Ok. I was working on something for a week and finally got 0.93 CV and 0.87 LB with Lower Resolution o.O.\nBut It will be completely difficult for me to extend this approach on Intermediate Resolution because of limited resources. Specially Kaggle Notebooks xD.\nBut I think there is a way to get such results even with lowest resolution images. \n(Hint : Approach is based on Augmentation and adjustment of tiles) \n(Just wanted to share)\nHave Fun.
", "messages": [ { "id": 895135, "postDate": "2020-06-21T06:01:01.507Z", "content": "Ok. I was working on something for a week and finally got 0.93 CV and 0.87 LB with Lower Resolution o.O.\nBut It will be completely difficult for me to extend this approach on Intermediate Resolution because of limited resources. Specially Kaggle Notebooks xD.\nBut I think there is a way to get such results even with lowest resolution images. \n(Hint : Approach is based on Augmentation and adjustment of tiles) \n(Just wanted to share)\nHave Fun.
", "rawMarkdown": "Ok. I was working on something for a week and finally got 0.93 CV and 0.87 LB with Lower Resolution o.O.\nBut It will be completely difficult for me to extend this approach on Intermediate Resolution because of limited resources. Specially Kaggle Notebooks xD.\nBut I think there is a way to get such results even with lowest resolution images. \n(Hint : Approach is based on Augmentation and adjustment of tiles) \n(Just wanted to share)\nHave Fun.", "votes": 13 }, { "id": 895727, "postDate": "2020-06-21T15:08:54.137Z", "content": "I had very surprising results on low resolution too. One of my first submission was at 0.81CV and somehow scored 0.85 LB.
\n\nMy guess is that the distribution of the private/public set is different. The organizer say that they have a good mix of data provider so one of my first guess is that the isup_grade * data_provider distribution is not the same creating some variation in kappa. There's also the problem of duplicates.
\n\nIntermediary resolution scores higher on LB and is somewhat more stable with the highest diffs being around 0.02 for me.
", "rawMarkdown": "I had very surprising results on low resolution too. One of my first submission was at 0.81CV and somehow scored 0.85 LB. \n\nMy guess is that the distribution of the private/public set is different. The organizer say that they have a good mix of data provider so one of my first guess is that the isup_grade * data_provider distribution is not the same creating some variation in kappa. There's also the problem of duplicates.\n\nIntermediary resolution scores higher on LB and is somewhat more stable with the highest diffs being around 0.02 for me.", "votes": 1, "replies": [ { "id": 896177, "postDate": "2020-06-21T23:45:42.910Z", "content": "Yes. I think Karolinska is effecting a lot in public LB.
", "rawMarkdown": "Yes. I think Karolinska is effecting a lot in public LB." } ] }, { "id": 904235, "postDate": "2020-06-27T13:09:44.287Z", "content": "is your current best score model trained using kaggle kernels only?
", "rawMarkdown": "is your current best score model trained using kaggle kernels only?", "replies": [ { "id": 904258, "postDate": "2020-06-27T13:31:55.313Z", "content": "Nop. But I have model with CV 0.87 with kaggle kernal only. I didn't submitted that. However, Now I am training on Kaggle TPUs lets see what's next.\nI am kind on considering to create a public notebook to train Iafoos method with TPU. \nBut It takes a lot of effort. 😅😅
", "rawMarkdown": "Nop. But I have model with CV 0.87 with kaggle kernal only. I didn't submitted that. However, Now I am training on Kaggle TPUs lets see what's next.\nI am kind on considering to create a public notebook to train Iafoos method with TPU. \nBut It takes a lot of effort. 😅😅" } ] }, { "id": 895316, "postDate": "2020-06-21T08:51:40.577Z", "content": "What was the low resolution dataset you used?
", "rawMarkdown": "What was the low resolution dataset you used?", "replies": [ { "id": 896176, "postDate": "2020-06-21T23:44:59.803Z", "content": "It means patches from lowest resolution image.
", "rawMarkdown": "It means patches from lowest resolution image." } ] }, { "id": 905341, "postDate": "2020-06-28T13:16:34.870Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 905348, "postDate": "2020-06-28T13:27:02.470Z", "content": "Currently I am using Ordinal Regression.
\n\n#\n\ny = dense(5, activation='sigmoid')\ny = summation(y)
\n\n#\n\nThis pseudo code might help you with this.
", "rawMarkdown": "Currently I am using Ordinal Regression.\n################\ny = dense(5, activation='sigmoid')\ny = summation(y)\n################\nThis pseudo code might help you with this." }, { "id": 905417, "postDate": "2020-06-28T14:36:48.827Z", "rawMarkdown": "", "isDeleted": true }, { "id": 905422, "postDate": "2020-06-28T14:38:53.187Z", "content": "You can check my discussion post about implementation
", "rawMarkdown": "You can check my discussion [post](https://www.kaggle.com/c/prostate-cancer-grade-assessment/discussion/162334) about implementation " }, { "id": 905579, "postDate": "2020-06-28T16:48:37.377Z", "rawMarkdown": "", "isDeleted": true } ] } ], "comments": [ { "id": 895727, "author_name": "Arnaud Roussel", "author_url": "", "post_date": "2020-06-21T15:08:54.137000", "content": "I had very surprising results on low resolution too. One of my first submission was at 0.81CV and somehow scored 0.85 LB.
\n\nMy guess is that the distribution of the private/public set is different. The organizer say that they have a good mix of data provider so one of my first guess is that the isup_grade * data_provider distribution is not the same creating some variation in kappa. There's also the problem of duplicates.
\n\nIntermediary resolution scores higher on LB and is somewhat more stable with the highest diffs being around 0.02 for me.
", "votes": 1, "replies": [ { "id": 896177, "author_name": "Salman", "author_url": "", "post_date": "2020-06-21T23:45:42.910000", "content": "Yes. I think Karolinska is effecting a lot in public LB.
", "votes": 0, "replies": [] } ] }, { "id": 904235, "author_name": "Mobassir", "author_url": "", "post_date": "2020-06-27T13:09:44.287000", "content": "is your current best score model trained using kaggle kernels only?
", "votes": 0, "replies": [ { "id": 904258, "author_name": "Salman", "author_url": "", "post_date": "2020-06-27T13:31:55.313000", "content": "Nop. But I have model with CV 0.87 with kaggle kernal only. I didn't submitted that. However, Now I am training on Kaggle TPUs lets see what's next.\nI am kind on considering to create a public notebook to train Iafoos method with TPU. \nBut It takes a lot of effort. 😅😅
", "votes": 0, "replies": [] } ] }, { "id": 895316, "author_name": "Kurian Benoy", "author_url": "", "post_date": "2020-06-21T08:51:40.577000", "content": "What was the low resolution dataset you used?
", "votes": 0, "replies": [ { "id": 896176, "author_name": "Salman", "author_url": "", "post_date": "2020-06-21T23:44:59.803000", "content": "It means patches from lowest resolution image.
", "votes": 0, "replies": [] } ] }, { "id": 905341, "author_name": "", "author_url": "", "post_date": "2020-06-28T13:16:34.870000", "content": "", "votes": 0, "replies": [ { "id": 905348, "author_name": "Salman", "author_url": "", "post_date": "2020-06-28T13:27:02.470000", "content": "Currently I am using Ordinal Regression.
\n\n#\n\ny = dense(5, activation='sigmoid')\ny = summation(y)
\n\n#\n\nThis pseudo code might help you with this.
", "votes": 0, "replies": [] }, { "id": 905417, "author_name": "", "author_url": "", "post_date": "2020-06-28T14:36:48.827000", "content": "", "votes": 0, "replies": [] }, { "id": 905422, "author_name": "Salman", "author_url": "", "post_date": "2020-06-28T14:38:53.187000", "content": "You can check my discussion post about implementation
", "votes": 0, "replies": [] }, { "id": 905579, "author_name": "", "author_url": "", "post_date": "2020-06-28T16:48:37.377000", "content": "", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "895135": "Ok. I was working on something for a week and finally got 0.93 CV and 0.87 LB with Lower Resolution o.O.\nBut It will be completely difficult for me to extend this approach on Intermediate Resolution because of limited resources. Specially Kaggle Notebooks xD.\nBut I think there is a way to get such results even with lowest resolution images. \n(Hint : Approach is based on Augmentation and adjustment of tiles) \n(Just wanted to share)\nHave Fun.", "895727": "I had very surprising results on low resolution too. One of my first submission was at 0.81CV and somehow scored 0.85 LB. \n\nMy guess is that the distribution of the private/public set is different. The organizer say that they have a good mix of data provider so one of my first guess is that the isup_grade * data_provider distribution is not the same creating some variation in kappa. There's also the problem of duplicates.\n\nIntermediary resolution scores higher on LB and is somewhat more stable with the highest diffs being around 0.02 for me.", "904235": "is your current best score model trained using kaggle kernels only?", "895316": "What was the low resolution dataset you used?", "905341": "" } }