![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2Fbdfae6f079a999416781bcf755727e9c%2FID_2bb252857.jpg?generation=1579153951682201&alt=media\")
Hi!
\n\nI was doing visualization of my predictions and compared focal loss outputs with regular loss hourglass outputs. Regular loss performed better on public LB, so I was wondering why. I found that although focal loss prediction is much better at recognizing positions of many cars undetected by regular loss model, the regression values for these cars are not well fitted. I think that might be a problem due to I was still using regular loss for regression values, so there might be a mismatch between these losses.
\n\nWhat do you think?
\n\nHere are some pictures, the first image is regular loss prediction, and the other two are focal loss predictions with different thresholds (-0.5 and -1)
\n\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2F46d38560ead6350111ef7451ff8f8c20%2FID_1ae5d76da.jpg?generation=1579153951771128&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2F233937df14ad712f52aa3330dea8e209%2FID_4ee22138a.jpg?generation=1579153951713757&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2Fb2c39370d197d11ad710d42123cd69ec%2FID_3c61e369f.jpg?generation=1579153951728934&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2Ff54b64ce1d12a0dbee2cc282dc83570a%2FID_3da1329cf.jpg?generation=1579153951931572&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2Ff17a3d75f50bbc0b0d14cd03e9b9cca1%2FID_3abfc5331.jpg?generation=1579153951978508&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2F195b3d75d4a2c7c1bb26ce9e7059431c%2FID_7a3dba583.jpg?generation=1579153951974002&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2F10b7bc4babcc4285665f30af2d7e5ac4%2FID_6b27966c4.jpg?generation=1579153952019810&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2F8016da880a521432942222a934d70ece%2FID_4e9ead80d.jpg?generation=1579153952170600&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2Fc279a20bdbbd8f680d2d46039c03f82f%2FID_0c8a9191b.jpg?generation=1579153952330815&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2F15d71f6e0d2a6b2613204519c70b667f%2FID_2eef26c90.jpg?generation=1579153952401644&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2F54013ea1796e7e240b836aba894eb5ca%2FID_0a5908893.jpg?generation=1579153952627376&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2F69e6bc907b4d8678440cef5db0d14fd4%2FID_0c3b1786e.jpg?generation=1579153952620987&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2F5b721d41b6d67393dffe948620da6842%2FID_6dc4db4c3.jpg?generation=1579153952695498&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2F32218d21f1ef295f51978e9d007a07b0%2FID_5d0ea0e27.jpg?generation=1579153952955299&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2Fcef17e9386a87f6372590db8a51dd8a3%2FID_2ed1703fa.jpg?generation=1579153953031646&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2F5db4cf45e84b5813d68bb79b00db849a%2FID_07a9fd761.jpg?generation=1579153952955357&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1150566%2Fc1c3b8786d1485251cb50565a1e2d737%2FID_5d73de275.jpg?generation=1579153953033180&alt=media\")
Hi!
\n\nI was doing visualization of my predictions and compared focal loss outputs with regular loss hourglass outputs. Regular loss performed better on public LB, so I was wondering why. I found that although focal loss prediction is much better at recognizing positions of many cars undetected by regular loss model, the regression values for these cars are not well fitted. I think that might be a problem due to I was still using regular loss for regression values, so there might be a mismatch between these losses.
\n\nWhat do you think?
\n\nHere are some pictures, the first image is regular loss prediction, and the other two are focal loss predictions with different thresholds (-0.5 and -1)
\n\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
HI i found that tuning the mix between the losses elps a lot. EHich model are you using exactly?
", "rawMarkdown": "HI i found that tuning the mix between the losses elps a lot. EHich model are you using exactly?", "votes": null }, { "id": "720681", "postDate": "01/16/2020 16:34:26", "content": "I'm using hourglass 104. What exactly do you mean mix between losses? @felipebihaiek
", "rawMarkdown": "I'm using hourglass 104. What exactly do you mean mix between losses? @felipebihaiek", "votes": null }, { "id": "720941", "postDate": "01/16/2020 21:47:04", "content": "Like the ratio between summing regr loss and mask loss e.g. L1 + BCE, or the mask loss themselves like focal + BCE, etc such as (weight) * BCE + (1 - weight) * focal
", "rawMarkdown": "Like the ratio between summing regr loss and mask loss e.g. L1 + BCE, or the mask loss themselves like focal + BCE, etc such as (weight) * BCE + (1 - weight) * focal", "votes": null }, { "id": "721002", "postDate": "01/16/2020 23:36:01", "content": "I see. Thanks! I'll try it out!
", "rawMarkdown": "I see. Thanks! I'll try it out!", "votes": null }, { "id": "721998", "postDate": "01/18/2020 01:10:16", "content": "Hi exactly what Brian said
", "rawMarkdown": "Hi exactly what Brian said", "votes": null } ], "comments": [ { "id": 720638, "author_name": "felipebihaiek", "author_url": "", "post_date": "01/16/2020 15:49:26", "content": "HI i found that tuning the mix between the losses elps a lot. EHich model are you using exactly?
", "votes": null, "replies": [ { "id": 720681, "author_name": "tonychenxyz", "author_url": "", "post_date": "01/16/2020 16:34:26", "content": "I'm using hourglass 104. What exactly do you mean mix between losses? @felipebihaiek
", "votes": null, "replies": [] }, { "id": 720941, "author_name": "joonl04", "author_url": "", "post_date": "01/16/2020 21:47:04", "content": "Like the ratio between summing regr loss and mask loss e.g. L1 + BCE, or the mask loss themselves like focal + BCE, etc such as (weight) * BCE + (1 - weight) * focal
", "votes": null, "replies": [] }, { "id": 721002, "author_name": "tonychenxyz", "author_url": "", "post_date": "01/16/2020 23:36:01", "content": "I see. Thanks! I'll try it out!
", "votes": null, "replies": [] }, { "id": 721998, "author_name": "felipebihaiek", "author_url": "", "post_date": "01/18/2020 01:10:16", "content": "Hi exactly what Brian said
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "720071": "Hi!\n\nI was doing visualization of my predictions and compared focal loss outputs with regular loss hourglass outputs. Regular loss performed better on public LB, so I was wondering why. I found that although focal loss prediction is much better at recognizing positions of many cars undetected by regular loss model, the regression values for these cars are not well fitted. I think that might be a problem due to I was still using regular loss for regression values, so there might be a mismatch between these losses.\n\nWhat do you think?\n\nHere are some pictures, the first image is regular loss prediction, and the other two are focal loss predictions with different thresholds (-0.5 and -1)\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n", "720638": "HI i found that tuning the mix between the losses elps a lot. EHich model are you using exactly?", "720681": "I'm using hourglass 104. What exactly do you mean mix between losses? @felipebihaiek", "720941": "Like the ratio between summing regr loss and mask loss e.g. L1 + BCE, or the mask loss themselves like focal + BCE, etc such as (weight) * BCE + (1 - weight) * focal", "721002": "I see. Thanks! I'll try it out!", "721998": "Hi exactly what Brian said" }, "source": "meta" }