Hello community,
\nI just finished adapting a version of the Faster R-CNN mentioned by @usharengaraju (Tensor Girl) in this discussion topic.
\nAs of the current version, it has a validation schema based on \"subsequences (see this topic about validation schemas).
\nIt's a full pipeline using purely pytorch and torchvision, so I hope it's interesting for the community and it serves as a good kick-starter for some of you.
\nIt has a LB=0.416
Hello community,
\nI just finished adapting a version of the Faster R-CNN mentioned by @usharengaraju (Tensor Girl) in this discussion topic.
\nAs of the current version, it has a validation schema based on \"subsequences (see this topic about validation schemas).
\nIt's a full pipeline using purely pytorch and torchvision, so I hope it's interesting for the community and it serves as a good kick-starter for some of you.
\nIt has a LB=0.416
Greate notebook! Thank you for sharing!
\nCould I ask you a nowise question - why pytroch FasterRCNN does not trainable on images without BBs?
Hi @meowmeowmeowmeowmeow thanks for commenting and upvoting.
\nIt seems like a problem that is deeply rooted and might not be trivially addressable. When I add samples with no annotations, I get the following error:
\nValueError: No ground-truth boxes available for one of the images during training\nIt looks like they are just plainly not supported by the architecture.
\nOn the other hand, I have seen that the other starters are dropping the images with no annotations as well, so it might be a common condition of objects detection models.
\nI don't know much so take what I'm saying with a grain of salt.
", "rawMarkdown": "Hi @meowmeowmeowmeowmeow thanks for commenting and upvoting.\n\nIt seems like a problem that is deeply rooted and might not be trivially addressable. When I add samples with no annotations, I get the following error:\n\n```python\nValueError: No ground-truth boxes available for one of the images during training\n```\n\nIt looks like they are just plainly not supported by the architecture. \n\nOn the other hand, I have seen that the other starters are dropping the images with no annotations as well, so it might be a common condition of objects detection models. \n\nI don't know much so take what I'm saying with a grain of salt.", "votes": null }, { "id": "1593307", "postDate": "11/23/2021 18:59:55", "content": "Thank you for the replay. I am actually pretty noob in object detection, do you know more elegant way of solving this problem rather than just dropping ?
", "rawMarkdown": "Thank you for the replay. I am actually pretty noob in object detection, do you know more elegant way of solving this problem rather than just dropping ?", "votes": null }, { "id": "1593668", "postDate": "11/24/2021 06:31:37", "content": "Hi @drhabib, unluckily I'm a noob too and I don't. I will address the problem at some point though, but it's not a first priority since there are a lot of low-hanging fruits right now.
", "rawMarkdown": "Hi @drhabib, unluckily I'm a noob too and I don't. I will address the problem at some point though, but it's not a first priority since there are a lot of low-hanging fruits right now.", "votes": null }, { "id": "1593944", "postDate": "11/24/2021 12:15:01", "content": "It requires tensorflow. so how to convert the pretrained pytorch weights to TF2 format?
", "rawMarkdown": "It requires tensorflow. so how to convert the pretrained pytorch weights to TF2 format?", "votes": null }, { "id": "1594010", "postDate": "11/24/2021 13:16:58", "content": "I have no practical experience with FasterRCNN (except one from detection), but I don't see any limitation by model architecture. If there is no annotation per sample, all proposed regions should be fined.
\nLooks like that limitation comes from PyTorch implementation.
Hi @dragonzhang
\nTensorflow is required for the \"Performance prize\" only, for the LB prize anything goes. As I am more of a pytorch guy, my solutions will be on this side for now.
Greate notebook! Thank you for sharing!
\nCould I ask you a nowise question - why pytroch FasterRCNN does not trainable on images without BBs?
Hi @meowmeowmeowmeowmeow thanks for commenting and upvoting.
\nIt seems like a problem that is deeply rooted and might not be trivially addressable. When I add samples with no annotations, I get the following error:
\nValueError: No ground-truth boxes available for one of the images during training\nIt looks like they are just plainly not supported by the architecture.
\nOn the other hand, I have seen that the other starters are dropping the images with no annotations as well, so it might be a common condition of objects detection models.
\nI don't know much so take what I'm saying with a grain of salt.
", "votes": null, "replies": [] }, { "id": 1593307, "author_name": "drhabib", "author_url": "", "post_date": "11/23/2021 18:59:55", "content": "Thank you for the replay. I am actually pretty noob in object detection, do you know more elegant way of solving this problem rather than just dropping ?
", "votes": null, "replies": [] }, { "id": 1593668, "author_name": "julian3833", "author_url": "", "post_date": "11/24/2021 06:31:37", "content": "Hi @drhabib, unluckily I'm a noob too and I don't. I will address the problem at some point though, but it's not a first priority since there are a lot of low-hanging fruits right now.
", "votes": null, "replies": [] }, { "id": 1594010, "author_name": "meowmeowmeowmeowmeow", "author_url": "", "post_date": "11/24/2021 13:16:58", "content": "I have no practical experience with FasterRCNN (except one from detection), but I don't see any limitation by model architecture. If there is no annotation per sample, all proposed regions should be fined.
\nLooks like that limitation comes from PyTorch implementation.
It requires tensorflow. so how to convert the pretrained pytorch weights to TF2 format?
", "votes": null, "replies": [ { "id": 1594454, "author_name": "julian3833", "author_url": "", "post_date": "11/24/2021 21:05:00", "content": "Hi @dragonzhang
\nTensorflow is required for the \"Performance prize\" only, for the LB prize anything goes. As I am more of a pytorch guy, my solutions will be on this side for now.