Fix ONNXRT calibration for Dml EP

[yuwenzho]

Type of Change

bug fix
API changed or not: no

Description

Fix ONNXRT calibration for Dml EP: calibrate initializer tensors and output tensors separately.

The bug is found in microsoft/Olive#852

How has this PR been tested?

CI

Dependency Change?

no

[yuwenzho]

extension test on Dml EP of ssd_mobilenet_v1, shufflenet-v2-12.

Acc regression of resnet50-v1-12 in CI model test is reasonable: The graph optimization level here is changed from 'ORT_ENABLE_ALL' (default) to 'ORT_DISABLE_ALL', which will casuse the intermediate output to change slightly.

[chensuyue]

extension test on Dml EP of ssd_mobilenet_v1, shufflenet-v2-12.

Acc regression of resnet50-v1-12 in CI model test is reasonable: The graph optimization level here is changed from 'ORT_ENABLE_ALL' (default) to 'ORT_DISABLE_ALL', which will casuse the intermediate output to change slightly.

Ok, then we need to update model refer before merge this PR. But why need to disable the graph optimization?

[yuwenzho]

extension test on Dml EP of ssd_mobilenet_v1, shufflenet-v2-12.
Acc regression of resnet50-v1-12 in CI model test is reasonable: The graph optimization level here is changed from 'ORT_ENABLE_ALL' (default) to 'ORT_DISABLE_ALL', which will casuse the intermediate output to change slightly.

Ok, then we need to update model refer before merge this PR. But why need to disable the graph optimization?

The graph optimization is disabled during calibration, becasue:

1. The model has been graph optimized before, so there is no need to optimize further during the calibration phase.
2. If the EP is a Dml EP and further graph optimization is performed, the model intermediate outputs required for calibration cannot be collected correctly.
3. Refer to the session creation func in onnxruntime calibration.