0.290 Private Score notebook
\nIn this competition we shake down from 2% to 11%. I am very sorry to my teammates @qiaoyuanfang. He did not win any medals because of me. He provided a lot of high-quality notebooks, and most of the results were in the silver medal. Because I am overly pursuing LB, and I adjusted the 15 classification parameters according to LB. This leads to serious overfitting. But 15 classification can indeed increase the performance.
\nOur highest private score during the competition was 0.286.
\nI chose the right way to average the confidence value and get a score of 0.290.
\nAnd both public LB and private LB increase at the same time.
My 15 classification uses a very simple model.
\nLabels: 15 classes, A class of an image is labelled 1 if 1 radiologists annotated the class on the image.
\nEfficientnet-b7 (imagenet) 5 fold:
\nimage size 600
\naug:
\nimg = tf.image.random_flip_left_right(img)
\nimg = tf.image.random_flip_up_down(img)
Efficientnet-b7 (noisy-student) 5 fold:
\nimage size 600
\naug:
\nimg = tf.image.random_flip_left_right(img)
\nimg = tf.image.random_flip_up_down(img)
densenet201 5 fold:
\nimage size 600
\nPre-training on the nih dataset
\naug:
\nimg = tf.image.random_flip_left_right(img)
\nimg = tf.image.random_flip_up_down(img)
\nimg = tf.image.random_hue(img, 0.2)
\nimg = tf.image.random_brightness(img, 0.2)
Finally, the confidence value is weighted average.
", "messages": [ { "id": 1258640, "postDate": "2021-03-31T19:08:21.583Z", "content": "0.290 Private Score notebook
\nIn this competition we shake down from 2% to 11%. I am very sorry to my teammates @qiaoyuanfang. He did not win any medals because of me. He provided a lot of high-quality notebooks, and most of the results were in the silver medal. Because I am overly pursuing LB, and I adjusted the 15 classification parameters according to LB. This leads to serious overfitting. But 15 classification can indeed increase the performance.
\nOur highest private score during the competition was 0.286.
\nI chose the right way to average the confidence value and get a score of 0.290.
\nAnd both public LB and private LB increase at the same time.
My 15 classification uses a very simple model.
\nLabels: 15 classes, A class of an image is labelled 1 if 1 radiologists annotated the class on the image.
\nEfficientnet-b7 (imagenet) 5 fold:
\nimage size 600
\naug:
\nimg = tf.image.random_flip_left_right(img)
\nimg = tf.image.random_flip_up_down(img)
Efficientnet-b7 (noisy-student) 5 fold:
\nimage size 600
\naug:
\nimg = tf.image.random_flip_left_right(img)
\nimg = tf.image.random_flip_up_down(img)
densenet201 5 fold:
\nimage size 600
\nPre-training on the nih dataset
\naug:
\nimg = tf.image.random_flip_left_right(img)
\nimg = tf.image.random_flip_up_down(img)
\nimg = tf.image.random_hue(img, 0.2)
\nimg = tf.image.random_brightness(img, 0.2)
Finally, the confidence value is weighted average.
", "rawMarkdown": "[0.290 Private Score notebook](https://www.kaggle.com/h053473666/0-286-norm-15class?scriptVersionId=58384072)\nIn this competition we shake down from 2% to 11%. I am very sorry to my teammates @qiaoyuanfang. He did not win any medals because of me. He provided a lot of high-quality notebooks, and most of the results were in the silver medal. Because I am overly pursuing LB, and I adjusted the 15 classification parameters according to LB. This leads to serious overfitting. But 15 classification can indeed increase the performance.\nOur highest private score during the competition was 0.286.\nI chose the right way to average the confidence value and get a score of 0.290.\nAnd both public LB and private LB increase at the same time.\n\nMy 15 classification uses a very simple model.\nLabels: 15 classes, A class of an image is labelled 1 if 1 radiologists annotated the class on the image.\nEfficientnet-b7 (imagenet) 5 fold:\nimage size 600\naug:\n`img = tf.image.random_flip_left_right(img)`\n`img = tf.image.random_flip_up_down(img)`\n\nEfficientnet-b7 (noisy-student) 5 fold:\nimage size 600\naug:\n`img = tf.image.random_flip_left_right(img)`\n`img = tf.image.random_flip_up_down(img)`\n\ndensenet201 5 fold:\nimage size 600\nPre-training on the nih dataset\naug:\n`img = tf.image.random_flip_left_right(img)`\n`img = tf.image.random_flip_up_down(img)`\n`img = tf.image.random_hue(img, 0.2)`\n`img = tf.image.random_brightness(img, 0.2)`\n\nFinally, the confidence value is weighted average.", "votes": 7 }, { "id": 1260561, "postDate": "2021-04-02T08:17:20.237Z", "content": "Could you please explain what do you mean by I adjusted the 15 classification parameters according to LB. ?
Different weights and different thresholds make public LB scores vary greatly.
\nhttps://www.kaggle.com/h053473666/0-284rcnn-norm-14class-1?scriptVersionId=58194040
\n0.298 Public Score
Ensemble your public notebooks(0.211) and @corochann(0.230) public notebooks + 15 classification can reach 0.284 public Score.
\nhttps://www.kaggle.com/h053473666/0318-5?scriptVersionId=57138569
I did not do local verification, but I adjusted the weights and thresholds according to the LB score.
", "rawMarkdown": "I did not do local verification, but I adjusted the weights and thresholds according to the LB score.", "votes": 1 } ] } ], "comments": [ { "id": 1260561, "author_name": "Awsaf", "author_url": "", "post_date": "2021-04-02T08:17:20.237000", "content": "Could you please explain what do you mean by I adjusted the 15 classification parameters according to LB. ?
Different weights and different thresholds make public LB scores vary greatly.
\nhttps://www.kaggle.com/h053473666/0-284rcnn-norm-14class-1?scriptVersionId=58194040
\n0.298 Public Score
Ensemble your public notebooks(0.211) and @corochann(0.230) public notebooks + 15 classification can reach 0.284 public Score.
\nhttps://www.kaggle.com/h053473666/0318-5?scriptVersionId=57138569
I did not do local verification, but I adjusted the weights and thresholds according to the LB score.
", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "1258640": "[0.290 Private Score notebook](https://www.kaggle.com/h053473666/0-286-norm-15class?scriptVersionId=58384072)\nIn this competition we shake down from 2% to 11%. I am very sorry to my teammates @qiaoyuanfang. He did not win any medals because of me. He provided a lot of high-quality notebooks, and most of the results were in the silver medal. Because I am overly pursuing LB, and I adjusted the 15 classification parameters according to LB. This leads to serious overfitting. But 15 classification can indeed increase the performance.\nOur highest private score during the competition was 0.286.\nI chose the right way to average the confidence value and get a score of 0.290.\nAnd both public LB and private LB increase at the same time.\n\nMy 15 classification uses a very simple model.\nLabels: 15 classes, A class of an image is labelled 1 if 1 radiologists annotated the class on the image.\nEfficientnet-b7 (imagenet) 5 fold:\nimage size 600\naug:\n`img = tf.image.random_flip_left_right(img)`\n`img = tf.image.random_flip_up_down(img)`\n\nEfficientnet-b7 (noisy-student) 5 fold:\nimage size 600\naug:\n`img = tf.image.random_flip_left_right(img)`\n`img = tf.image.random_flip_up_down(img)`\n\ndensenet201 5 fold:\nimage size 600\nPre-training on the nih dataset\naug:\n`img = tf.image.random_flip_left_right(img)`\n`img = tf.image.random_flip_up_down(img)`\n`img = tf.image.random_hue(img, 0.2)`\n`img = tf.image.random_brightness(img, 0.2)`\n\nFinally, the confidence value is weighted average.", "1260561": "Could you please explain what do you mean by `I adjusted the 15 classification parameters according to LB.` ?" } }