[WIP] Improve performance by adding forward indexes. #1035

xadupre · 2020-08-03T11:00:41Z

Add forward index to easily find nodes ingesting a node. Following profile measure the improvment for the conversion of model EfficientNetB2, with and without this PR.

Before:

61.469 profile  conversion_time.py:44
`- 61.469 convert  conversion_time.py:38
   |- 35.363 spy_convert  conversion_time.py:26
   |  |- 34.500 spy_convert_in  conversion_time.py:30
   |  |  `- 34.497 process_tf_graph  tf2onnx/tfonnx.py:334
   |  |     |- 16.260 tensorflow_onnx_mapping  tf2onnx/tfonnx.py:211
   |  |     |  |- 5.029 version_11  tf2onnx/onnx_opset/nn.py:360
   |  |     |  |  `- 5.029 version_1  tf2onnx/onnx_opset/nn.py:325
   |  |     |  |     `- 4.989 conv_convert_inputs  tf2onnx/onnx_opset/nn.py:64
   |  |     |  |        |- 2.890 insert_new_node_on_output  tf2onnx/graph.py:1252
   |  |     |  |        |  `- 2.773 replace_all_inputs  tf2onnx/graph.py:1291
   |  |     |  |        |     `- 1.462 replace_all_inputs  tf2onnx/graph.py:1291
   |  |     |  |        `- 1.747 find_output_consumers  tf2onnx/graph.py:1277
   |  |     |  |           `- 0.920 find_output_consumers  tf2onnx/graph.py:1277
   |  |     |  |              `- 0.634 get_body_graphs  tf2onnx/graph.py:332
   |  |     |  |- 4.203 version_1  tf2onnx/onnx_opset/tensor.py:129
   |  |     |  |  `- 4.193 replace_all_inputs  tf2onnx/graph.py:1291
   |  |     |  |     |- 2.185 replace_all_inputs  tf2onnx/graph.py:1291
   |  |     |  |     |  |- 0.897 [self]  
   |  |     |  |     |  `- 0.796 get_body_graphs  tf2onnx/graph.py:332
   |  |     |  |     |- 0.775 [self]  
   |  |     |  |     `- 0.675 get_body_graphs  tf2onnx/graph.py:332
   |  |     |  |- 2.079 version_1  tf2onnx/onnx_opset/tensor.py:110
   |  |     |  |  `- 2.065 replace_all_inputs  tf2onnx/graph.py:1291
   |  |     |  |     |- 0.792 [self]  
   |  |     |  |     `- 0.755 get_body_graphs  tf2onnx/graph.py:332
   |  |     |  |- 1.845 version_1  tf2onnx/onnx_opset/controlflow.py:420
   |  |     |  |  `- 1.788 wire_if_branch  tf2onnx/onnx_opset/controlflow.py:725
   |  |     |  |     `- 1.721 prefix_graph  tf2onnx/onnx_opset/controlflow.py:795
   |  |     |  |        `- 1.031 replace_all_inputs  tf2onnx/graph.py:1291
   |  |     |  |- 0.775 version_1  tf2onnx/onnx_opset/nn.py:443
   |  |     |  |  `- 0.768 conv_convert_inputs  tf2onnx/onnx_opset/nn.py:64
   |  |     |  |     `- 0.696 insert_new_node_on_output  tf2onnx/graph.py:1252
   |  |     |  |        `- 0.673 replace_all_inputs  tf2onnx/graph.py:1291
   |  |     |  |- 0.746 version_10  tf2onnx/onnx_opset/tensor.py:760
   |  |     |  |  `- 0.744 version_1  tf2onnx/onnx_opset/tensor.py:647
   |  |     |  |     `- 0.710 insert_new_node_on_output  tf2onnx/graph.py:1252
   |  |     |  |        `- 0.687 replace_all_inputs  tf2onnx/graph.py:1291
   |  |     |  |- 0.718 version_1  tf2onnx/onnx_opset/tensor.py:1030
   |  |     |  |  `- 0.680 replace_all_inputs  tf2onnx/graph.py:1291
   |  |     |  `- 0.713 version_1  tf2onnx/onnx_opset/tensor.py:1176
   |  |     |     `- 0.698 insert_new_node_on_output  tf2onnx/graph.py:1252
   |  |     |        `- 0.670 replace_all_inputs  tf2onnx/graph.py:1291
   |  |     |- 6.489 process_tf_graph  tf2onnx/tfonnx.py:334
   |  |     |  |- 3.558 run_rewriters  tf2onnx/tfonnx.py:302
   |  |     |  |  `- 0.979 reset_nodes  tf2onnx/graph.py:665
   |  |     |  |- 1.042 tensorflow_onnx_mapping  tf2onnx/tfonnx.py:211
   |  |     |  `- 0.899 tensorflow_to_onnx  tf2onnx/tf_utils.py:225
   |  |     |     `- 0.897 tflist_to_onnx  tf2onnx/tf_utils.py:132
   |  |     |- 6.242 resolve_functions  tf2onnx/tf_loader.py:417
   |  |     |  |- 3.780 tflist_to_onnx  tf2onnx/tf_utils.py:132
   |  |     |  |  `- 2.336 tf_to_onnx_tensor  tf2onnx/tf_utils.py:53
   |  |     |  |     `- 2.241 get_tf_tensor_data  tf2onnx/tf_utils.py:69
   |  |     |  |        `- 2.060 _array_str_implementation  numpy/core/arrayprint.py:1489
   |  |     |  |              [102 frames hidden]  numpy
   |  |     |  `- 2.188 function_def_to_graph  tensorflow/python/framework/function_def_to_graph.py:33
   |  |     |        [518 frames hidden]  tensorflow, contextlib, threading, fu...
   |  |     |- 3.083 tensorflow_to_onnx  tf2onnx/tf_utils.py:225
   |  |     |  `- 3.083 tflist_to_onnx  tf2onnx/tf_utils.py:132
   |  |     |     `- 2.201 tf_to_onnx_tensor  tf2onnx/tf_utils.py:53
   |  |     |        `- 2.127 get_tf_tensor_data  tf2onnx/tf_utils.py:69
   |  |     |           `- 2.005 _array_str_implementation  numpy/core/arrayprint.py:1489
   |  |     |                 [100 frames hidden]  numpy
   |  |     |- 1.100 run_rewriters  tf2onnx/tfonnx.py:302
   |  |     `- 0.871 update_proto  tf2onnx/graph.py:795
   |  |        `- 0.870 update_proto  tf2onnx/graph.py:344
   |  |           `- 0.703 make_graph  tf2onnx/graph.py:1027
   |  `- 0.863 new_func  tensorflow/python/util/deprecation.py:473
   |        [265 frames hidden]  tensorflow, contextlib, inspect
   `- 26.091 spy_model  conversion_time.py:9
      |- 23.741 EfficientNetB2  tensorflow/python/keras/applications/efficientnet.py:569
      |     [38166 frames hidden]  tensorflow, _collections_abc, six, th...
      `- 1.940 new_func  tensorflow/python/util/deprecation.py:311
            [363 frames hidden]  tensorflow, abc, contextlib, copy, lo...

After:

40.406 profile  conversion_time.py:45
`- 40.406 convert  conversion_time.py:39
   |- 24.676 spy_model  conversion_time.py:10
   |  |- 22.778 EfficientNetB2  tensorflow/python/keras/applications/efficientnet.py:569
   |  |     [35899 frames hidden]  tensorflow, _collections_abc, six, th...
   |  `- 1.508 new_func  tensorflow/python/util/deprecation.py:311
   |        [343 frames hidden]  tensorflow, abc, copy, logging, absl,...
   `- 15.715 spy_convert  conversion_time.py:27
      |- 14.877 spy_convert_in  conversion_time.py:31
      |  `- 14.874 process_tf_graph  tf2onnx/tfonnx.py:334
      |     |- 5.679 process_tf_graph  tf2onnx/tfonnx.py:334
      |     |  |- 3.258 run_rewriters  tf2onnx/tfonnx.py:302
      |     |  |  |- 0.850 reset_nodes  tf2onnx/graph.py:665
      |     |  |  |  `- 0.491 [self]  
      |     |  |  `- 0.424 rewrite_flatten  tf2onnx/rewriter/flatten_rewriter.py:17
      |     |  |     `- 0.409 match_ops  tf2onnx/graph_matcher.py:243
      |     |  |        `- 0.406 match_op  tf2onnx/graph_matcher.py:226
      |     |  |- 0.837 tensorflow_onnx_mapping  tf2onnx/tfonnx.py:211
      |     |  `- 0.743 tensorflow_to_onnx  tf2onnx/tf_utils.py:225
      |     |     `- 0.741 tflist_to_onnx  tf2onnx/tf_utils.py:132
      |     |- 3.993 resolve_functions  tf2onnx/tf_loader.py:417
      |     |  |- 1.949 function_def_to_graph  tensorflow/python/framework/function_def_to_graph.py:33
      |     |  |     [524 frames hidden]  tensorflow, contextlib, threading, fu...
      |     |  `- 1.775 tflist_to_onnx  tf2onnx/tf_utils.py:132
      |     |     `- 0.446 get_tf_node_attr  tf2onnx/tf_utils.py:120
      |     |        `- 0.433 get_attr  tensorflow/python/framework/ops.py:2497
      |     |              [18 frames hidden]  tensorflow, contextlib
      |     |- 2.178 tensorflow_onnx_mapping  tf2onnx/tfonnx.py:211
      |     |  `- 1.492 version_1  tf2onnx/onnx_opset/controlflow.py:420
      |     |     `- 1.440 wire_if_branch  tf2onnx/onnx_opset/controlflow.py:725
      |     |        `- 1.375 prefix_graph  tf2onnx/onnx_opset/controlflow.py:795
      |     |           |- 0.693 replace_all_inputs  tf2onnx/graph.py:1298
      |     |           `- 0.519 make_node  tf2onnx/graph.py:549
      |     |- 1.073 run_rewriters  tf2onnx/tfonnx.py:302
      |     |  `- 0.503 reset_nodes  tf2onnx/graph.py:665
      |     |- 0.806 tensorflow_to_onnx  tf2onnx/tf_utils.py:225
      |     |  `- 0.806 tflist_to_onnx  tf2onnx/tf_utils.py:132
      |     `- 0.754 update_proto  tf2onnx/graph.py:795
      |        `- 0.753 update_proto  tf2onnx/graph.py:344
      |           `- 0.596 make_graph  tf2onnx/graph.py:1027
      `- 0.838 new_func  tensorflow/python/util/deprecation.py:473
            [270 frames hidden]  tensorflow, contextlib, threading, in...

lgtm-com · 2020-08-03T11:11:40Z

This pull request introduces 1 alert when merging 1de2a95 into 36d7413 - view on LGTM.com

new alerts:

1 for Unused import

lgtm-com · 2020-08-03T12:39:57Z

This pull request introduces 1 alert when merging 4167ad1 into 36d7413 - view on LGTM.com

new alerts:

1 for Unused import

lgtm-com · 2020-08-03T14:03:57Z

This pull request introduces 1 alert when merging 6fc785d into 36d7413 - view on LGTM.com

new alerts:

1 for Unused import

lgtm-com · 2020-08-03T15:56:10Z

This pull request introduces 2 alerts when merging 63d12a7 into 36d7413 - view on LGTM.com

new alerts:

1 for Unused local variable
1 for Unused import

lgtm-com · 2020-08-03T16:55:14Z

This pull request introduces 2 alerts when merging b612427 into b4126fb - view on LGTM.com

new alerts:

1 for Unused local variable
1 for Unused import

lgtm-com · 2020-08-04T14:03:32Z

This pull request introduces 2 alerts when merging 57a3d6c into b4126fb - view on LGTM.com

new alerts:

1 for Unused local variable
1 for Unused import

lgtm-com · 2020-08-04T14:18:10Z

This pull request introduces 1 alert when merging 922ce19 into b4126fb - view on LGTM.com

new alerts:

1 for Unused local variable

lgtm-com · 2020-08-04T14:34:40Z

This pull request introduces 1 alert when merging 00091ff into b4126fb - view on LGTM.com

new alerts:

1 for Unused local variable

lgtm-com · 2020-08-04T15:08:51Z

This pull request introduces 1 alert when merging 91fb1c1 into b4126fb - view on LGTM.com

new alerts:

1 for Unused local variable

lgtm-com · 2020-08-04T15:54:13Z

This pull request introduces 1 alert when merging 4949509 into b4126fb - view on LGTM.com

new alerts:

1 for Unused local variable

lgtm-com · 2020-08-04T16:32:21Z

This pull request introduces 1 alert when merging 7047b47 into b4126fb - view on LGTM.com

new alerts:

1 for Unused local variable

lgtm-com · 2020-08-04T17:11:38Z

This pull request introduces 1 alert when merging b81b2f8 into b4126fb - view on LGTM.com

new alerts:

1 for Unused local variable

lgtm-com · 2020-08-04T18:13:29Z

This pull request introduces 1 alert when merging 78a7ac3 into b4126fb - view on LGTM.com

new alerts:

1 for Unused local variable

lgtm-com · 2020-08-04T21:32:35Z

This pull request introduces 1 alert when merging 252fd21 into b4126fb - view on LGTM.com

new alerts:

1 for Unused local variable

… findex

lgtm-com · 2020-08-10T12:17:42Z

This pull request introduces 1 alert when merging 0f6a1cc into c867e52 - view on LGTM.com

new alerts:

1 for Unused import

lgtm-com · 2020-08-10T12:52:49Z

This pull request introduces 1 alert when merging ebff87d into c867e52 - view on LGTM.com

new alerts:

1 for Unused import

… findex

lgtm-com · 2020-08-11T15:28:05Z

This pull request introduces 2 alerts when merging 6570acd into 0110037 - view on LGTM.com

new alerts:

2 for Unused local variable

lgtm-com · 2020-08-11T16:02:35Z

This pull request introduces 2 alerts when merging 1b03b07 into 0110037 - view on LGTM.com

new alerts:

2 for Unused local variable

lgtm-com · 2020-08-12T09:11:53Z

This pull request introduces 2 alerts when merging 785e8aa into 0110037 - view on LGTM.com

new alerts:

2 for Unused local variable

lgtm-com · 2020-08-12T09:53:48Z

This pull request introduces 2 alerts when merging 16b5284 into 0110037 - view on LGTM.com

new alerts:

2 for Unused local variable

… findex

lgtm-com · 2020-08-12T19:46:57Z

This pull request introduces 2 alerts when merging 09119b9 into d3d301a - view on LGTM.com

new alerts:

2 for Unused local variable

lgtm-com · 2020-08-12T21:17:39Z

This pull request introduces 2 alerts when merging ffac86a into d3d301a - view on LGTM.com

new alerts:

2 for Unused local variable

… findex

First step to add forward indexes.

1de2a95

fix failing unit test

4167ad1

Update graph.py

6fc785d

better consistency with forward index

63d12a7

remove unnecessary prints

b612427

xadupre changed the title ~~Improve performance by adding forward indexes.~~ [WIP] Improve performance by adding forward indexes. Aug 4, 2020

specifies positions of replaced input

57a3d6c

pep8

922ce19

lint

00091ff

Update graph.py

91fb1c1

remove verifications

4949509

Update transpose_optimizer.py

7047b47

Update back_to_back_optimizer.py

b81b2f8

Verifies we don't need ops.

78a7ac3

improvments

252fd21

Update graph.py

f0f5fca

sdpython added 2 commits August 10, 2020 13:34

Merge branch 'master' of https://github.com/onnx/tensorflow-onnx into…

3a8bc96

… findex

pylint

0f6a1cc

fix unnecessary modifications

ebff87d

sdpython added 8 commits August 10, 2020 23:11

fix one bug in find_output_consumers

e38ebd5

goes through all nodes

0ac32db

Update graph.py

d5f8811

Update graph.py

da22263

update benchark

2b93937

Merge branch 'master' of https://github.com/onnx/tensorflow-onnx into…

b76fafa

… findex

add a method to modify input index

2be64a6

refactoring

6570acd

add missing line

1b03b07

sdpython added 2 commits August 12, 2020 10:37

Fix a bug when multiple input are removed

b570374

Update tensor.py

785e8aa

sdpython added 2 commits August 12, 2020 11:33

simplifies the API

9f60686

Update tensor.py

16b5284

Merge branch 'master' of https://github.com/onnx/tensorflow-onnx into…

09119b9

… findex

minor updates

ffac86a

sdpython added 3 commits September 4, 2020 11:05

Merge branch 'master' of https://github.com/onnx/tensorflow-onnx into…

09f597a

… findex

Update graph.py

a874730

Update graph.py

794fa8f

xadupre closed this Sep 9, 2020

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[WIP] Improve performance by adding forward indexes. #1035

[WIP] Improve performance by adding forward indexes. #1035

xadupre commented Aug 3, 2020 •

edited

lgtm-com bot commented Aug 3, 2020

lgtm-com bot commented Aug 3, 2020

lgtm-com bot commented Aug 3, 2020

lgtm-com bot commented Aug 3, 2020

lgtm-com bot commented Aug 3, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 10, 2020

lgtm-com bot commented Aug 10, 2020

lgtm-com bot commented Aug 11, 2020

lgtm-com bot commented Aug 11, 2020

lgtm-com bot commented Aug 12, 2020

lgtm-com bot commented Aug 12, 2020

lgtm-com bot commented Aug 12, 2020

lgtm-com bot commented Aug 12, 2020

[WIP] Improve performance by adding forward indexes. #1035

[WIP] Improve performance by adding forward indexes. #1035

Conversation

xadupre commented Aug 3, 2020 • edited

lgtm-com bot commented Aug 3, 2020

lgtm-com bot commented Aug 3, 2020

lgtm-com bot commented Aug 3, 2020

lgtm-com bot commented Aug 3, 2020

lgtm-com bot commented Aug 3, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 4, 2020

lgtm-com bot commented Aug 10, 2020

lgtm-com bot commented Aug 10, 2020

lgtm-com bot commented Aug 11, 2020

lgtm-com bot commented Aug 11, 2020

lgtm-com bot commented Aug 12, 2020

lgtm-com bot commented Aug 12, 2020

lgtm-com bot commented Aug 12, 2020

lgtm-com bot commented Aug 12, 2020

xadupre commented Aug 3, 2020 •

edited