Error with Gen Settings in Python
zd819 opened this issue · comments
Describe the bug
Running Gen_Settings function on some of my ML models produces the following errors:
CIN Model (Attached)
Failed to generate settings: Failed analyse for node #8 "/Mul" Mul
IIN Model (Attached)
Failed to generate settings: Failed analyse for node #48 "/target_attention.0/Unsqueeze" Unsqueeze13
LIN Model (Attached)
Failed to generate settings: Failed analyse for node #16 "/Squeeze" Squeeze13
Expected behaviors
Gen_Settings to run correctly and produce a settings file
Steps to reproduce the bug
- Load Models and Corresponding Input file
- Generate and calibrate settings using the attached onnx and input.json files
Device and Operating System
OS: Windows 10 Home
Version: 10.0.19045 Build 19045
Processor: Intel(R) Core(TM) i5-1035G1 CPU
Memory: 20 GB
Additional Information
Python 3.10.4
Ezkl 1.7.19
torch 1.13.1
onnx 1.14.0
Files：
models.zip
inputs.zip
@zd819 before I file an issue -- can you see if this issue persists when exporting the model with Onnx versions 10 through 12 ?
@zd819 before I file an issue -- can you see if this issue persists when exporting the model with Onnx versions 10 through 12 ?
we have verified that this issue persists with onnx version 10.
Opened a separate issue for IIN Model.
sonos/tract#1156
The LIN model isn't really properly specified from what I can tell. It mixes variable batch_size with a specified batch_size of 20. It failing seems like something we would want tbh
Thanks for all the focus and quick fixes Alex and the ezkl team. I am testing the new changes on the CIN Model, and for the LIN model we have changed it since. Will run tests and reply with results, but as you've tested should mostly be fixed for CIN.
Will await the IIN fix
Seems like IIN also has inconsistent shapes, see here: sonos/tract#1156 (comment) . Will need to resolve those to produce a coherent / consistent graph.
Will close this as we resolved issues with CIN and LIN