Add debug mode tutorial notebook

[ksztenderski]

Signed-off-by: ksztenderski ksztenderski@nvidia.com

Category:

Other (e.g. Documentation, Tests, Configuration)

Description:

Add tutorial jupyter notebook showing use cases and features of debug mode.

Additional information:

Affected modules and functionalities:

Key points relevant for the review:

Checklist

Tests

• Existing tests apply
• New tests added
  • Python tests
  • GTests
  • Benchmark
  • Other
• N/A

Documentation

• Existing documentation applies
• Documentation updated
  • Docstring
  • Doxygen
  • RST
  • Jupyter
  • Other
• N/A

DALI team only

Requirements

• Implements new requirements
• Affects existing requirements
• N/A

REQ IDs: N/A

JIRA TASK: N/A

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[ksztenderski]

!build

[dali-automaton]

CI MESSAGE: [3840125]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [3840132]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [3840166]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [3840166]: BUILD FAILED

[ksztenderski]

!build

[dali-automaton]

CI MESSAGE: [3846256]: BUILD STARTED

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

klecki commented on 2022-01-28T11:21:48Z
----------------------------------------------------------------

Maybe we can link to the DataNode doc: https://docs.nvidia.com/deeplearning/dali/user-guide/docs/pipeline.html#datanode

And/or mention the fact that we are talking about the outputs of operators? I guess not many people know what the DataNode is in DALI.

ksztenderski commented on 2022-01-28T17:29:36Z
----------------------------------------------------------------

done

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

klecki commented on 2022-01-28T11:21:49Z
----------------------------------------------------------------

Maybe we should add one sentence about why we need to change to the experimental.pipeline_def, or that apart from allowing us to turn on the debug mode it is basically the same as the old one.

Also I would add a filler sentence like:

Let's take a simple pipeline that reads images and rotates them and try out the debug mode on it.

ksztenderski commented on 2022-01-28T17:29:53Z
----------------------------------------------------------------

done

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

klecki commented on 2022-01-28T11:21:50Z
----------------------------------------------------------------

Line #2.    def simple_pipeline():

Maybe

def simple_pipeline_with_print

?

ksztenderski commented on 2022-01-28T17:31:28Z
----------------------------------------------------------------

done

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

klecki commented on 2022-01-28T11:21:51Z
----------------------------------------------------------------

I think that is worth to summarize that there would be difference in running that pipeline in debug and non-debug mode:

• in debug mode the print happens in every run()
• in regular mode it happens during build() only once

Maybe it is worth to highlight the fact that in debug mode, we can see the operators executed one by one in Python, when we call run() as opposed to scheduling the execution of the graph in backend.

ksztenderski commented on 2022-01-28T17:37:03Z
----------------------------------------------------------------

done

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

klecki commented on 2022-01-28T11:21:52Z
----------------------------------------------------------------

Should we say that it gives the access to the TensorListCPU/GPU for the current iteration?

ksztenderski commented on 2022-01-28T17:37:10Z
----------------------------------------------------------------

done

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

klecki commented on 2022-01-28T11:21:53Z
----------------------------------------------------------------

I would maybe expand it a bit, and say how we can modify the data? And what format do they expect?

ksztenderski commented on 2022-01-28T17:37:35Z
----------------------------------------------------------------

done

[klecki]

Also I would add a section about limitations:

• mention the speed impact,
• mention the seeds - it can have different results than regular pipeline,
• mention the execution order
• anything else?

We should also clearly state somewhere, that those operators work only in a scope of the pipeline_def and one cannot use them in numpy style for eager execution.

[dali-automaton]

CI MESSAGE: [3846256]: BUILD PASSED

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

JanuszL commented on 2022-01-28T13:45:27Z
----------------------------------------------------------------

I think we should stress out the performance implication.

> This allows for better performanc

Sound like it the graph mode is just a bit faster. I would say that in this mode the execution is expected to be slow, is not intended to be used in production code nor for benchmarking. Maybe you can even write that it is equivalent to creating a separate pipeline for each operator and running it synchronously one by one.

[ksztenderski]

done

---

View entire conversation on ReviewNB

[ksztenderski]

done

---

View entire conversation on ReviewNB

[ksztenderski]

done

---

View entire conversation on ReviewNB

[ksztenderski]

done

---

View entire conversation on ReviewNB

[ksztenderski]

done

---

View entire conversation on ReviewNB

[ksztenderski]

done

---

View entire conversation on ReviewNB

[ksztenderski]

!build

[dali-automaton]

CI MESSAGE: [3848351]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [3848351]: BUILD FAILED

[ksztenderski]

!build

[dali-automaton]

CI MESSAGE: [3857453]: BUILD STARTED

[ksztenderski]

!build

[dali-automaton]

CI MESSAGE: [3924231]: BUILD STARTED

[JanuszL]

Maybe in the first example it can stay cpu and add mixed in the simple_pipeline_with_print with a note in the text that it works for GPU as well.

---

View entire conversation on ReviewNB

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

JanuszL commented on 2022-02-09T10:34:52Z
----------------------------------------------------------------

I think you should better print images not jpegs here.

[ksztenderski]

Done, but I'm guessing that you might also wanted it to show print on TensorListGPU, but I intentionally printed jpegs here so it would take less space.

---

View entire conversation on ReviewNB

[JanuszL]

Up to you.

---

View entire conversation on ReviewNB

[dali-automaton]

CI MESSAGE: [3924231]: BUILD FAILED

[ksztenderski]

!build

[dali-automaton]

CI MESSAGE: [3926081]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [3926081]: BUILD FAILED

[dali-automaton]

CI MESSAGE: [3926081]: BUILD PASSED