{ "id": 123685, "title": "playing with logits?", "url": "/competitions/pku-autonomous-driving/discussion/123685", "author_name": "", "post_date": "2019-12-29T16:14:49.220437900Z", "votes": 7, "comment_count": 15, "views": 0, "content": "

are you playing with logits threshold or you are using default logits > 0 for Model predictions threshold?\nplaying with logits can increase public lb score but IMO it might will hurt us badly in private lb, your thoughts?\nremember the public lb uses only 9% data\ni noticed few times that local cv increases but public lb score decreases!\nhow much data you are using for validation?

", "messages": [ { "id": "705883", "postDate": "12/29/2019 16:14:49", "content": "

are you playing with logits threshold or you are using default logits > 0 for Model predictions threshold?\nplaying with logits can increase public lb score but IMO it might will hurt us badly in private lb, your thoughts?\nremember the public lb uses only 9% data\ni noticed few times that local cv increases but public lb score decreases!\nhow much data you are using for validation?

", "rawMarkdown": "are you playing with logits threshold or you are using default **logits > 0** for Model predictions threshold?\nplaying with logits can increase public lb score but IMO it might will hurt us badly in private lb, your thoughts?\nremember the public lb uses only 9% data\ni noticed few times that local cv increases but public lb score decreases!\nhow much data you are using for validation?", "votes": null }, { "id": "706959", "postDate": "12/31/2019 04:45:41", "content": "
\n

playing with logits can increase public lb score but IMO it might will hurt us badly in private lb

\n
\n\n

I agree. I have fixed prediction threshold to 0.4 after converting logits to probability.

", "rawMarkdown": "> playing with logits can increase public lb score but IMO it might will hurt us badly in private lb\n\nI agree. I have fixed prediction threshold to 0.4 after converting logits to probability.", "votes": null }, { "id": "707005", "postDate": "12/31/2019 06:29:11", "content": "

Did you mean the threshold for output heatmap? 0.4 is quite high.

", "rawMarkdown": "Did you mean the threshold for output heatmap? 0.4 is quite high.", "votes": null }, { "id": "707011", "postDate": "12/31/2019 06:32:41", "content": "

I'm using 20% of training data to be the validation set. But my model is not able to set the threshold that low.. It will predict several center points on cars which are bigger in the image.

", "rawMarkdown": "I'm using 20% of training data to be the validation set. But my model is not able to set the threshold that low.. It will predict several center points on cars which are bigger in the image.", "votes": null }, { "id": "707180", "postDate": "12/31/2019 11:28:04", "content": "

I use th=-0.5 for raw logits, similar

", "rawMarkdown": "I use th=-0.5 for raw logits, similar", "votes": null }, { "id": "707183", "postDate": "12/31/2019 11:32:25", "content": "

One thing I consider might help is visualization heatmap, with lower threshold for heatmap\n- lower false negative, or less missing cars\n- some overlapped points\nIsn't it like a trade-off?

", "rawMarkdown": "One thing I consider might help is visualization heatmap, with lower threshold for heatmap\n- lower false negative, or less missing cars\n- some overlapped points\nIsn't it like a trade-off?", "votes": null }, { "id": "707219", "postDate": "12/31/2019 12:39:08", "content": "

The centernet paper says:

\n\n
\n

We detect all responses whose value is greater or equal to its 8-connected neighbors

\n
\n\n

Did you try it ?

", "rawMarkdown": "The centernet paper says:\n\n> We detect all responses whose value is greater or equal to its 8-connected neighbors\n\nDid you try it ?", "votes": null }, { "id": "707230", "postDate": "12/31/2019 12:56:16", "content": "

Yes, before output the heatmap prediction, I did the maxpooling(3x3) to get this effect. But it still has some repeat center predictions on the same car. I guess the problem might due to the gaussian heatmap label is too big on some cars. But thanks a lot for your remind!

", "rawMarkdown": "Yes, before output the heatmap prediction, I did the maxpooling(3x3) to get this effect. But it still has some repeat center predictions on the same car. I guess the problem might due to the gaussian heatmap label is too big on some cars. But thanks a lot for your remind!", "votes": null }, { "id": "707232", "postDate": "12/31/2019 12:58:21", "content": "

wow, looks like I need to try with larger threshold. I set the threshold to 0.2 right now.

", "rawMarkdown": "wow, looks like I need to try with larger threshold. I set the threshold to 0.2 right now.", "votes": null }, { "id": "707235", "postDate": "12/31/2019 13:03:01", "content": "

May I ask you guy's local CV score? Mine local CV is 0.15, and LB 0.069

", "rawMarkdown": "May I ask you guy's local CV score? Mine local CV is 0.15, and LB 0.069", "votes": null }, { "id": "707379", "postDate": "12/31/2019 18:04:16", "content": "

Dear @niuddd if i see the visualization then i can see some missing cars,i mean the model didn't detect few cars and most of them are large vehicle's\ndid you face same issue? how to deal with this problem?

", "rawMarkdown": "Dear @niuddd if i see the visualization then i can see some missing cars,i mean the model didn't detect few cars and most of them are large vehicle's\ndid you face same issue? how to deal with this problem?", "votes": null }, { "id": "707405", "postDate": "12/31/2019 18:42:53", "content": "

is th=-0.5 reliable?

", "rawMarkdown": "is th=-0.5 reliable?", "votes": null }, { "id": "707406", "postDate": "12/31/2019 18:43:36", "content": "

it seems like my model doesn't detect large cars well,i tried 8% data for validation but my focal loss shows inconsistency in validation

", "rawMarkdown": "it seems like my model doesn't detect large cars well,i tried 8% data for validation but my focal loss shows inconsistency in validation", "votes": null }, { "id": "707582", "postDate": "01/01/2020 05:25:00", "content": "

I am trying to predict width and height,then using width and height to solve the more points in a same car...just starting,need time to evaluation.\nWhat do you think of this? width and height?

", "rawMarkdown": "I am trying to predict width and height,then using width and height to solve the more points in a same car...just starting,need time to evaluation.\nWhat do you think of this? width and height?", "votes": null }, { "id": "707667", "postDate": "01/01/2020 09:15:24", "content": "

@mobassir My focal loss always get overfitting very easily. Im still trying to solve this issue.\n@guobaozi I think the multi-points on same car issue can be decreased by simply adjusting the heatmap label size, tuning the focal loss parameters(alpha, beta) and increasing threshold of raw prediction.

", "rawMarkdown": "mobassir My focal loss always get overfitting very easily. Im still trying to solve this issue.\n@guobaozi I think the multi-points on same car issue can be decreased by simply adjusting the heatmap label size, tuning the focal loss parameters(alpha, beta) and increasing threshold of raw prediction.", "votes": null }, { "id": "707717", "postDate": "01/01/2020 10:58:56", "content": "

Thank you for your reply. 1,simply adjusting the heatmap label size, tuning the focal loss parameters.I think this is training method.\n2, Predicting the width and height for post process,the function \"clear_duplicates\" maybe performe better.

\n\n

When training width and height, we can reduce the w&h loss factor to reduce the width and height block's impact to other blocks.

", "rawMarkdown": "Thank you for your reply. 1,simply adjusting the heatmap label size, tuning the focal loss parameters.I think this is training method.\n2, Predicting the width and height for post process,the function \"clear_duplicates\" maybe performe better.\n\nWhen training width and height, we can reduce the w&h loss factor to reduce the width and height block's impact to other blocks.", "votes": null } ], "comments": [ { "id": 706959, "author_name": "nihei123", "author_url": "", "post_date": "12/31/2019 04:45:41", "content": "
\n

playing with logits can increase public lb score but IMO it might will hurt us badly in private lb

\n
\n\n

I agree. I have fixed prediction threshold to 0.4 after converting logits to probability.

", "votes": null, "replies": [ { "id": 707005, "author_name": "xiejialun", "author_url": "", "post_date": "12/31/2019 06:29:11", "content": "

Did you mean the threshold for output heatmap? 0.4 is quite high.

", "votes": null, "replies": [] }, { "id": 707180, "author_name": "niuddd", "author_url": "", "post_date": "12/31/2019 11:28:04", "content": "

I use th=-0.5 for raw logits, similar

", "votes": null, "replies": [] }, { "id": 707232, "author_name": "xiejialun", "author_url": "", "post_date": "12/31/2019 12:58:21", "content": "

wow, looks like I need to try with larger threshold. I set the threshold to 0.2 right now.

", "votes": null, "replies": [] }, { "id": 707235, "author_name": "xiejialun", "author_url": "", "post_date": "12/31/2019 13:03:01", "content": "

May I ask you guy's local CV score? Mine local CV is 0.15, and LB 0.069

", "votes": null, "replies": [] }, { "id": 707405, "author_name": "mobassir", "author_url": "", "post_date": "12/31/2019 18:42:53", "content": "

is th=-0.5 reliable?

", "votes": null, "replies": [] } ] }, { "id": 707011, "author_name": "xiejialun", "author_url": "", "post_date": "12/31/2019 06:32:41", "content": "

I'm using 20% of training data to be the validation set. But my model is not able to set the threshold that low.. It will predict several center points on cars which are bigger in the image.

", "votes": null, "replies": [ { "id": 707219, "author_name": "nihei123", "author_url": "", "post_date": "12/31/2019 12:39:08", "content": "

The centernet paper says:

\n\n
\n

We detect all responses whose value is greater or equal to its 8-connected neighbors

\n
\n\n

Did you try it ?

", "votes": null, "replies": [] }, { "id": 707230, "author_name": "xiejialun", "author_url": "", "post_date": "12/31/2019 12:56:16", "content": "

Yes, before output the heatmap prediction, I did the maxpooling(3x3) to get this effect. But it still has some repeat center predictions on the same car. I guess the problem might due to the gaussian heatmap label is too big on some cars. But thanks a lot for your remind!

", "votes": null, "replies": [] }, { "id": 707406, "author_name": "mobassir", "author_url": "", "post_date": "12/31/2019 18:43:36", "content": "

it seems like my model doesn't detect large cars well,i tried 8% data for validation but my focal loss shows inconsistency in validation

", "votes": null, "replies": [] }, { "id": 707582, "author_name": "guobaozi", "author_url": "", "post_date": "01/01/2020 05:25:00", "content": "

I am trying to predict width and height,then using width and height to solve the more points in a same car...just starting,need time to evaluation.\nWhat do you think of this? width and height?

", "votes": null, "replies": [] }, { "id": 707667, "author_name": "xiejialun", "author_url": "", "post_date": "01/01/2020 09:15:24", "content": "

@mobassir My focal loss always get overfitting very easily. Im still trying to solve this issue.\n@guobaozi I think the multi-points on same car issue can be decreased by simply adjusting the heatmap label size, tuning the focal loss parameters(alpha, beta) and increasing threshold of raw prediction.

", "votes": null, "replies": [] }, { "id": 707717, "author_name": "guobaozi", "author_url": "", "post_date": "01/01/2020 10:58:56", "content": "

Thank you for your reply. 1,simply adjusting the heatmap label size, tuning the focal loss parameters.I think this is training method.\n2, Predicting the width and height for post process,the function \"clear_duplicates\" maybe performe better.

\n\n

When training width and height, we can reduce the w&h loss factor to reduce the width and height block's impact to other blocks.

", "votes": null, "replies": [] } ] }, { "id": 707183, "author_name": "niuddd", "author_url": "", "post_date": "12/31/2019 11:32:25", "content": "

One thing I consider might help is visualization heatmap, with lower threshold for heatmap\n- lower false negative, or less missing cars\n- some overlapped points\nIsn't it like a trade-off?

", "votes": null, "replies": [ { "id": 707379, "author_name": "mobassir", "author_url": "", "post_date": "12/31/2019 18:04:16", "content": "

Dear @niuddd if i see the visualization then i can see some missing cars,i mean the model didn't detect few cars and most of them are large vehicle's\ndid you face same issue? how to deal with this problem?

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "705883": "are you playing with logits threshold or you are using default **logits > 0** for Model predictions threshold?\nplaying with logits can increase public lb score but IMO it might will hurt us badly in private lb, your thoughts?\nremember the public lb uses only 9% data\ni noticed few times that local cv increases but public lb score decreases!\nhow much data you are using for validation?", "706959": "> playing with logits can increase public lb score but IMO it might will hurt us badly in private lb\n\nI agree. I have fixed prediction threshold to 0.4 after converting logits to probability.", "707005": "Did you mean the threshold for output heatmap? 0.4 is quite high.", "707011": "I'm using 20% of training data to be the validation set. But my model is not able to set the threshold that low.. It will predict several center points on cars which are bigger in the image.", "707180": "I use th=-0.5 for raw logits, similar", "707183": "One thing I consider might help is visualization heatmap, with lower threshold for heatmap\n- lower false negative, or less missing cars\n- some overlapped points\nIsn't it like a trade-off?", "707219": "The centernet paper says:\n\n> We detect all responses whose value is greater or equal to its 8-connected neighbors\n\nDid you try it ?", "707230": "Yes, before output the heatmap prediction, I did the maxpooling(3x3) to get this effect. But it still has some repeat center predictions on the same car. I guess the problem might due to the gaussian heatmap label is too big on some cars. But thanks a lot for your remind!", "707232": "wow, looks like I need to try with larger threshold. I set the threshold to 0.2 right now.", "707235": "May I ask you guy's local CV score? Mine local CV is 0.15, and LB 0.069", "707379": "Dear @niuddd if i see the visualization then i can see some missing cars,i mean the model didn't detect few cars and most of them are large vehicle's\ndid you face same issue? how to deal with this problem?", "707405": "is th=-0.5 reliable?", "707406": "it seems like my model doesn't detect large cars well,i tried 8% data for validation but my focal loss shows inconsistency in validation", "707582": "I am trying to predict width and height,then using width and height to solve the more points in a same car...just starting,need time to evaluation.\nWhat do you think of this? width and height?", "707667": "mobassir My focal loss always get overfitting very easily. Im still trying to solve this issue.\n@guobaozi I think the multi-points on same car issue can be decreased by simply adjusting the heatmap label size, tuning the focal loss parameters(alpha, beta) and increasing threshold of raw prediction.", "707717": "Thank you for your reply. 1,simply adjusting the heatmap label size, tuning the focal loss parameters.I think this is training method.\n2, Predicting the width and height for post process,the function \"clear_duplicates\" maybe performe better.\n\nWhen training width and height, we can reduce the w&h loss factor to reduce the width and height block's impact to other blocks." }, "source": "meta" }