Since I don't have much time, I'm dropping out of this competition.
\nBefore that I would like to share some stuff I did and would be great if you can share yours as well!
First, I tried to create a Image Level PE detection network, which is a simple binary classifier.
\n(I was thinking of training an another exam-level network for the exam level diagnosis.)
To start with, tested out with some good-old effnets, but the accuracy is not so good compared to baselines.
\nThe CV is around 0.31, even with the best model (haven't submitted to LB).
| \n | Accuracy | \n
|---|---|
| Baseline (All zeros) | \n94.3 | \n
| Baseline2 (OsciArt) | \n?(around 97?) | \n
| Effnet-b0 | \n94.5 | \n
| Effnet-b3 | \n96.7 | \n
| Effnet-b5 | \n97.0 | \n
Should we improve the exam-level accuracy instead..?
\nTry b7? (will take couple of days to train..)
\nCan we get a better accuracy with higher resolution images? (Ian's is only 256*256)
\nShould try TPU with that, given the tfrecords published :)
\nUsing z_pos in image-level CNN may be required.
Since I don't have much time, I'm dropping out of this competition.
\nBefore that I would like to share some stuff I did and would be great if you can share yours as well!
First, I tried to create a Image Level PE detection network, which is a simple binary classifier.
\n(I was thinking of training an another exam-level network for the exam level diagnosis.)
To start with, tested out with some good-old effnets, but the accuracy is not so good compared to baselines.
\nThe CV is around 0.31, even with the best model (haven't submitted to LB).
| \n | Accuracy | \n
|---|---|
| Baseline (All zeros) | \n94.3 | \n
| Baseline2 (OsciArt) | \n?(around 97?) | \n
| Effnet-b0 | \n94.5 | \n
| Effnet-b3 | \n96.7 | \n
| Effnet-b5 | \n97.0 | \n
Should we improve the exam-level accuracy instead..?
\nTry b7? (will take couple of days to train..)
\nCan we get a better accuracy with higher resolution images? (Ian's is only 256*256)
\nShould try TPU with that, given the tfrecords published :)
\nUsing z_pos in image-level CNN may be required.
Since the dataset is really imbalanced, acc doesn't mean much, do you have a recall or f1 score?. Have you tried training on just the patients with a positive exam for pe?
\nOn another note, this post getting zero comments is seriously one of the saddest things I ever witnessed. Don't get me wrong, you have all the right to stay silent and keep your solution to your selves. but I mean come on people.
", "rawMarkdown": "Since the dataset is *really* imbalanced, acc doesn't mean much, do you have a recall or f1 score?. Have you tried training on just the patients with a positive exam for pe?\n\nOn another note, this post getting zero comments is seriously one of the saddest things I ever witnessed. Don't get me wrong, you have all the right to stay silent and keep your solution to your selves. but I mean come on people." } ], "comments": [ { "id": 1057511, "author_name": "DarkCube", "author_url": "", "post_date": "2020-10-22T18:00:42.660000", "content": "Since the dataset is really imbalanced, acc doesn't mean much, do you have a recall or f1 score?. Have you tried training on just the patients with a positive exam for pe?
\nOn another note, this post getting zero comments is seriously one of the saddest things I ever witnessed. Don't get me wrong, you have all the right to stay silent and keep your solution to your selves. but I mean come on people.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1049014": "Since I don't have much time, I'm dropping out of this competition.\nBefore that I would like to share some stuff I did and would be great if you can share yours as well!\n\nFirst, I tried to create a Image Level PE detection network, which is a simple binary classifier.\n(I was thinking of training an another exam-level network for the exam level diagnosis.)\n\nTo start with, tested out with some good-old effnets, but the accuracy is not so good compared to baselines.\nThe CV is around 0.31, even with the best model (haven't submitted to LB).\n\n* Notes\nI used Ian's processed jpg data to save time.\nCan upload my training pipeline if needed ([it's based on the previous PANDA competition](https://github.com/kentaroy47/Kaggle-PANDA-1st-place-solution))\nToggling with augumentations, mixup, cutmix did not help much.\nI also tried oversampling positive data, did not help much.\nTrained for 30 epochs with cos annealing. bs=32, lr=1e-4, adam.\n\n| | Accuracy |\n|:--------------------:|:--------:|\n| Baseline (All zeros) | 94.3 |\n| Baseline2 (OsciArt) | ?(around 97?) |\n| Effnet-b0 | 94.5 |\n| Effnet-b3 | 96.7 |\n| Effnet-b5 | 97.0 |\n\n## Thoughts\nShould we improve the exam-level accuracy instead..?\nTry b7? (will take couple of days to train..)\nCan we get a better accuracy with higher resolution images? (Ian's is only 256*256)\nShould try TPU with that, given the tfrecords published :)\nUsing z_pos in image-level CNN may be required.", "1057511": "Since the dataset is *really* imbalanced, acc doesn't mean much, do you have a recall or f1 score?. Have you tried training on just the patients with a positive exam for pe?\n\nOn another note, this post getting zero comments is seriously one of the saddest things I ever witnessed. Don't get me wrong, you have all the right to stay silent and keep your solution to your selves. but I mean come on people." } }