Fix pipeline serialization with make_continious inside#848
Fix pipeline serialization with make_continious inside#848JanuszL merged 1 commit intoNVIDIA:masterfrom JanuszL:fix_make_continues_serial
Conversation
|
build! |
|
CI MESSAGE: [726719]: BUILD STARTED |
|
The last change will work only with NVIDIA/DALI_extra#10. |
|
build! |
|
CI MESSAGE: [726719]: BUILD FAILED |
Kh4L
left a comment
Kh4L
left a comment
There was a problem hiding this comment.
What happens if user try to ˋbuilda pipeline afterserialize`?
dali/pipeline/pipeline.h
Outdated
There was a problem hiding this comment.
| * for_serialization tells if skip some graph processing not that is not neccesary | |
| * for_serialization tells if skip some graph processing that is not neccesary |
It won't work, other solution is to just serialize everything, and I will try to do that to cover that case too. |
dali/pipeline/pipeline.cc
Outdated
There was a problem hiding this comment.
Be consistent with the &&. Either at the beginning or at the end of the line
dali/test/python/test_pipeline.py
Outdated
There was a problem hiding this comment.
how about testing that both pipelines produce the same outputs for the same seed?
There was a problem hiding this comment.
Done in next test case. In that one I want to have pipe deleted before creating new one.
|
build! |
|
CI MESSAGE: [727852]: BUILD STARTED |
|
CI MESSAGE: [727852]: BUILD FAILED |
- when DALI pipeline is serialized make_continious nodes are skipped (inserting such node into graph updates names of the following operator input to consume the output from make_continious), while updated input names of the following operators are not reverted back to the original name. In effect when the pipeline is deserialized input cannot match outputs in the graph - stores separate, unmodified graph description only for the serialization purpose Signed-off-by: Janusz Lisiecki <jlisiecki@nvidia.com>
|
CI MESSAGE: [727979]: BUILD STARTED |
|
CI MESSAGE: [727979]: BUILD FAILED |
|
CI MESSAGE: [728007]: BUILD STARTED |
|
CI MESSAGE: [728007]: BUILD PASSED |
- when DALI pipeline is serialized make_continious nodes are skipped (inserting such node into graph updates names of the following operator input to consume the output from make_continious), while updated input names of the following operators are not reverted back to the original name. In effect when the pipeline is deserialized input cannot match outputs in the graph - stores separate, unmodified graph description only for the serialization purpose Signed-off-by: Janusz Lisiecki <jlisiecki@nvidia.com> Signed-off-by: kychennv <kychen@nvidia.com>
- when DALI pipeline is serialized make_continious nodes are skipped (inserting such node into graph updates names of the following operator input to consume the output from make_continious), while updated input names of the following operators are not reverted back to the original name. In effect when the pipeline is deserialized input cannot match outputs in the graph - stores separate, unmodified graph description only for the serialization purpose Signed-off-by: Janusz Lisiecki <jlisiecki@nvidia.com>
5 participants
skipped (inserting such node into graph updates names of the
following operator input to consume the output from make_continious),
while updated input names of following operators are not
reverted back to the original name. In effect when the pipeline
is deserialized input cannot match outputs in the graph
Signed-off-by: Janusz Lisiecki jlisiecki@nvidia.com