I come here everyday to check if anyone has posted, and it seems oddly quiet. Maybe because the competition is still in the early phases, but I don't know.
\nAnyways, I'll share what we've been working on lately.
\nIn my free time, I have been working on creating a small utility package to visualize the results of my predictions. I think I'll share it a notebook soon to help anyone who wants to participate in the competition.
\nI didn't have time to try out fancy models or techniques, so I've only created a small 2.5D pipeline using fastai that gets not so bad results on my validation set, and I'll probably share it this week.
\nMy thoughts are all going towards creating a custom Yolo dataset and training a Yolo model to predict the bounding boxes of contacts, then assigning pairs based on the predicted bboxes.
\nIf you'd like to share what you're doing, please do and let's get the conversation started.
", "messages": [ { "id": 2100710, "postDate": "2023-01-15T11:29:17.383Z", "content": "I come here everyday to check if anyone has posted, and it seems oddly quiet. Maybe because the competition is still in the early phases, but I don't know.
\nAnyways, I'll share what we've been working on lately.
\nIn my free time, I have been working on creating a small utility package to visualize the results of my predictions. I think I'll share it a notebook soon to help anyone who wants to participate in the competition.
\nI didn't have time to try out fancy models or techniques, so I've only created a small 2.5D pipeline using fastai that gets not so bad results on my validation set, and I'll probably share it this week.
\nMy thoughts are all going towards creating a custom Yolo dataset and training a Yolo model to predict the bounding boxes of contacts, then assigning pairs based on the predicted bboxes.
\nIf you'd like to share what you're doing, please do and let's get the conversation started.
", "rawMarkdown": "I come here everyday to check if anyone has posted, and it seems oddly quiet. Maybe because the competition is still in the early phases, but I don't know.\n\nAnyways, I'll share what we've been working on lately. \n\nIn my free time, I have been working on creating a small utility package to visualize the results of my predictions. I think I'll share it a notebook soon to help anyone who wants to participate in the competition.\n\nI didn't have time to try out fancy models or techniques, so I've only created a small 2.5D pipeline using fastai that gets not so bad results on my validation set, and I'll probably share it this week.\n\nMy thoughts are all going towards creating a custom Yolo dataset and training a Yolo model to predict the bounding boxes of contacts, then assigning pairs based on the predicted bboxes. \n\nIf you'd like to share what you're doing, please do and let's get the conversation started.", "votes": 14 }, { "id": 2101611, "postDate": "2023-01-16T03:44:29.383Z", "content": "I would recommend doing \"quick\" experiments.
\nIf you take hours on 1 epoch, the number of experiments you can do will be limited.
\nIf you model and preprocess wisely, you can easily improve the experiment time; my 1epoch time is around 5 min using CNNs.
Good reference would be the DFL competition, which discuss a lot of wise approaches upon 3D data.
\nhttps://www.kaggle.com/competitions/dfl-bundesliga-data-shootout/discussion
how could your 1 epoch time is only 5 mins even possible? I still stuck on the bottleneck of the opencv is to slow in reading images with its neighbor frames? or are using another approach?
I think for starters you can use LRU caching while reading images, over sampling contact data, and sampling a small percentage of the game plays to make your experiments faster
", "rawMarkdown": "I think for starters you can use LRU caching while reading images, over sampling contact data, and sampling a small percentage of the game plays to make your experiments faster", "votes": 2 } ] } ] }, { "id": 2101546, "postDate": "2023-01-16T01:54:10.607Z", "content": "I'm trying to crop image with helmet and person, then train a classification model to predict contact,but it's easily going overfitting.
", "rawMarkdown": "I'm trying to crop image with helmet and person, then train a classification model to predict contact,but it's easily going overfitting.", "votes": 3 }, { "id": 2100736, "postDate": "2023-01-15T11:58:35.630Z", "content": "I am working on optimizing my training process atm even tho my backbone is resnext50 as a pretrained with frozen layers and mlp with 8 layers. Training time of 1 fold without even epochs is 7hrs and the same code works on colab but dont work on kaggle so I feel like Clown
", "rawMarkdown": "I am working on optimizing my training process atm even tho my backbone is resnext50 as a pretrained with frozen layers and mlp with 8 layers. Training time of 1 fold without even epochs is 7hrs and the same code works on colab but dont work on kaggle so I feel like Clown ", "votes": 3 }, { "id": 2101319, "postDate": "2023-01-15T20:27:39.050Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 2101611, "author_name": "arutema47", "author_url": "", "post_date": "2023-01-16T03:44:29.383000", "content": "I would recommend doing \"quick\" experiments.
\nIf you take hours on 1 epoch, the number of experiments you can do will be limited.
\nIf you model and preprocess wisely, you can easily improve the experiment time; my 1epoch time is around 5 min using CNNs.
Good reference would be the DFL competition, which discuss a lot of wise approaches upon 3D data.
\nhttps://www.kaggle.com/competitions/dfl-bundesliga-data-shootout/discussion
how could your 1 epoch time is only 5 mins even possible? I still stuck on the bottleneck of the opencv is to slow in reading images with its neighbor frames? or are using another approach?
I think for starters you can use LRU caching while reading images, over sampling contact data, and sampling a small percentage of the game plays to make your experiments faster
", "votes": 2, "replies": [] } ] } ] }, { "id": 2101546, "author_name": "ant1", "author_url": "", "post_date": "2023-01-16T01:54:10.607000", "content": "I'm trying to crop image with helmet and person, then train a classification model to predict contact,but it's easily going overfitting.
", "votes": 3, "replies": [] }, { "id": 2100736, "author_name": "Dominik Mińkowski", "author_url": "", "post_date": "2023-01-15T11:58:35.630000", "content": "I am working on optimizing my training process atm even tho my backbone is resnext50 as a pretrained with frozen layers and mlp with 8 layers. Training time of 1 fold without even epochs is 7hrs and the same code works on colab but dont work on kaggle so I feel like Clown
", "votes": 3, "replies": [] }, { "id": 2101319, "author_name": "", "author_url": "", "post_date": "2023-01-15T20:27:39.050000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2100710": "I come here everyday to check if anyone has posted, and it seems oddly quiet. Maybe because the competition is still in the early phases, but I don't know.\n\nAnyways, I'll share what we've been working on lately. \n\nIn my free time, I have been working on creating a small utility package to visualize the results of my predictions. I think I'll share it a notebook soon to help anyone who wants to participate in the competition.\n\nI didn't have time to try out fancy models or techniques, so I've only created a small 2.5D pipeline using fastai that gets not so bad results on my validation set, and I'll probably share it this week.\n\nMy thoughts are all going towards creating a custom Yolo dataset and training a Yolo model to predict the bounding boxes of contacts, then assigning pairs based on the predicted bboxes. \n\nIf you'd like to share what you're doing, please do and let's get the conversation started.", "2101611": "I would recommend doing \"quick\" experiments.\nIf you take hours on 1 epoch, the number of experiments you can do will be limited.\nIf you model and preprocess wisely, you can easily improve the experiment time; my 1epoch time is around 5 min using CNNs.\n\nGood reference would be the DFL competition, which discuss a lot of wise approaches upon 3D data.\nhttps://www.kaggle.com/competitions/dfl-bundesliga-data-shootout/discussion", "2101546": "I'm trying to crop image with helmet and person, then train a classification model to predict contact,but it's easily going overfitting.", "2100736": "I am working on optimizing my training process atm even tho my backbone is resnext50 as a pretrained with frozen layers and mlp with 8 layers. Training time of 1 fold without even epochs is 7hrs and the same code works on colab but dont work on kaggle so I feel like Clown ", "2101319": "" } }