Conversation
tools/pytorch2onnx.py
Outdated
|
|
||
|
|
||
| def _update_input_img(img_list, img_meta_list): | ||
| def _update_input_img(img_list, img_meta_list, update_origin=False): |
There was a problem hiding this comment.
What does update_origin means?
There was a problem hiding this comment.
ori_shape in img_meta is used to rescale output to origin shape.
_update_input_img will update img_meta_list according to img_list. And this flag is used to control if the segmentator needs to rescale according to the input tensor shape.
| # update img_meta | ||
| img_list, img_meta_list = _update_input_img( | ||
| img_list, img_meta_list) | ||
| img_list, img_meta_list, test_mode == 'whole') |
There was a problem hiding this comment.
Why arg is test_mode == 'whole'?
There was a problem hiding this comment.
The input tensor has been interpolated with scale_factor=1.5 to test dynamic shape support.
We only support dynamic shape on these whole mode models.
Codecov Report
@@ Coverage Diff @@
## master #547 +/- ##
==========================================
- Coverage 86.79% 86.76% -0.04%
==========================================
Files 101 101
Lines 5249 5250 +1
Branches 848 849 +1
==========================================
- Hits 4556 4555 -1
- Misses 533 535 +2
Partials 160 160
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
* fix verify bugs * rename args
* resolve comments * update changelog * init stdet_demo * frame extraction & human det * update code * rename label_map.txt as k400_label_map.txt * finish demo * after check * resolve comments & + docstring
Hi
This PR fixes
shape mismatchwhen verifying thewhole modemodel inpytorch2onnx.