I was trying to explore the impact of using classifiers for removing encoded masks. The idea is already explained in other available notebooks. Primarily I was checking the notebook by @mobassir. I used his shared kernel which is Keras efficientnetb2 for classifying cloud.
\n\nI found that classifier has a good impact on public LB. Just using the same classifier from the above link I saw an improvement from 0.001 to 0.003 on various submissions. I think a better classifier might have a higher impact.
\n\nI would like to know your thoughts and suggestions on this. Are you trying classifiers? If you are trying which architecture you are using? Any more thoughts?
", "messages": [ { "id": "665161", "postDate": "11/04/2019 17:22:19", "content": "I was trying to explore the impact of using classifiers for removing encoded masks. The idea is already explained in other available notebooks. Primarily I was checking the notebook by @mobassir. I used his shared kernel which is Keras efficientnetb2 for classifying cloud.
\n\nI found that classifier has a good impact on public LB. Just using the same classifier from the above link I saw an improvement from 0.001 to 0.003 on various submissions. I think a better classifier might have a higher impact.
\n\nI would like to know your thoughts and suggestions on this. Are you trying classifiers? If you are trying which architecture you are using? Any more thoughts?
", "rawMarkdown": "I was trying to explore the impact of using classifiers for removing encoded masks. The idea is already explained in other available notebooks. Primarily I was checking the notebook by @mobassir. I used his shared kernel which is [Keras efficientnetb2 for classifying cloud](https://www.kaggle.com/mobassir/keras-efficientnetb2-for-classifying-cloud). \n\nI found that classifier has a good impact on public LB. Just using the same classifier from the above link I saw an improvement from 0.001 to 0.003 on various submissions. I think a better classifier might have a higher impact. \n\nI would like to know your thoughts and suggestions on this. Are you trying classifiers? If you are trying which architecture you are using? Any more thoughts?", "votes": null }, { "id": "665169", "postDate": "11/04/2019 17:33:13", "content": "Did you see impact on CV too?
", "rawMarkdown": "Did you see impact on CV too?", "votes": null }, { "id": "665173", "postDate": "11/04/2019 17:35:50", "content": "I haven't test t that yet. I will test that and let you know.
", "rawMarkdown": "I haven't test t that yet. I will test that and let you know.", "votes": null }, { "id": "665450", "postDate": "11/05/2019 02:48:53", "content": "LB 0.002 enhance with efficientnetb2 classifier
", "rawMarkdown": "LB 0.002 enhance with efficientnetb2 classifier", "votes": null }, { "id": "665451", "postDate": "11/05/2019 02:50:20", "content": "I think I have more confidence on the classifier than segmentation model , so I will keep it at final submission.
", "rawMarkdown": "I think I have more confidence on the classifier than segmentation model , so I will keep it at final submission.", "votes": null }, { "id": "667337", "postDate": "11/07/2019 06:08:20", "content": "Nice! Did you count how many mask removed by the classifier?
", "rawMarkdown": "Nice! Did you count how many mask removed by the classifier?", "votes": null }, { "id": "667361", "postDate": "11/07/2019 06:34:27", "content": "I didn't count it, since the predicting process is kind of merging together. I will count it next time I submit the kernel then share to here.
", "rawMarkdown": "I didn't count it, since the predicting process is kind of merging together. I will count it next time I submit the kernel then share to here.", "votes": null }, { "id": "667634", "postDate": "11/07/2019 13:31:28", "content": "@niuddd The prediction with classifier was : Fish 1315, Flower 1385, Gravel 1348, Sugar 2313\nThe prediction without classifier is : Fish 1481, Flower 1420, Gravel 1541, Sugar 2471
", "rawMarkdown": "niuddd The prediction with classifier was : Fish\t1315, Flower\t1385, Gravel\t1348, Sugar\t2313\nThe prediction without classifier is : Fish\t1481, Flower\t1420, Gravel\t1541, Sugar\t2471", "votes": null }, { "id": "667658", "postDate": "11/07/2019 14:16:18", "content": "Thanks! Your segment model predict 46.7% nonempty mask, with classifier you predict 43.0% nonempty mask.\nMy segment models predicting around 43% masks got the best public score, i'll try lower pixel thresholds and use classifier to remove false positives.
", "rawMarkdown": "Thanks! Your segment model predict 46.7% nonempty mask, with classifier you predict 43.0% nonempty mask.\nMy segment models predicting around 43% masks got the best public score, i'll try lower pixel thresholds and use classifier to remove false positives.", "votes": null }, { "id": "667660", "postDate": "11/07/2019 14:20:03", "content": "Thanks for the information! \nI have pretty low threshold on sugar class (only 0.4), others are higher than 0.6.
", "rawMarkdown": "Thanks for the information! \nI have pretty low threshold on sugar class (only 0.4), others are higher than 0.6.", "votes": null }, { "id": "667662", "postDate": "11/07/2019 14:24:44", "content": "pixel label of image is noisy and in consistent.
\n\nbut at image label, image label is more consistent and less noisy.\nhence it is possible that classification at image level may give better accuracy
", "rawMarkdown": "pixel label of image is noisy and in consistent.\n\nbut at image label, image label is more consistent and less noisy.\nhence it is possible that classification at image level may give better accuracy", "votes": null }, { "id": "667747", "postDate": "11/07/2019 16:33:34", "content": "LB +0.003 using classifier. on CV around + 0.001.
", "rawMarkdown": "LB +0.003 using classifier. on CV around + 0.001.", "votes": null }, { "id": "668905", "postDate": "11/09/2019 04:06:18", "content": "You are right. I think classifier will play an important role in this competition.
", "rawMarkdown": "You are right. I think classifier will play an important role in this competition.", "votes": null }, { "id": "672607", "postDate": "11/14/2019 02:34:00", "content": "I just want to state that contribution of the post-processing classifier (the mentioned kernel) come from Raman @samusram . He did a hard and great work on his original kernel. It is quite a pity that the original author got quite lower credits than forked kernels with few lines of changes.
", "rawMarkdown": "I just want to state that contribution of the post-processing classifier (the mentioned kernel) come from Raman @samusram . He did a hard and great work on [his original kernel](https://www.kaggle.com/samusram/cloud-classifier-for-post-processing). It is quite a pity that the original author got quite lower credits than forked kernels with few lines of changes.", "votes": null } ], "comments": [ { "id": 665169, "author_name": "igormunizims", "author_url": "", "post_date": "11/04/2019 17:33:13", "content": "Did you see impact on CV too?
", "votes": null, "replies": [ { "id": 665173, "author_name": "mykttu", "author_url": "", "post_date": "11/04/2019 17:35:50", "content": "I haven't test t that yet. I will test that and let you know.
", "votes": null, "replies": [] } ] }, { "id": 665450, "author_name": "xiejialun", "author_url": "", "post_date": "11/05/2019 02:48:53", "content": "LB 0.002 enhance with efficientnetb2 classifier
", "votes": null, "replies": [ { "id": 665451, "author_name": "xiejialun", "author_url": "", "post_date": "11/05/2019 02:50:20", "content": "I think I have more confidence on the classifier than segmentation model , so I will keep it at final submission.
", "votes": null, "replies": [] }, { "id": 667337, "author_name": "niuddd", "author_url": "", "post_date": "11/07/2019 06:08:20", "content": "Nice! Did you count how many mask removed by the classifier?
", "votes": null, "replies": [] }, { "id": 667361, "author_name": "xiejialun", "author_url": "", "post_date": "11/07/2019 06:34:27", "content": "I didn't count it, since the predicting process is kind of merging together. I will count it next time I submit the kernel then share to here.
", "votes": null, "replies": [] }, { "id": 667634, "author_name": "xiejialun", "author_url": "", "post_date": "11/07/2019 13:31:28", "content": "@niuddd The prediction with classifier was : Fish 1315, Flower 1385, Gravel 1348, Sugar 2313\nThe prediction without classifier is : Fish 1481, Flower 1420, Gravel 1541, Sugar 2471
", "votes": null, "replies": [] }, { "id": 667658, "author_name": "niuddd", "author_url": "", "post_date": "11/07/2019 14:16:18", "content": "Thanks! Your segment model predict 46.7% nonempty mask, with classifier you predict 43.0% nonempty mask.\nMy segment models predicting around 43% masks got the best public score, i'll try lower pixel thresholds and use classifier to remove false positives.
", "votes": null, "replies": [] }, { "id": 667660, "author_name": "xiejialun", "author_url": "", "post_date": "11/07/2019 14:20:03", "content": "Thanks for the information! \nI have pretty low threshold on sugar class (only 0.4), others are higher than 0.6.
", "votes": null, "replies": [] } ] }, { "id": 667662, "author_name": "hengck23", "author_url": "", "post_date": "11/07/2019 14:24:44", "content": "pixel label of image is noisy and in consistent.
\n\nbut at image label, image label is more consistent and less noisy.\nhence it is possible that classification at image level may give better accuracy
", "votes": null, "replies": [ { "id": 668905, "author_name": "mykttu", "author_url": "", "post_date": "11/09/2019 04:06:18", "content": "You are right. I think classifier will play an important role in this competition.
", "votes": null, "replies": [] } ] }, { "id": 667747, "author_name": "tugstugi", "author_url": "", "post_date": "11/07/2019 16:33:34", "content": "LB +0.003 using classifier. on CV around + 0.001.
", "votes": null, "replies": [] }, { "id": 672607, "author_name": "ratthachat", "author_url": "", "post_date": "11/14/2019 02:34:00", "content": "I just want to state that contribution of the post-processing classifier (the mentioned kernel) come from Raman @samusram . He did a hard and great work on his original kernel. It is quite a pity that the original author got quite lower credits than forked kernels with few lines of changes.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "665161": "I was trying to explore the impact of using classifiers for removing encoded masks. The idea is already explained in other available notebooks. Primarily I was checking the notebook by @mobassir. I used his shared kernel which is [Keras efficientnetb2 for classifying cloud](https://www.kaggle.com/mobassir/keras-efficientnetb2-for-classifying-cloud). \n\nI found that classifier has a good impact on public LB. Just using the same classifier from the above link I saw an improvement from 0.001 to 0.003 on various submissions. I think a better classifier might have a higher impact. \n\nI would like to know your thoughts and suggestions on this. Are you trying classifiers? If you are trying which architecture you are using? Any more thoughts?", "665169": "Did you see impact on CV too?", "665173": "I haven't test t that yet. I will test that and let you know.", "665450": "LB 0.002 enhance with efficientnetb2 classifier", "665451": "I think I have more confidence on the classifier than segmentation model , so I will keep it at final submission.", "667337": "Nice! Did you count how many mask removed by the classifier?", "667361": "I didn't count it, since the predicting process is kind of merging together. I will count it next time I submit the kernel then share to here.", "667634": "niuddd The prediction with classifier was : Fish\t1315, Flower\t1385, Gravel\t1348, Sugar\t2313\nThe prediction without classifier is : Fish\t1481, Flower\t1420, Gravel\t1541, Sugar\t2471", "667658": "Thanks! Your segment model predict 46.7% nonempty mask, with classifier you predict 43.0% nonempty mask.\nMy segment models predicting around 43% masks got the best public score, i'll try lower pixel thresholds and use classifier to remove false positives.", "667660": "Thanks for the information! \nI have pretty low threshold on sugar class (only 0.4), others are higher than 0.6.", "667662": "pixel label of image is noisy and in consistent.\n\nbut at image label, image label is more consistent and less noisy.\nhence it is possible that classification at image level may give better accuracy", "667747": "LB +0.003 using classifier. on CV around + 0.001.", "668905": "You are right. I think classifier will play an important role in this competition.", "672607": "I just want to state that contribution of the post-processing classifier (the mentioned kernel) come from Raman @samusram . He did a hard and great work on [his original kernel](https://www.kaggle.com/samusram/cloud-classifier-for-post-processing). It is quite a pity that the original author got quite lower credits than forked kernels with few lines of changes." }, "source": "meta" }