Rewrite memory manager log_info in Python.

Context

This MR fix !276 (merged)

It began by looking at the memory management of the dinov2 model

Realizing that it is not readable, I rewrited it in python in order to easily tweak it.

Here is the new memory info after rewritting it:

Add display_names parameter to aidge_core.generate_optimized_memory_info which allows to diplay the names or not:

Benefice

This MR remove a dependencie to gnuplot

Log function is easy to use now!

Miscellaneous

After using MetaOperators to simplify the graph we get the follwing memory placement strategy which surprised me as it produces a lower memory peak!

This make sens since by fusing operation we load in RAM data and do not need to save intermediate output.

A little sad to not have found this out before webinar

For reproductibility the code used:

import aidge_core
import aidge_backend_cpu
import aidge_onnx

dinov2_model = aidge_onnx.load_onnx("hf_dinov2_sim.onnx")

aidge_core.fuse_to_metaops(dinov2_model, "MatMul-*>Add", "Linear")

aidge_core.fuse_to_metaops(dinov2_model, "ReduceMean-*>Sub#1~>(Pow#1->ReduceMean-*>Add#1->Sqrt)-*>Div#1-*>Mul#1-*>Add#2;"
                                        "Sub#1~*>Div#1;"
                                        "Pow#1<1~Producer;"
                                        "Add#1<*~Producer;"
                                        "Mul#1<*~Producer;"
                                        "Add#2<*~Producer;"
                                        "Sub#1~>$", "LayerNorm")

aidge_core.fuse_to_metaops(dinov2_model, "MatMul->Div#1->Softmax-*>MatMul;"
                                        "Div#1<1~Producer", "ScaledDotProductAttention")

aidge_core.fuse_to_metaops(dinov2_model, "ScaledDotProductAttention#1->Transpose->Reshape#1->Linear;"
                                        "Reshape#1<1~Producer;"
                                        "ScaledDotProductAttention#1<0-(Transpose<-Reshape#2<-Add#1);"
                                        "ScaledDotProductAttention#1<1-(Transpose<-Reshape#3<-Add#2);"
                                        "ScaledDotProductAttention#1<2-(Transpose<-Reshape#4<-Add#3);"
                                        "Reshape#2<1~Producer;"
                                        "Add#1<*-0-Split#1;"
                                        "Add#2<*-1-Split#1;"
                                        "Add#3<*-2-Split#1;"
                                        "Split#1<-MatMul;"
                                        "Split#1<1~Producer", "MultiHeadAttention")

aidge_core.fuse_to_metaops(dinov2_model, "Div#1->Erf->Add#1-*>Mul->Mul#2;"
                                        "Div#1<1~Producer;"
                                        "Add#1<*~Producer;"
                                        "Mul#2<*~Producer", "GeLU")
dinov2_model.set_ordered_outputs([dinov2_model.get_ordered_outputs()[0][0].inputs()[0], dinov2_model.get_ordered_outputs()[0]])
dinov2_model.set_backend("cpu")
dinov2_model.set_datatype(aidge_core.dtype.float32)
dinov2_model.forward_dims([[1,3,224,224]], True)

s = aidge_core.SequentialScheduler(dinov2_model)
s.generate_scheduling()
aidge_core.generate_optimized_memory_info(s, "stats_dyno", wrapping=False, display_names=False)

## Major modifications

- enhance structure: rewrite `generate_optimized_memory_info`
- add `display_names` parameter to `generate_optimized_memory_info` which allows to diplay the names or not
- add matplotlib dependence for Python package i `pyproject.toml`
- remove a dependency to gnuplot

requested review from @olivierbichler

assigned to @cmoineau

@idealbuq This MR may interest you since you work on memory mapping

@olivierbichler I did not fully reproduced the log method as I don't handle memory wrapping.

However, to be honnest I don't understand what the code represent ...

I commented the previous code but I think that I will completely remove the log function from CPP if we were to merge this.

Is it this code that you did not reproduce?

                fmt::print(gnuplot.get(), "set arrow from {},{} to {},{} nohead\n",
                    startX, (contiguousOffset / 1024.0),
                    (startX + 0.1), (contiguousOffset / 1024.0));

                fmt::print(gnuplot.get(), "set arrow from {},{} to {},{} nohead\n",
                    (startX + 0.05), ((contiguousOffset + contiguousSize) / 1024.0),
                    (startX + 0.05), (wrappedOffset / 1024.0));

Yes it is.

It is not critical. It actually shows where the wrapping is by displaying an arrow from the end of the first section to the beginning of the wrapped section. It was not that clear anyway, there is probably a clever way to materialize that in the graph...

added 1 commit

• 55ce8814 - [Min] MemManager plot, remove axis and center more the fig.

Compare with previous version

Tweaked a little bit for improving readability

marked this merge request as draft

added PriorityLow Refactoring🎨 StatusWork in Progress labels

mentioned in issue #211 (closed)

changed the description

added 1 commit

• 4e662abd - Add matplotlib as dependance.

Compare with previous version

added 1 commit

• ae972675 - Add matplotlib as dependance.

Compare with previous version

requested review from @idealbuq

@idealbuq @olivierbichler I will remove gnuplot dependencies and push this once @pineapple merged the 0.4

mentioned in merge request aidge!86 (merged)

added 2 commits

• 01f65937 - Fix error with meminfo save fig.
• 1be99fe8 - Remove MemoryManager.log() method.

Compare with previous version

marked this merge request as ready

added 93 commits

• 1be99fe8...a983179d - 88 commits from branch dev
• 59bffdf1 - Rewrite memory manager log_info in Python.
• 80ea57b4 - [Min] MemManager plot, remove axis and center more the fig.
• 4f0ceb4d - Add matplotlib as dependance.
• 28470824 - Fix error with meminfo save fig.
• 61e9bf2a - Remove MemoryManager.log() method.

Compare with previous version

enabled an automatic merge when all merge checks for 61e9bf2a pass

aborted the automatic merge because the source branch was updated. Learn more.