Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Remove unreferenced variables from ProgramDesc in prune() #7890

Merged
merged 3 commits into from
Jan 29, 2018
Merged

Remove unreferenced variables from ProgramDesc in prune() #7890

merged 3 commits into from
Jan 29, 2018

Conversation

kexinzhao
Copy link
Contributor

Fixes #7417

By following similar test procedure as in #7097, we get the testing result as follows:

FLAGS_dirname: /paddle/python/paddle/v2/fluid/tests/book/recognize_digits_mlp.inference.model
WARNING: Logging before InitGoogleLogging() is written to STDERR
I0125 23:58:51.136490 29702 inference.cc:26] loading model from /paddle/python/paddle/v2/fluid/tests/book/recognize_digits_mlp.inference.model/__model__
I0125 23:58:51.136664 29702 inference.cc:32] program_desc_str's size: 1351
I0125 23:58:51.136858 29702 inference.cc:78] parameter's name: fc_1.b_0
I0125 23:58:51.136916 29702 inference.cc:78] parameter's name: fc_0.b_0
I0125 23:58:51.136937 29702 inference.cc:78] parameter's name: fc_0.w_0
I0125 23:58:51.136955 29702 inference.cc:78] parameter's name: fc_1.w_0
I0125 23:58:51.136973 29702 inference.cc:78] parameter's name: fc_2.w_0
I0125 23:58:51.136991 29702 inference.cc:78] parameter's name: fc_2.b_0
W0125 23:58:51.137250 29702 init.cc:56] 'GPU' is not supported, Please re-compile with WITH_GPU option
I0125 23:58:51.137315 29702 scope.cc:48] Create variable fc_2.b_0
I0125 23:58:51.137331 29702 executor.cc:99] Create Variable fc_2.b_0 global, which pointer is 0x2024d70
I0125 23:58:51.137348 29702 scope.cc:48] Create variable fc_1.b_0
I0125 23:58:51.137359 29702 executor.cc:99] Create Variable fc_1.b_0 global, which pointer is 0x2024d90
I0125 23:58:51.137369 29702 scope.cc:48] Create variable fc_0.b_0
I0125 23:58:51.137399 29702 executor.cc:99] Create Variable fc_0.b_0 global, which pointer is 0x2025000
I0125 23:58:51.137413 29702 scope.cc:48] Create variable fc_2.w_0
I0125 23:58:51.137421 29702 executor.cc:99] Create Variable fc_2.w_0 global, which pointer is 0x2025080
I0125 23:58:51.137432 29702 scope.cc:48] Create variable fc_0.w_0
I0125 23:58:51.137441 29702 executor.cc:99] Create Variable fc_0.w_0 global, which pointer is 0x202fdd0
I0125 23:58:51.137451 29702 scope.cc:48] Create variable fc_1.w_0
I0125 23:58:51.137465 29702 executor.cc:99] Create Variable fc_1.w_0 global, which pointer is 0x202ff20
I0125 23:58:51.137636 29702 executor.cc:126] Op(load), inputs:{}, outputs:{Out[fc_1.b_0[64]({})]}.
I0125 23:58:51.137719 29702 executor.cc:126] Op(load), inputs:{}, outputs:{Out[fc_0.b_0[128]({})]}.
I0125 23:58:51.138204 29702 executor.cc:126] Op(load), inputs:{}, outputs:{Out[fc_0.w_0[784, 128]({})]}.
I0125 23:58:51.138305 29702 executor.cc:126] Op(load), inputs:{}, outputs:{Out[fc_1.w_0[128, 64]({})]}.
I0125 23:58:51.138361 29702 executor.cc:126] Op(load), inputs:{}, outputs:{Out[fc_2.w_0[64, 10]({})]}.
I0125 23:58:51.138411 29702 executor.cc:126] Op(load), inputs:{}, outputs:{Out[fc_2.b_0[10]({})]}.
I0125 23:58:51.139745 29702 feed_fetch_method.h:28] SetFeedVariable name=feed index=0
I0125 23:58:51.139770 29702 scope.cc:48] Create variable feed
I0125 23:58:51.139794 29702 scope.cc:48] Create variable fc_1.tmp_1
I0125 23:58:51.139806 29702 executor.cc:104] Create Variable fc_1.tmp_1 locally, which pointer is 0x2035410
I0125 23:58:51.139817 29702 scope.cc:48] Create variable fc_0.tmp_0
I0125 23:58:51.139827 29702 executor.cc:104] Create Variable fc_0.tmp_0 locally, which pointer is 0x2030560
I0125 23:58:51.139837 29702 scope.cc:48] Create variable fc_1.tmp_0
I0125 23:58:51.139847 29702 executor.cc:104] Create Variable fc_1.tmp_0 locally, which pointer is 0x2030260
I0125 23:58:51.139856 29702 scope.cc:48] Create variable fc_0.tmp_2
I0125 23:58:51.139868 29702 executor.cc:104] Create Variable fc_0.tmp_2 locally, which pointer is 0x20357f0
I0125 23:58:51.139881 29702 scope.cc:48] Create variable fc_0.tmp_1
I0125 23:58:51.139889 29702 executor.cc:104] Create Variable fc_0.tmp_1 locally, which pointer is 0x2035900
I0125 23:58:51.139899 29702 scope.cc:48] Create variable fc_1.tmp_2
I0125 23:58:51.139909 29702 executor.cc:104] Create Variable fc_1.tmp_2 locally, which pointer is 0x2035a50
I0125 23:58:51.139919 29702 executor.cc:99] Create Variable fc_1.b_0 global, which pointer is 0x2024d90
I0125 23:58:51.139928 29702 scope.cc:48] Create variable fc_2.tmp_1
I0125 23:58:51.139938 29702 executor.cc:104] Create Variable fc_2.tmp_1 locally, which pointer is 0x2035b60
I0125 23:58:51.139946 29702 executor.cc:99] Create Variable feed global, which pointer is 0x2024cd0
I0125 23:58:51.139953 29702 scope.cc:48] Create variable fetch
I0125 23:58:51.139962 29702 executor.cc:99] Create Variable fetch global, which pointer is 0x2035c70
I0125 23:58:51.139972 29702 scope.cc:48] Create variable fc_2.tmp_2
I0125 23:58:51.139981 29702 executor.cc:104] Create Variable fc_2.tmp_2 locally, which pointer is 0x2035d10
I0125 23:58:51.139991 29702 scope.cc:48] Create variable fc_2.tmp_0
I0125 23:58:51.140000 29702 executor.cc:104] Create Variable fc_2.tmp_0 locally, which pointer is 0x2035e20
I0125 23:58:51.140010 29702 executor.cc:99] Create Variable fc_0.b_0 global, which pointer is 0x2025000
I0125 23:58:51.140019 29702 executor.cc:99] Create Variable fc_0.w_0 global, which pointer is 0x202fdd0
I0125 23:58:51.140028 29702 executor.cc:99] Create Variable fc_1.w_0 global, which pointer is 0x202ff20
I0125 23:58:51.140038 29702 scope.cc:48] Create variable x
I0125 23:58:51.140048 29702 executor.cc:104] Create Variable x locally, which pointer is 0x2035f30
I0125 23:58:51.140064 29702 executor.cc:99] Create Variable fc_2.w_0 global, which pointer is 0x2025080
I0125 23:58:51.140072 29702 executor.cc:99] Create Variable fc_2.b_0 global, which pointer is 0x2024d70
I0125 23:58:51.140100 29702 feed_op.cc:44] Feed Var feed's 0 column to var x
I0125 23:58:51.140112 29702 tensor_util.h:36] Copy 1, 784 from CPUPlace to CPUPlace
I0125 23:58:51.140149 29702 executor.cc:126] Op(feed), inputs:{X[feed[-1]({{}})]}, outputs:{Out[x[1, 784]({})]}.
I0125 23:58:51.140187 29702 mul_op.cc:36] mul operator x.shape=1, 784 y.shape=784, 128 x_num_col_dims=1 y_num_col_dims=1
I0125 23:58:51.140211 29702 operator.cc:488] expected_kernel_key:data_type[5]:data_layout[ANY_LAYOUT]:place[CPUPlace]:library_type[PLAIN]
I0125 23:58:51.152216 29702 executor.cc:126] Op(mul), inputs:{X[x[1, 784]({})], Y[fc_0.w_0[784, 128]({})]}, outputs:{Out[fc_0.tmp_0[1, 128]({})]}.
I0125 23:58:51.152279 29702 operator.cc:488] expected_kernel_key:data_type[5]:data_layout[ANY_LAYOUT]:place[CPUPlace]:library_type[PLAIN]
I0125 23:58:51.152340 29702 executor.cc:126] Op(elementwise_add), inputs:{X[fc_0.tmp_0[1, 128]({})], Y[fc_0.b_0[128]({})]}, outputs:{Out[fc_0.tmp_1[1, 128]({})]}.
I0125 23:58:51.152426 29702 operator.cc:488] expected_kernel_key:data_type[5]:data_layout[ANY_LAYOUT]:place[CPUPlace]:library_type[PLAIN]
I0125 23:58:51.152494 29702 executor.cc:126] Op(relu), inputs:{X[fc_0.tmp_1[1, 128]({})]}, outputs:{Out[fc_0.tmp_2[1, 128]({})]}.
I0125 23:58:51.152621 29702 mul_op.cc:36] mul operator x.shape=1, 128 y.shape=128, 64 x_num_col_dims=1 y_num_col_dims=1
I0125 23:58:51.152664 29702 operator.cc:488] expected_kernel_key:data_type[5]:data_layout[ANY_LAYOUT]:place[CPUPlace]:library_type[PLAIN]
I0125 23:58:51.152767 29702 executor.cc:126] Op(mul), inputs:{X[fc_0.tmp_2[1, 128]({})], Y[fc_1.w_0[128, 64]({})]}, outputs:{Out[fc_1.tmp_0[1, 64]({})]}.
I0125 23:58:51.152817 29702 operator.cc:488] expected_kernel_key:data_type[5]:data_layout[ANY_LAYOUT]:place[CPUPlace]:library_type[PLAIN]
I0125 23:58:51.152873 29702 executor.cc:126] Op(elementwise_add), inputs:{X[fc_1.tmp_0[1, 64]({})], Y[fc_1.b_0[64]({})]}, outputs:{Out[fc_1.tmp_1[1, 64]({})]}.
I0125 23:58:51.152909 29702 operator.cc:488] expected_kernel_key:data_type[5]:data_layout[ANY_LAYOUT]:place[CPUPlace]:library_type[PLAIN]
I0125 23:58:51.152951 29702 executor.cc:126] Op(relu), inputs:{X[fc_1.tmp_1[1, 64]({})]}, outputs:{Out[fc_1.tmp_2[1, 64]({})]}.
I0125 23:58:51.152990 29702 mul_op.cc:36] mul operator x.shape=1, 64 y.shape=64, 10 x_num_col_dims=1 y_num_col_dims=1
I0125 23:58:51.153020 29702 operator.cc:488] expected_kernel_key:data_type[5]:data_layout[ANY_LAYOUT]:place[CPUPlace]:library_type[PLAIN]
I0125 23:58:51.153084 29702 executor.cc:126] Op(mul), inputs:{X[fc_1.tmp_2[1, 64]({})], Y[fc_2.w_0[64, 10]({})]}, outputs:{Out[fc_2.tmp_0[1, 10]({})]}.
I0125 23:58:51.153143 29702 operator.cc:488] expected_kernel_key:data_type[5]:data_layout[ANY_LAYOUT]:place[CPUPlace]:library_type[PLAIN]
I0125 23:58:51.153270 29702 executor.cc:126] Op(elementwise_add), inputs:{X[fc_2.tmp_0[1, 10]({})], Y[fc_2.b_0[10]({})]}, outputs:{Out[fc_2.tmp_1[1, 10]({})]}.
I0125 23:58:51.153374 29702 operator.cc:488] expected_kernel_key:data_type[5]:data_layout[ANY_LAYOUT]:place[CPUPlace]:library_type[PLAIN]
I0125 23:58:51.153532 29702 executor.cc:126] Op(softmax), inputs:{X[fc_2.tmp_1[1, 10]({})]}, outputs:{Out[fc_2.tmp_2[1, 10]({})]}.
I0125 23:58:51.153575 29702 tensor_util.h:36] Copy 1, 10 from CPUPlace to CPUPlace
I0125 23:58:51.153626 29702 fetch_op.cc:62] Fetch variable fc_2.tmp_2 to fetch
I0125 23:58:51.153672 29702 executor.cc:126] Op(fetch), inputs:{X[fc_2.tmp_2[1, 10]({})]}, outputs:{Out[fetch[-1]({{}})]}.
I0125 23:58:51.153836 29704 scope.cc:33] Destroy variable x
I0125 23:58:51.154021 29704 scope.cc:33] Destroy variable fc_2.tmp_0
I0125 23:58:51.154045 29704 scope.cc:33] Destroy variable fc_2.tmp_2
I0125 23:58:51.154078 29704 scope.cc:33] Destroy variable fc_2.tmp_1
I0125 23:58:51.154228 29704 scope.cc:33] Destroy variable fc_1.tmp_2
I0125 23:58:51.154316 29704 scope.cc:33] Destroy variable fc_0.tmp_1
I0125 23:58:51.154392 29704 scope.cc:33] Destroy variable fc_1.tmp_1
I0125 23:58:51.154435 29704 scope.cc:33] Destroy variable fc_0.tmp_0
I0125 23:58:51.154466 29704 scope.cc:33] Destroy variable fc_1.tmp_0
I0125 23:58:51.154597 29704 scope.cc:33] Destroy variable fc_0.tmp_2
I0125 23:58:51.154650 29702 feed_fetch_method.h:51] Fetch fetch with index 0 shape= 1, 10
I0125 23:58:51.154670 29702 scope.cc:33] Destroy variable fetch
I0125 23:58:51.154847 29702 scope.cc:33] Destroy variable feed
I0125 23:58:51.154989 29702 scope.cc:33] Destroy variable fc_2.b_0
I0125 23:58:51.155105 29702 scope.cc:33] Destroy variable fc_1.b_0
I0125 23:58:51.155323 29702 scope.cc:33] Destroy variable fc_0.b_0
I0125 23:58:51.155359 29702 scope.cc:33] Destroy variable fc_1.w_0
I0125 23:58:51.155376 29702 scope.cc:33] Destroy variable fc_0.w_0
I0125 23:58:51.155407 29702 scope.cc:33] Destroy variable fc_2.w_0
dims_i: 1 10
result: 0.106794 0.1657 0.0748172 0.0881855 0.119018 0.0405249 0.111904 0.103682 0.0953393 0.0940347

@Xreki Xreki added the 预测 原名Inference，包含Capi预测问题等 label Jan 26, 2018
Xreki
Xreki reviewed Jan 26, 2018
View reviewed changes
paddle/framework/prune.cc
auto* var_field = output->mutable_blocks(block_id)->mutable_vars();
for (const auto& var : *var_field) {
var_map[var.name()] = var;
}
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maybe, there is no need for an extra map. We can directly use var map of input ProgramDesc. We can use input.FindVar(name)->Proto() to get the proto::VarDesc.

Copy link
Contributor Author

@kexinzhao kexinzhao Jan 26, 2018 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There are two different classes one is framework::proto::xxxDesc, one is framework::xxxDesc. The latter is basically a wrapper about the former.

For this prune function, the input is proto::ProgramDesc, so we only have limited functionality provided by the protobuf library itself. After you compile the framework.proto, you will get framework.pb.h + cc, which lists all the available functions for proto::ProgramDesc and etc.

So I don't think we can use input.FindVar() here.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Later, when we are going to implement out transpiler function for inference optimization on the c++ side after loading a inference desc from file or buffer, we can use your suggestion, as in that case, we can take a framework::ProgramDesc as input.

@Xreki what do you think?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

For this prune function, the input is proto::ProgramDesc, so we only have limited functionality provided by the protobuf library itself.

You are right. Sorry for didn't notice that.

paddle/framework/prune.cc Outdated
for (const auto& op : *op_field) {
// add VarDescs of all input arguments for each OpDesc
auto& input_field = op.inputs();
for (auto& input : input_field) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

input -> input_var? Because the first argument of this function is named input.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

paddle/framework/prune.cc Outdated
}
// add VarDescs of all output arguments for each OpDesc
auto& output_field = op.outputs();
for (auto& output : output_field) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

output -> output_var? Because the second argument of this function is named output.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

Xreki
Xreki approved these changes Jan 29, 2018
View reviewed changes
Copy link
Contributor

@Xreki Xreki left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

paddle/framework/prune.cc
auto* var_field = output->mutable_blocks(block_id)->mutable_vars();
for (const auto& var : *var_field) {
var_map[var.name()] = var;
}
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

For this prune function, the input is proto::ProgramDesc, so we only have limited functionality provided by the protobuf library itself.

You are right. Sorry for didn't notice that.

@kexinzhao kexinzhao merged commit 3f2aa98 into PaddlePaddle:develop Jan 29, 2018
@kexinzhao kexinzhao deleted the remove_unreferenced_var branch April 4, 2018 02:13
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@Xreki Xreki Xreki approved these changes

@sidgoyal78 sidgoyal78 Awaiting requested review from sidgoyal78

@tonyyang-svail tonyyang-svail Awaiting requested review from tonyyang-svail

Assignees
No one assigned
Labels
预测 原名Inference，包含Capi预测问题等
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@kexinzhao @Xreki