![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1182060%2F6746e084eb0e4f713714b50e707b9a45%2FScreenshot%20from%202019-11-10%2014-12-46.png?generation=1573405996013520&alt=media\")
Hi everyone, I was inspecting some of my models' predictions and I came across this sample:
\n\n
I think will be really hard for a model to get this right, if you look at the left bottom part, the image has 3 different classes.
\n\nI know that this is a noisy dataset, the labels are subjective and all that stuff, but what I would like to know is how are you improving model results in this kind of situation? so far I'm just hoping it can learn.
", "messages": [ { "id": "669919", "postDate": "11/10/2019 17:18:55", "content": "Hi everyone, I was inspecting some of my models' predictions and I came across this sample:
\n\n
I think will be really hard for a model to get this right, if you look at the left bottom part, the image has 3 different classes.
\n\nI know that this is a noisy dataset, the labels are subjective and all that stuff, but what I would like to know is how are you improving model results in this kind of situation? so far I'm just hoping it can learn.
", "rawMarkdown": "Hi everyone, I was inspecting some of my models' predictions and I came across this sample:\n\n\n\nI think will be really hard for a model to get this right, if you look at the left bottom part, the image has 3 different classes.\n\nI know that this is a noisy dataset, the labels are subjective and all that stuff, but what I would like to know is how are you improving model results in this kind of situation? so far I'm just hoping it can learn.", "votes": null }, { "id": "670040", "postDate": "11/10/2019 23:30:10", "content": "I think that using separated binary segmentation model for each class could help 😉
", "rawMarkdown": "I think that using separated binary segmentation model for each class could help 😉", "votes": null }, { "id": "670400", "postDate": "11/11/2019 12:26:44", "content": "I tried separated models for each class, but it was worse, did you had more luck?
", "rawMarkdown": "I tried separated models for each class, but it was worse, did you had more luck?", "votes": null }, { "id": "670533", "postDate": "11/11/2019 14:57:07", "content": "No, I didn't 😂 . But I will try other architectures and, maybe, using ensemble learning on them to find out whether a separated model for each could surpass one single model for all classes.
", "rawMarkdown": "No, I didn't 😂 . But I will try other architectures and, maybe, using ensemble learning on them to find out whether a separated model for each could surpass one single model for all classes.", "votes": null }, { "id": "670638", "postDate": "11/11/2019 17:11:47", "content": "Yes 😄 , that's also what I'm planning to do.
", "rawMarkdown": "Yes 😄 , that's also what I'm planning to do.", "votes": null }, { "id": "671352", "postDate": "11/12/2019 15:05:04", "content": "Hi Dimitre, here’s my simple idea : if you have an accurate classifier, it should be able to tell you that there is some fish in the picture. After that, one non-ideal method is to use GradCAM mask from the classifier for fish prediction (like Raman’s or Chris’ methods in the other topics)
", "rawMarkdown": "Hi Dimitre, here’s my simple idea : if you have an accurate classifier, it should be able to tell you that there is some fish in the picture. After that, one non-ideal method is to use GradCAM mask from the classifier for fish prediction (like Raman’s or Chris’ methods in the other topics)", "votes": null }, { "id": "671378", "postDate": "11/12/2019 15:42:33", "content": "I see @ratthachat , I think I should evaluate the my classification models together with the segmentations.
", "rawMarkdown": "I see @ratthachat , I think I should evaluate the my classification models together with the segmentations.", "votes": null }, { "id": "671507", "postDate": "11/12/2019 20:01:24", "content": "May be data augmentation can help you.
", "rawMarkdown": "May be data augmentation can help you.", "votes": null }, { "id": "671575", "postDate": "11/12/2019 22:25:44", "content": "Yes @shishu1421 , I'm also experimenting with that, but I think we should accept that some samples will be near impossible to predict right 😄
", "rawMarkdown": "Yes @shishu1421 , I'm also experimenting with that, but I think we should accept that some samples will be near impossible to predict right 😄", "votes": null }, { "id": "671741", "postDate": "11/13/2019 05:23:09", "content": "Yeaahhhh! @dimitreoliveira
", "rawMarkdown": "Yeaahhhh! @dimitreoliveira", "votes": null }, { "id": "672212", "postDate": "11/13/2019 16:24:22", "content": "Couldn't we try fine tuning VGG to output 4 values with multiple one hot encodings, say multiple classifications??
", "rawMarkdown": "Couldn't we try fine tuning VGG to output 4 values with multiple one hot encodings, say multiple classifications??", "votes": null }, { "id": "672408", "postDate": "11/13/2019 21:04:49", "content": "do you mean a multi-label model? because that's what I use for the classification phase.
", "rawMarkdown": "do you mean a multi-label model? because that's what I use for the classification phase.", "votes": null }, { "id": "673328", "postDate": "11/14/2019 20:40:43", "content": "yes, how did that work out???
", "rawMarkdown": "yes, how did that work out???", "votes": null }, { "id": "673334", "postDate": "11/14/2019 21:08:00", "content": "It improved my results for sure, it seems that removing false positives has a big impact on this competition, this was also pointed by others.
", "rawMarkdown": "It improved my results for sure, it seems that removing false positives has a big impact on this competition, this was also pointed by others.", "votes": null } ], "comments": [ { "id": 670040, "author_name": "phunghieu", "author_url": "", "post_date": "11/10/2019 23:30:10", "content": "I think that using separated binary segmentation model for each class could help 😉
", "votes": null, "replies": [ { "id": 670400, "author_name": "dimitreoliveira", "author_url": "", "post_date": "11/11/2019 12:26:44", "content": "I tried separated models for each class, but it was worse, did you had more luck?
", "votes": null, "replies": [] }, { "id": 670533, "author_name": "phunghieu", "author_url": "", "post_date": "11/11/2019 14:57:07", "content": "No, I didn't 😂 . But I will try other architectures and, maybe, using ensemble learning on them to find out whether a separated model for each could surpass one single model for all classes.
", "votes": null, "replies": [] }, { "id": 670638, "author_name": "dimitreoliveira", "author_url": "", "post_date": "11/11/2019 17:11:47", "content": "Yes 😄 , that's also what I'm planning to do.
", "votes": null, "replies": [] } ] }, { "id": 671352, "author_name": "ratthachat", "author_url": "", "post_date": "11/12/2019 15:05:04", "content": "Hi Dimitre, here’s my simple idea : if you have an accurate classifier, it should be able to tell you that there is some fish in the picture. After that, one non-ideal method is to use GradCAM mask from the classifier for fish prediction (like Raman’s or Chris’ methods in the other topics)
", "votes": null, "replies": [ { "id": 671378, "author_name": "dimitreoliveira", "author_url": "", "post_date": "11/12/2019 15:42:33", "content": "I see @ratthachat , I think I should evaluate the my classification models together with the segmentations.
", "votes": null, "replies": [] } ] }, { "id": 671507, "author_name": "shishu1421", "author_url": "", "post_date": "11/12/2019 20:01:24", "content": "May be data augmentation can help you.
", "votes": null, "replies": [ { "id": 671575, "author_name": "dimitreoliveira", "author_url": "", "post_date": "11/12/2019 22:25:44", "content": "Yes @shishu1421 , I'm also experimenting with that, but I think we should accept that some samples will be near impossible to predict right 😄
", "votes": null, "replies": [] }, { "id": 671741, "author_name": "shishu1421", "author_url": "", "post_date": "11/13/2019 05:23:09", "content": "Yeaahhhh! @dimitreoliveira
", "votes": null, "replies": [] } ] }, { "id": 672212, "author_name": "dynamite2055", "author_url": "", "post_date": "11/13/2019 16:24:22", "content": "Couldn't we try fine tuning VGG to output 4 values with multiple one hot encodings, say multiple classifications??
", "votes": null, "replies": [ { "id": 672408, "author_name": "dimitreoliveira", "author_url": "", "post_date": "11/13/2019 21:04:49", "content": "do you mean a multi-label model? because that's what I use for the classification phase.
", "votes": null, "replies": [] }, { "id": 673328, "author_name": "dynamite2055", "author_url": "", "post_date": "11/14/2019 20:40:43", "content": "yes, how did that work out???
", "votes": null, "replies": [] }, { "id": 673334, "author_name": "dimitreoliveira", "author_url": "", "post_date": "11/14/2019 21:08:00", "content": "It improved my results for sure, it seems that removing false positives has a big impact on this competition, this was also pointed by others.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "669919": "Hi everyone, I was inspecting some of my models' predictions and I came across this sample:\n\n\n\nI think will be really hard for a model to get this right, if you look at the left bottom part, the image has 3 different classes.\n\nI know that this is a noisy dataset, the labels are subjective and all that stuff, but what I would like to know is how are you improving model results in this kind of situation? so far I'm just hoping it can learn.", "670040": "I think that using separated binary segmentation model for each class could help 😉", "670400": "I tried separated models for each class, but it was worse, did you had more luck?", "670533": "No, I didn't 😂 . But I will try other architectures and, maybe, using ensemble learning on them to find out whether a separated model for each could surpass one single model for all classes.", "670638": "Yes 😄 , that's also what I'm planning to do.", "671352": "Hi Dimitre, here’s my simple idea : if you have an accurate classifier, it should be able to tell you that there is some fish in the picture. After that, one non-ideal method is to use GradCAM mask from the classifier for fish prediction (like Raman’s or Chris’ methods in the other topics)", "671378": "I see @ratthachat , I think I should evaluate the my classification models together with the segmentations.", "671507": "May be data augmentation can help you.", "671575": "Yes @shishu1421 , I'm also experimenting with that, but I think we should accept that some samples will be near impossible to predict right 😄", "671741": "Yeaahhhh! @dimitreoliveira", "672212": "Couldn't we try fine tuning VGG to output 4 values with multiple one hot encodings, say multiple classifications??", "672408": "do you mean a multi-label model? because that's what I use for the classification phase.", "673328": "yes, how did that work out???", "673334": "It improved my results for sure, it seems that removing false positives has a big impact on this competition, this was also pointed by others." }, "source": "meta" }