Export WordEmbeddingsTransform to ONNX #1249
Conversation
Using placeholders for ONNX initiliazer nodes
|
|
vaeksare
changed the title
vaeksare
changed the title
| // | | | | ||
| // J[j] K[j] L[j] | ||
| // | | | | ||
| // --------------------Concat (axis = 1) --------------------------- |
There was a problem hiding this comment.
``suggestion
// '-------------------Concat (axis = 1) ---------------------------'
There was a problem hiding this comment.
There was a problem hiding this comment.
Looks good.
@vaeksare : Would it be worth writing a unit test which checks output equality between the WordEmbedding transform using ML.NET & its ONNX export? This will ensure the ONNX export stays in sync w/ the behavior of the transform in the future.
Currently, I think you are testing that the ONNX output is the expected value, but I'm concerned that when we do checkins which change the test values, we simply blindly update the expected test outputs without checking for equality between the ML.NET model & its ONNX conversion.
Possible cases (same type you manually checked): #1249 (comment)
| // / / | | \ \ | ||
| // | Cast (to = int64) | Cast (to = float) Not | | ||
| // | | | | | | | ||
| // '------------ Add -----------' | Scale (scale = 2.0) Cast (to = int32) | |
There was a problem hiding this comment.
Missing variable names between operators! Operators are never connected directly. There must be some variables between them. Please also make sure those names are consistent to your comments for your code. Thanks. #Resolved
There was a problem hiding this comment.
|
@justinormont This was something I considered and talked to other people about initially. The issue that occurs is that our OnnxTransform (the only real way we have to run ONNX models directly) currently only works on Windows. As such, these tests would have to be made to only runs on Windows builds, which kind of defeats the purpose and means we still need the original ones. When the ability to run ONNX models is added to all platforms, this is likely something we should look into. But at the moment, I don't believe investing time into this is worth it. |
Co-Authored-By: vaeksare <vaeksare@microsoft.com>
Co-Authored-By: vaeksare <vaeksare@microsoft.com>
Co-Authored-By: vaeksare <vaeksare@microsoft.com>
Co-Authored-By: vaeksare <vaeksare@microsoft.com>
Co-Authored-By: vaeksare <vaeksare@microsoft.com>
Co-Authored-By: vaeksare <vaeksare@microsoft.com>
Co-Authored-By: vaeksare <vaeksare@microsoft.com>
Co-Authored-By: vaeksare <vaeksare@microsoft.com>
…inelearning into savewordembeddings
* add sweepable api example * fix tests * add grid search tuner * use grid search
* add sweepable api example * fix tests * add grid search tuner * use grid search
ghost
locked as resolved and limited conversation to collaborators
Implements the ability to export the WordEmbeddingsTransform by converting it to an ONNX model, as well as expanding the functionality of some existing structures to allow for more efficient conversion implementation. The detailed conversion strategy can be found in comments inline. Fixes #1248
Testing was done through running the model on the same input using ML.NET and Lotus runtime directly using python bindings, producing the same results. The verified ONNX model (run in Lotus to check results) saved in Json format is used as a baseline for the formal tests.
The resulting ONNX model looks as follows:
