid/operators/detail/variable_response.cc
Unescape View File

@ -210,15 +210,15 @@ bool ParseLodData(::google::protobuf::io::CodedInputStream* input,
}
if (wt == WIRETYPE_LENGTH_DELIMITED) {
int length = 0;
if (!input->ReadVarintSizeAsInt(&length)) {
int num_bytes = 0;
if (!input->ReadVarintSizeAsInt(&num_bytes)) {
return tag;
}
for (int i = 0; i < length; i++) {
int start_pos = input->CurrentPosition();
while (input->CurrentPosition() - start_pos < num_bytes) {
uint64_t v;
if (!input->ReadVarint64(&v)) {
return false;
return tag;
}
lod->push_back(v);
}
@ -275,8 +275,8 @@ int VariableResponse::Parse(Source* source) {
break;
}
case sendrecv::VariableMessage::kTypeFieldNumber: {
uint64_t v;
if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) {
uint32_t v;
if ((wt != WIRETYPE_VARINT) || !input.ReadVarint32(&v)) {
return tag;
}
@ -284,8 +284,8 @@ int VariableResponse::Parse(Source* source) {
break;
}
case sendrecv::VariableMessage::kDataTypeFieldNumber: {
uint64_t v = 0;
if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) {
uint32_t v = 0;
if ((wt != WIRETYPE_VARINT) || !input.ReadVarint32(&v)) {
return tag;
}
@ -305,11 +305,12 @@ int VariableResponse::Parse(Source* source) {
// packed
if (wt == WIRETYPE_LENGTH_DELIMITED) {
int length = 0;
if (!input.ReadVarintSizeAsInt(&length)) {
int num_bytes = 0;
if (!input.ReadVarintSizeAsInt(&num_bytes)) {
return tag;
}
for (int i = 0; i < length; i++) {
int start_pos = input.CurrentPosition();
while (input.CurrentPosition() - start_pos < num_bytes) {
uint64_t v;
if (!input.ReadVarint64(&v)) {
return tag;
@ -318,7 +319,6 @@ int VariableResponse::Parse(Source* source) {
}
break;
}
return tag;
}
case sendrecv::VariableMessage::kLodLevelFieldNumber: {
@ -372,9 +372,9 @@ int VariableResponse::Parse(Source* source) {
meta_.varname() != "",
"meta info should be got first!");
int length = 0;
int num_bytes = 0;
if (wt != WIRETYPE_LENGTH_DELIMITED ||
!ReadVarintSizeAsInt(&input, &length)) {
!ReadVarintSizeAsInt(&input, &num_bytes)) {
return tag;
}
@ -382,14 +382,14 @@ int VariableResponse::Parse(Source* source) {
if (meta_.type() == sendrecv::LOD_TENSOR) {
PADDLE_ENFORCE(meta_.lod_size() >= 0,
"lod info should be got first!");
if (!CopyLodTensorData(&input, *dev_ctx_, dims, length)) {
if (!CopyLodTensorData(&input, *dev_ctx_, dims, num_bytes)) {
return tag;
}
break;
}
if (meta_.type() == sendrecv::SELECTED_ROWS) {
if (!CopySelectRowsTensorData(&input, *dev_ctx_, dims, length)) {
if (!CopySelectRowsTensorData(&input, *dev_ctx_, dims, num_bytes)) {
return tag;
}
break;
@ -403,13 +403,13 @@ int VariableResponse::Parse(Source* source) {
meta_.varname() != "",
"meta info should be got first!");
int length = 0;
int num_bytes = 0;
if (wt != WIRETYPE_LENGTH_DELIMITED ||
!ReadVarintSizeAsInt(&input, &length)) {
!ReadVarintSizeAsInt(&input, &num_bytes)) {
return tag;
}
if (!CopySelectRowsData(&input, *dev_ctx_, length)) {
if (!CopySelectRowsData(&input, *dev_ctx_, num_bytes)) {
return tag;
}
break;

86
python/paddle/fluid/transpiler/distribute_transpiler.py
Unescape View File

@ -18,7 +18,9 @@ import math
import distributed_splitter as splitter
from .. import core
from ..framework import Program, default_main_program, Variable, Parameter
from ..framework import Program, default_main_program, \
default_startup_program, \
Variable, Parameter, grad_var_name
LOOKUP_TABLE_TYPE = "lookup_table"
LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad"
@ -153,43 +155,43 @@ class DistributeTranspiler:
split_method=splitter.round_robin,
sync_mode=True):
"""
Transpile the program to distributed data-parallelism programs.
The main_program will be transformed to use a remote parameter server
to do parameter optimization. And the optimization graph will be put
into a parameter server program.
Use different methods to split trainable variables to different
parameter servers.
Steps to transpile trainer:
1. split variable to multiple blocks, aligned by product(dim[1:]) (width).
2. rename splited grad variables to add trainer_id suffix ".trainer_%d".
3. modify trainer program add split_op to each grad variable.
4. append send_op to send splited variables to server and fetch
params(splited blocks or origin param) from server.
5. append concat_op to merge splited blocks to update local weights.
Steps to transpile pserver:
1. create new program for parameter server.
2. create params and grad variables that assigned to current server instance.
3. create a sub-block in the server side program
4. append ops that should run on current server instance.
5. add listen_and_serv op
:param trainer_id: one unique id for each trainer in a job.
:type trainer_id: int
:param program: program to transpile, default is default_main_program
:type program: Program
:param pservers: parameter server endpoints like "m1:6174,m2:6174"
:type pservers: string
:param trainers: total number of workers/trainers in the job
:type trainers: int
:param split_method: A function to determin how to split variables
to different servers equally.
:type split_method: function
:type sync_mode: boolean default True
:param sync_mode: if sync_mode is set True, it means that dist transpiler
will transpile the program into sync_mode pserver and trainer program.
Transpile the program to distributed data-parallelism programs.
The main_program will be transformed to use a remote parameter server
to do parameter optimization. And the optimization graph will be put
into a parameter server program.
Use different methods to split trainable variables to different
parameter servers.
Steps to transpile trainer:
1. split variable to multiple blocks, aligned by product(dim[1:]) (width).
2. rename splited grad variables to add trainer_id suffix ".trainer_%d".
3. modify trainer program add split_op to each grad variable.
4. append send_op to send splited variables to server and fetch
params(splited blocks or origin param) from server.
5. append concat_op to merge splited blocks to update local weights.
Steps to transpile pserver:
1. create new program for parameter server.
2. create params and grad variables that assigned to current server instance.
3. create a sub-block in the server side program
4. append ops that should run on current server instance.
5. add listen_and_serv op
:param trainer_id: one unique id for each trainer in a job.
:type trainer_id: int
:param program: program to transpile, default is default_main_program
:type program: Program
:param pservers: parameter server endpoints like "m1:6174,m2:6174"
:type pservers: string
:param trainers: total number of workers/trainers in the job
:type trainers: int
:param split_method: A function to determin how to split variables
to different servers equally.
:type split_method: function
:type sync_mode: boolean default True
:param sync_mode: if sync_mode is set True, it means that dist transpiler
will transpile the program into sync_mode pserver and trainer program.
"""
assert (callable(split_method))
if program is None:
@ -244,7 +246,7 @@ class DistributeTranspiler:
]
grad_list = [
grad for grad in grad_list
if grad.name != framework.grad_var_name(self.table_name)
if grad.name != grad_var_name(self.table_name)
]
self.table_param_grad = [
param_grad for param_grad in params_grads
@ -494,7 +496,7 @@ class DistributeTranspiler:
were split to several blocks.
"""
s_prog = Program()
orig_s_prog = framework.default_startup_program()
orig_s_prog = default_startup_program()
params = self.param_grad_ep_mapping[endpoint]["params"]
def _get_splited_name_and_shape(varname):
@ -619,7 +621,7 @@ class DistributeTranspiler:
# 2. add split_ids_op and send_vars_op to send gradient to pservers
# there should only be one table_name
all_ops = program.global_block().ops
table_grad_name = framework.grad_var_name(self.table_name)
table_grad_name = grad_var_name(self.table_name)
for op in all_ops:
if table_grad_name in op.output_arg_names:
op_index = list(all_ops).index(op)
@ -692,7 +694,7 @@ class DistributeTranspiler:
persistable=True)
grad_var = _clone_var(
pserver_program.global_block(),
self.origin_program.global_block().vars[framework.grad_var_name(
self.origin_program.global_block().vars[grad_var_name(
self.table_name)],
persistable=False)

Write Preview
Loading…
Cancel
Save