Conversation
|
The build failure occurs because it doesn't have the newer batchnorm yet. |
|
I guess with mlpack/mlpack#2474 merged, this should build now? Let's rerun the tests. |
Yes, it builds now. |
|
Can you take a look at why the style build is failing? |
|
Sure, I will resolve it with the next commit. Thanks for pointing it out. |
|
Hey @kartikdutt18 Can you rebase this PR? |
models/yolo/yolo_impl.hpp
Outdated
| if (weights == "voc") | ||
| { | ||
| // Download weights here. | ||
| LoadModel("./../weights/YOLO/yolo" + YOLOVersion + "_voc.bin"); |
There was a problem hiding this comment.
Please let us know if the weight file is ready. Do you intend to use the weight converter?
There was a problem hiding this comment.
Yes I do intend to use the weights. I am done with the implementation of detector too, It makes some extra copies which can be avoided which I'll remove and then I'll leave a comment here when everything is ready.
There was a problem hiding this comment.
I'm done with the detector part as well and I'm moving forward with testing the model. If everything works well, I'll add the YOLO model in converter repo for you to test it and then we can merge this PR.
There was a problem hiding this comment.
I have added a simple tests in this PR. I haven't added tests for weights because of mlpack/mlpack#2561. I think we can merge this PR since the model is working and any changes would be internal implementation in mlpack. To see if the model is working correctly you can run the following commands :
git clone https://github.com/kartikdutt18/mlpack-PyTorch-Weight-Translator.git
cd mlpack-PyTorch-Weight-Translator
./run.sh yolov1_tiny
With this you will see the model weights being created and transferred into the YOLO model and being used for prediction. You can see that the iou of predictions made by PyTorch model and YOLO model are same i.e. 1.0. Kindly let me know if you face any issues. If this makes sense, I think we can merge this PR and then later add a PR with tests for weights once mlpack/mlpack#2561 is resolved.
(Similar comment from DarkNet).
There was a problem hiding this comment.
@kartikdutt18 Rather than creating a weight file, how about save it to the weight directory for making it ready-for-use?
There was a problem hiding this comment.
This would face the same problem as DarkNet since this also uses BatchNorm Layer which has a serialization issue.
Co-authored-by: sy0814k <sy0814k@gmail.com>
Co-authored-by: sy0814k <sy0814k@gmail.com>
KimSangYeon-DGU
left a comment
KimSangYeon-DGU
left a comment
There was a problem hiding this comment.
No further comments from my side.
KimSangYeon-DGU
left a comment
KimSangYeon-DGU
left a comment
There was a problem hiding this comment.
Left some comments about styles
KimSangYeon-DGU
left a comment
KimSangYeon-DGU
left a comment
There was a problem hiding this comment.
No further comments from my side.
![mlpack-bot[bot]](https://avatars.githubusercontent.com/in/23934?s=60&v=4){kind=small}
|
Thanks for another great contribution 👍 |
|
Awesome, Thanks a lot @KimSangYeon-DGU and @saksham189 for all the reviews, suggestions and the help. |
4 participants
Hey everyone,
This PR aims to YOLO class to the models repo. The following is the TO-DO list:
Add PreProcessor Function (Convert DataLoader annotations to YOLO targets). (I'll add this by 10.07.20)
Add Detection stage for the YOLO class (Where loss is calculated, feature maps are converted to understandable format).
Add YOLO model architecture.