[systemml] branch master updated: [SYSTEMDS-263] Initial design ONNX graph importer

[mb...@apache.org]

This is an automated email from the ASF dual-hosted git repository.

mboehm7 pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/systemml.git


The following commit(s) were added to refs/heads/master by this push:
     new 0dae427  [SYSTEMDS-263] Initial design ONNX graph importer
0dae427 is described below

commit 0dae42705f91b00abc03be09d810b3a9286338c5
Author: Lukas Timpl <lu...@student.tugraz.at>
AuthorDate: Sun Apr 12 20:43:47 2020 +0200

    [SYSTEMDS-263] Initial design ONNX graph importer
    
    Since ONNX does support conditional operators (loop, if), I've tailored
    the design towards a command-line tool that generates a DML script as
    discussed.
    
    AMLS project SS2020.
    Closes #885.
---
 docs/Tasks.txt               |  1 +
 docs/onnx-systemds-design.md | 46 ++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 47 insertions(+)

diff --git a/docs/Tasks.txt b/docs/Tasks.txt
index cfcab1a..316971d 100644
--- a/docs/Tasks.txt
+++ b/docs/Tasks.txt
@@ -205,6 +205,7 @@ SYSTEMDS-250 Extended Slice Finding
 SYSTEMDS-260 Misc Tools
  * 261 Stable marriage algorithm                                      OK
  * 262 Data augmentation tool for data cleaning                       OK
+ * 263 ONNX graph importer/exporter
 
 SYSTEMDS-270 Compressed Matrix Blocks
  * 271 Reintroduce compressed matrix blocks from SystemML             OK
diff --git a/docs/onnx-systemds-design.md b/docs/onnx-systemds-design.md
new file mode 100644
index 0000000..9650f9c
--- /dev/null
+++ b/docs/onnx-systemds-design.md
@@ -0,0 +1,46 @@
+# onnx-systemds
+
+A tool for importing/exporting [ONNX](https://github.com/onnx/onnx/blob/master/docs/IR.md) graphs into/from SystemDS DML scripts.
+
+
+## Goals
+
+* Support for importing [operators of the ONNX base definition](https://github.com/onnx/onnx/blob/master/docs/Operators.md)
+
+* Support for importing [operators defined by ONNX-ML](https://github.com/onnx/onnx/blob/master/docs/Operators-ml.md)
+
+* Support for exporting DML script to ONNX graphs
+
+## Limitations
+
+* Not able to support all data types / operators as they are not currently supported by SystemDS
+
+
+
+## Suggested Implementation
+
+Since the ONNX specification includes the conditional operators [loop](https://github.com/onnx/onnx/blob/master/docs/Operators.md#Loop) and [if](https://github.com/onnx/onnx/blob/master/docs/Operators.md#If), a direct conversion from ONNX to the internal HOP might not be ideal. 
+
+Hence my suggested implementation is a dedicated tool invoked from command line which generates DML scripts. This also enables optimizations performed by the compiler at both graph and program level.
+
+### Example Call
+
+```bash
+onnx-systemds model.onx --out model_script.dml
+```
+
+
+### Tooling
+
+* Due to the availability of a [Python API](https://github.com/onnx/onnx/blob/master/docs/PythonAPIOverview.md) for ONNX, I would suggest implementing the tool in Python
+* Another advantage of Python is good support for template engines e.g. [Jinja](https://jinja.palletsprojects.com/en/2.11.x/)
+* An implementation could use templates for various operators which are then combined into a script
+
+### Implementation Details
+
+ONNX is a [serialized graph](https://github.com/onnx/onnx/blob/master/docs/IR.md#graphs) structured as a sorted list of nodes that form a DAG (directed acyclic graph).
+
+1. Loading in the serialized structure
+2. [Checking](https://github.com/onnx/onnx/blob/master/docs/PythonAPIOverview.md#checking-an-onnx-model) model and [converting](https://github.com/onnx/onnx/blob/master/docs/PythonAPIOverview.md#converting-version-of-an-onnx-model-within-default-domain-aionnx) models to a common version
+3. Building a simple internal graph structure (for arbitrary operators)
+4. Generating the DML script while traversing this graph (provided information in doc_strings and other description variables are added as comments to improve human-readability of the generated script)