Fix #565 #582
Fix #565 #582
Conversation
| @@ -432,13 +432,15 @@ def _load_conv_params(conv, state_dict_tv, module_name_tv, | |||
| """ | |||
|
|
|||
| weight_tv_name = module_name_tv + '.weight' | |||
| conv.weight.data.copy_(state_dict_tv[weight_tv_name]) | |||
| loaded_param_names.append(weight_tv_name) | |||
| if conv.weight.data.shape == state_dict_tv[weight_tv_name].shape: | |||
There was a problem hiding this comment.
do we need to issue some warning if the shapes are mismatched?
There was a problem hiding this comment.
I think no warning is OK, since:
- The shape mismatch may only happen when in_channels != 3. In this case, the shape mismatch is not unexpected. Users assume that pre-train weights are for other modules (except conv1).
- There is already a warning message in
_load_torchvision_checkpoint, which indicates that parameters ofconv1andbn1are not loaded.
There was a problem hiding this comment.
Its okay if pytorch could warn the user. As long as there are warnings, I'm ok with the change.
Codecov Report
@@ Coverage Diff @@
## master #582 +/- ##
==========================================
- Coverage 84.17% 84.13% -0.04%
==========================================
Files 121 121
Lines 8613 8617 +4
Branches 1427 1431 +4
==========================================
Hits 7250 7250
- Misses 1014 1015 +1
- Partials 349 352 +3
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
|
Changes are not covered by tests |
I think it is covered: I add a testing case in which the ResNet is initialized with |
|
Then there must be tracking issues in Codcov |
|
In our unit test refactor doc, |
Sorry I didn't see it before, I will reorganize related unittests. |
No description provided.