Add an ability to return a duplicated outputs from the DALI pipeline #1556
Conversation
|
CI MESSAGE: [1024038]: BUILD STARTED |
dali/python/nvidia/dali/pipeline.py
Outdated
| for name in outputs: | ||
| idx = output_map.get(name, next_free_idx + 1) | ||
| if idx > next_free_idx: | ||
| output_map[name] = next_free_idx | ||
| idx = next_free_idx | ||
| next_free_idx += 1 | ||
| unique_outputs.append(name) | ||
| self._outputs_list.append(idx) |
There was a problem hiding this comment.
I wanted to suggest something that is clearer when indexing, but it does lookup the dictionary twice.
for name in outputs:
if name not in output_map:
output_map[name] = next_free_idx
next_free_idx +=1
unique_outputs.append(name)
self._outputs_list.append(output_map[name])
There was a problem hiding this comment.
It is not perf critical. Done.
dali/python/nvidia/dali/pipeline.py
Outdated
| pipeline_tls = tls() | ||
|
|
||
| def show_deprecation_warning(deprecated, in_favor_of): |
There was a problem hiding this comment.
| def show_deprecation_warning(deprecated, in_favor_of): | |
| def _show_deprecation_warning(deprecated, in_favor_of): |
|
CI MESSAGE: [1024053]: BUILD STARTED |
|
CI MESSAGE: [1024038]: BUILD FAILED |
|
CI MESSAGE: [1024053]: BUILD FAILED |
|
CI MESSAGE: [1024285]: BUILD STARTED |
|
CI MESSAGE: [1024285]: BUILD FAILED |
|
CI MESSAGE: [1024302]: BUILD STARTED |
|
CI MESSAGE: [1024302]: BUILD FAILED |
|
CI MESSAGE: [1024321]: BUILD STARTED |
|
CI MESSAGE: [1024323]: BUILD STARTED |
|
CI MESSAGE: [1024321]: BUILD FAILED |
|
CI MESSAGE: [1024321]: BUILD PASSED |
|
CI MESSAGE: [1024323]: BUILD PASSED |
dali/pipeline/pipeline.cc
Outdated
| .AddOutput("contiguous_" + name, "cpu"); | ||
| PrepareOpSpec(&spec, GetNextInternalLogicalId()); | ||
|
|
||
| graph_.AddOp(spec, "__MakeContiguous_" + name); |
There was a problem hiding this comment.
| graph_.AddOp(spec, "__MakeContiguous_" + name); | |
| graph_.AddOp(spec, "__MakeContiguous_" + name); | |
| it->second.has_contiguous = true; |
This should solve the issue without the need for an extra unordererd_set - and I think it's missing anyway.
|
CI MESSAGE: [1029963]: BUILD STARTED |
|
CI MESSAGE: [1030120]: BUILD STARTED |
|
CI MESSAGE: [1030120]: BUILD FAILED |
|
CI MESSAGE: [1029963]: BUILD FAILED |
- adds a code that prevents adding "MakeContiguous" multiple times to the output - adds tests for CPU, GPU and Mixed operators outputs that duplicates - some small clean up in used API checker in the Pipeline Signed-off-by: Janusz Lisiecki <jlisiecki@nvidia.com>
|
CI MESSAGE: [1030348]: BUILD STARTED |
|
CI MESSAGE: [1030348]: BUILD PASSED |
Signed-off-by: Janusz Lisiecki jlisiecki@nvidia.com
Why we need this PR?
What happened in this PR?
JIRA TASK: [475]