Multi-GPU Training Support #42
Conversation
☂️ Python Coverage
Overall Coverage
New FilesNo new covered files... Modified Files
|
luxonis_train/nodes/efficientnet.py
Outdated
| x = self.backbone.conv_head(x) | ||
| x = self.backbone.bn2(x) | ||
| x = self.backbone.act2(x) | ||
| x = self.backbone.global_pool(x) |
There was a problem hiding this comment.
Is there a reason we append flattened features here instead of before global pool?
It seems also we have a list of size 5?
There was a problem hiding this comment.
Whereas elsewhere it's a list of size 4?
There was a problem hiding this comment.
Thanks, I think it should be added before global pool.
Regarding the list size of 4, would it make sense to remove one of the out_indices for example 2, so we have a list size of 4?
There was a problem hiding this comment.
I think maybe we check the paper if they define "blocks" to find out which features we should take?
Test Results 4 files 4 suites 1h 0m 37s ⏱️ Results for commit 267936b. |
kozlov721
changed the title
|
@kozlov721 Do we require anywhere downstream that backbone outputs exactly 4 feature maps? |
No, I don't think so |
|
How would neck accept this then, the last 3 feature maps for example? @kozlov721 |
The |
|
Okay, I think we can merge this then. |
Co-authored-by: GitHub Actions <[email protected]>
ddp_find_unused_parameters_trueerror, which occurs when some backbone components are not utilized in the forward pass.