![\"\"](\"https://i.ibb.co/LkJGQmW/Selection-217.png\")
i am preparing starter kit over the next few weeks.
\nIf you have recommended papers or methods, please let me know. I will include it in the starter kit.
google drive link for code, model, train, log, etc:
\nhttps://drive.google.com/drive/folders/1oXfTayH6eRG36DAKmvwVyFE5cmPR6K6W?usp=sharing
version.1 2020-14-jun: completed and ready for download!
\nafter rescore
\n
version.2: … in preparation …
\ndue date: 21-jun
(let me try object detection first and decide if i should focus more on image classification or object detection)
customize bifpn from https://github.com/rwightman/efficientdet-pytorch
use efficientNet v2 as backbone: https://arxiv.org/pdf/2104.00298.pdf
use adversial-prop-det for training: https://arxiv.org/pdf/2103.13886.pdf
reference: https://amaarora.github.io/2021/01/13/efficientdet-pytorch.html
\nversion.1 shows that pure classification does not work better than one with joint aux loss.
\nthe reason is overfitting
hence we need to implement something like this ….
\n![\"\"](\"https://i.ibb.co/XtywGb9/Selection-239.png\")
later version
\nyolo-2021 : https://twitter.com/alexeyab84/status/1398443022619189248?s=20
reference: https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229615
\nit would be interesting to use multi-task yolo by extracting extra labels from radiologist report in external data …
\n![\"\"](\"https://i.ibb.co/sgKhJwF/Selection-289.png\")
etc (random notes for myself):
\nFB MoCo self-supervised contrastive learning?
\ntransformers ????
\nmeta-pseudo label adapted for detection, transMIL (transformer MIL, ranking as a \"sorted\" sequence prediction)
\nSemi-Supervised AUC Optimization without Guessing Labels of Unlabeled Data
\ntrasnformer yolo https://www.gitmemory.com/issue/ultralytics/yolov5/2329/808852867
\nre-ranking trasnformer for ensemble
\nresize artificats, adversial noise
\nConstrained Optimization to Train Neural Networks
\non Critical and Under-Represented Classes
transformer hybrid:
\nhttps://www.kaggle.com/bamps53/lb0-87-part-of-3rd-place-solution?scriptVersionId=65265405
\nhttps://www.kaggle.com/c/bms-molecular-translation/discussion/243932
why to convert weak segmentation ground truth (bbox) into a stronger one? … pool most probable pixels? lung prior? read papers on the weak labels for segmentation.
\nkeys to winning
\na better loss function for mAP image classification and localization (same or different one?) …
someone can recommend a ranking-based loss or large margin function for mAP?
use of external data (especially MIDRC-RICORD Data, Kaggle previous data etc)
small data size is one of the main problems in this challenge. Your results may be unstable (overfitting, etc) … you need to think of a way to solve this, e.g. deep/extra supervision, augmentation
i am preparing starter kit over the next few weeks.
\nIf you have recommended papers or methods, please let me know. I will include it in the starter kit.
google drive link for code, model, train, log, etc:
\nhttps://drive.google.com/drive/folders/1oXfTayH6eRG36DAKmvwVyFE5cmPR6K6W?usp=sharing
version.1 2020-14-jun: completed and ready for download!
\nafter rescore
\n
version.2: … in preparation …
\ndue date: 21-jun
(let me try object detection first and decide if i should focus more on image classification or object detection)
customize bifpn from https://github.com/rwightman/efficientdet-pytorch
use efficientNet v2 as backbone: https://arxiv.org/pdf/2104.00298.pdf
use adversial-prop-det for training: https://arxiv.org/pdf/2103.13886.pdf
reference: https://amaarora.github.io/2021/01/13/efficientdet-pytorch.html
\nversion.1 shows that pure classification does not work better than one with joint aux loss.
\nthe reason is overfitting
hence we need to implement something like this ….
\n
later version
\nyolo-2021 : https://twitter.com/alexeyab84/status/1398443022619189248?s=20
reference: https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229615
\nit would be interesting to use multi-task yolo by extracting extra labels from radiologist report in external data …
\n
etc (random notes for myself):
\nFB MoCo self-supervised contrastive learning?
\ntransformers ????
\nmeta-pseudo label adapted for detection, transMIL (transformer MIL, ranking as a \"sorted\" sequence prediction)
\nSemi-Supervised AUC Optimization without Guessing Labels of Unlabeled Data
\ntrasnformer yolo https://www.gitmemory.com/issue/ultralytics/yolov5/2329/808852867
\nre-ranking trasnformer for ensemble
\nresize artificats, adversial noise
\nConstrained Optimization to Train Neural Networks
\non Critical and Under-Represented Classes
transformer hybrid:
\nhttps://www.kaggle.com/bamps53/lb0-87-part-of-3rd-place-solution?scriptVersionId=65265405
\nhttps://www.kaggle.com/c/bms-molecular-translation/discussion/243932
why to convert weak segmentation ground truth (bbox) into a stronger one? … pool most probable pixels? lung prior? read papers on the weak labels for segmentation.
\nkeys to winning
\na better loss function for mAP image classification and localization (same or different one?) …
someone can recommend a ranking-based loss or large margin function for mAP?
use of external data (especially MIDRC-RICORD Data, Kaggle previous data etc)
small data size is one of the main problems in this challenge. Your results may be unstable (overfitting, etc) … you need to think of a way to solve this, e.g. deep/extra supervision, augmentation
I really appreciate your work and thanks for sharing information
", "rawMarkdown": "I really appreciate your work and thanks for sharing information", "votes": 5 }, { "id": 1363385, "postDate": "2021-06-24T06:11:58.033Z", "content": "after numerous experiments:
\nconfirmed that aux loss using the mask can improve results significantly.
\nI can get local CV up to 0.39+ for single fold using different network (e.g. transformer), different pooling (attention-based, global average), different loss (lovasz, sigmoid, softamx, more weighing in aux loss, hard mining top +ve/-ve pixels etc)
freezing at fine tuning
increasing difficulty at training and fine-tuning
but it is sometimes unstable (i.e. different methods improve results for different folds/subsets. I can't find a method that improves for all cases)
this is because the mask is \"approximate\". we need a way to better refine the labeling of the pixel to mark them as indicating lung opacity/non-opacity or other.
you can think of pixels inside the box as \"likely to contain lung opacity\", but \"not all\" pixels are like that (e.g. box contains non-lung region). There is a similar argument for pixels that is outside the box. you need a pooling loss.
the number of train images is too small. Strong network overfits easily
I stacked top 50 object detections boxes with confidence to generate the \"finer\" mask. Regions with high-confidence opacity will be brighter whereas negative sample will be darker but not completely blank. Almost perfectly filtered distractive objects such as letter and metals. But got same result for appearance classification. Without external data, I doubt whether multitasking model will result in an improvement.
", "rawMarkdown": "I stacked top 50 object detections boxes with confidence to generate the \"finer\" mask. Regions with high-confidence opacity will be brighter whereas negative sample will be darker but not completely blank. Almost perfectly filtered distractive objects such as letter and metals. But got same result for appearance classification. Without external data, I doubt whether multitasking model will result in an improvement." }, { "id": 1365025, "postDate": "2021-06-25T11:55:16.643Z", "content": "\n\nbut it is sometimes unstable (i.e. different methods improve results for different folds/subsets. I can't find a method that improves for all cases)
\n
Exactly! In my case too even average of folds is not the ideal improvement on leaderboard, I select my final best model as average of all my CV folds and LB. Even changing the seed changes the scores a lot (made it only for worse in my trails).
\n\n\nyou can think of pixels inside the box as \"likely to contain lung opacity\", but \"not all\" pixels are like that (e.g. box contains non-lung region). There is a similar argument for pixels that is outside the box. you need a pooling loss.
\n
I am using something like pooling loss working out great for me!
\n\n\nthe number of train images is too small. Strong network overfits easily
\n
I wonder if the extra data pretraining might solve that for us although that might mean even stronger augmentations or higher drop rates to make it not overfit.
", "rawMarkdown": "> but it is sometimes unstable (i.e. different methods improve results for different folds/subsets. I can't find a method that improves for all cases)\n\nExactly! In my case too even average of folds is not the ideal improvement on leaderboard, I select my final best model as average of all my CV folds and LB. Even changing the seed changes the scores a lot (made it only for worse in my trails).\n\n> you can think of pixels inside the box as \"likely to contain lung opacity\", but \"not all\" pixels are like that (e.g. box contains non-lung region). There is a similar argument for pixels that is outside the box. you need a pooling loss.\n\nI am using something like pooling loss working out great for me!\n\n> the number of train images is too small. Strong network overfits easily\n\nI wonder if the extra data pretraining might solve that for us although that might mean even stronger augmentations or higher drop rates to make it not overfit." }, { "id": 1365783, "postDate": "2021-06-26T05:56:54.817Z", "content": "I trained a MoCo model on CheXpert dataset(license issue, only for experiments), MoCo models was trained well with top 1 accuracy over 90%(k=65536 and 4096). But finetuned on this dataset doesn't improve the performance compare to imagenet pretrained weights. Naive knn on MoCo representations doesn't work as well, as expected.
", "rawMarkdown": "I trained a MoCo model on CheXpert dataset(license issue, only for experiments), MoCo models was trained well with top 1 accuracy over 90%(k=65536 and 4096). But finetuned on this dataset doesn't improve the performance compare to imagenet pretrained weights. Naive knn on MoCo representations doesn't work as well, as expected." } ] }, { "id": 1462936, "postDate": "2021-08-10T03:39:04.480Z", "content": "to make this thread complete, i have updated:
\n[1] notebook to show interference code and results
\n(My submission couldn't complete in time and hence I missed the submission deadline)
\nhttps://www.kaggle.com/hengck23/final-v-00-private-public-score-0-615-0-628?scriptVersionId=70933214
this is bare minium model (only one k-fold model for study classification and one k-fold detection model)
\nwith decent results
Private Score 0.615
\nPublic Score 0.628
![\"https://i.ibb.co/pb1SB85/Selection-665.png\"](\"https://i.ibb.co/pb1SB85/Selection-665.png\")
[2] code and trained model for effcienetDet (see folder \"2021-08-11-effdet\")
\nhttps://drive.google.com/drive/folders/16JzkuBpW5cwEexM5awq0wOK6t9gYLeKc
[3] more results on effcienetDet
\n![\"https://i.ibb.co/vBr88pr/Selection-668.png\"](\"https://i.ibb.co/vBr88pr/Selection-668.png\")
voc12 mAP computation, python code: https://github.com/Cartucho/mAP/blob/master/main.py
", "rawMarkdown": "voc12 mAP computation, python code: https://github.com/Cartucho/mAP/blob/master/main.py", "votes": 3, "replies": [ { "id": 1362347, "postDate": "2021-06-23T11:46:44.013Z", "content": "![\"https://insidebigdata.com/wp-content/uploads/2021/03/arXiv_2021_Feb6.png\"](\"https://insidebigdata.com/wp-content/uploads/2021/03/arXiv_2021_Feb6.png\")
Constrained Optimization for Training Deep Neural Networks Under Class Imbalance
\nhttps://arxiv.org/pdf/2102.12894v1.pdf
we love neighbors:
\nhttps://www.kaggle.com/c/siim-covid19-detection/discussion/240250
\n\"FIGURE 7: Lung zones: proposed methodology for dividing a frontal CXR into 3 zones per lung (total of 6 zones). The upper zone extends from the apices to the superior portion of the hilum. The mid zone spans the space between the superior and inferior hilar margins. The lower zone extends from the inferior hilar margins to the costophrenic sulci.\"
\n![\"\"](\"https://images.journals.lww.com/thoracicimaging/ArticleViewerPreview.00005382-202011000-00004.F7.jpeg\")
see also:
\n![\"https://i.ibb.co/2dWFsCr/Selection-323.png\"](\"https://i.ibb.co/2dWFsCr/Selection-323.png\")
![\"https://i.ibb.co/LvGJQVV/Selection-324.png\"](\"https://i.ibb.co/LvGJQVV/Selection-324.png\")
Note: bbox aux loss help image classification.
\nconversely, 4-class image label aux loss should help opacity localization/detection.
\n(you need an aux-loss yolo or efficientDet)
modified meta pesudo label
\n![\"\"](\"https://i.ibb.co/6P7dnsn/Selection-317.png\")
insprations from recent cvpr 2021
\n[CVPR 2021] Background Splitting: Finding Rare Categories in a Sea of Background
![\"https://i.ibb.co/SVKfTvG/Selection-342.png\"](\"https://i.ibb.co/SVKfTvG/Selection-342.png\")
Thanks this is interesting, however in this case the background images are pseudo-labeled by an imagenet-pretrained model. Can this apply to our case.. ?
", "rawMarkdown": "Thanks this is interesting, however in this case the background images are pseudo-labeled by an imagenet-pretrained model. Can this apply to our case.. ?" }, { "id": 1361158, "postDate": "2021-06-22T16:00:35.597Z", "content": "![\"\"](\"https://i.ibb.co/Kx1nZyT/boxes.png\")
\nget your best object detection model trained with default background box 0,0,0,0 -> label \"negative\" samples and make the predicted opacity box as background -> train again. From the image above (covid negative), you can see that vanilla model does makes some \"reasonable\" but wrong predictions on the negative image, but it also looks at the foreign objects such as the \"BIPE\" and spine. [0,0,0,0] fake background box tells the model \"your prediction is very wrong!\", but doesn't tell how & where to improve. Without a proper regularization, model will be easily overfitted to looking at irrelevant objects, like metal, heart, patient pose, etc. Pseudolabelling background with good boxes instead of 0,0,0,0 fake box (should) help model understand the corresponding region is not opacity and makes the loss smoother.
![\"https://i.ibb.co/LxL1ybP/Selection-343.png\"](\"https://i.ibb.co/LxL1ybP/Selection-343.png\")
\nBuilding High Performance Chest X-Ray Classification Models | Talk 1 | CVPR 2021 | WIANLP Tutorial
Topic VI: Data Efficient Deep Learning for Medical Image Analysis | CVPR 2021 | VITA Workshop
\nhttps://www.youtube.com/watch?v=HnBvpwP0Wp4
success in experiments in transformer opens up a world of new possibilities …
\ncode coming up ….
![\"https://i.ibb.co/2FYmFh6/Selection-335.png\"](\"https://i.ibb.co/2FYmFh6/Selection-335.png\")
\n![\"https://i.ibb.co/VDnGj2n/Selection-337.png\"](\"https://i.ibb.co/VDnGj2n/Selection-337.png\")
\n![\"https://i.ibb.co/hfnyBM1/Selection-338.png\"](\"https://i.ibb.co/hfnyBM1/Selection-338.png\")
\n![\"https://i.ibb.co/6WnyF6X/Selection-339.png\"](\"https://i.ibb.co/6WnyF6X/Selection-339.png\")
Have you share the code with transformer?
", "rawMarkdown": "Have you share the code with transformer?" } ] }, { "id": 1349610, "postDate": "2021-06-15T00:50:44.503Z", "content": "important conclusion from version.1 2020-14-jun:
\n(do not take any words for granted. i encourage you to verify the claims below and discuss here)
Confirmed. Different backbones of EffNet with aux loss consistently resulted in CV mAP 0.38x. Model get overfitted easily without aux loss.
", "rawMarkdown": "Confirmed. Different backbones of EffNet with aux loss consistently resulted in CV mAP 0.38x. Model get overfitted easily without aux loss." }, { "id": 1350981, "postDate": "2021-06-16T01:50:15.457Z", "content": "you may also want to check that at CV mAP 0.38x. , the validation and training loss gap is not large (much better than the case without aux loss). this indicates aux loss improve generalisation.
\nbut the loss curves are bumpy and fluctuating. there is still some room for improvement if you can smooth and stabilise the curve
", "rawMarkdown": "you may also want to check that at CV mAP 0.38x. , the validation and training loss gap is not large (much better than the case without aux loss). this indicates aux loss improve generalisation.\n\nbut the loss curves are bumpy and fluctuating. there is still some room for improvement if you can smooth and stabilise the curve", "votes": 1 }, { "id": 1351581, "postDate": "2021-06-16T12:32:59.537Z", "content": "I can confirm that augmentation helps a lot. But regard aux loss, my resnet101 model(224x224) started to overfit heavily when I add aux loss. The training accuracy is nearly 95% with aux loss and 70% without aux loss, validation accuracy is about 67% percent in both cases. I tried using all images and ignore all images(except negative class) without boxes, still no luck ;(
\nI'll try to use your input pipeline to investigate this further, maybe segmentation head only works better only if your augmentation is strong enough?
resnet perform worse than efficient net in my experiments
", "rawMarkdown": "resnet perform worse than efficient net in my experiments", "votes": 1 }, { "id": 1351728, "postDate": "2021-06-16T14:51:38.603Z", "content": "@artnotintelligence you probably need to specify high drop_rate and drop_path_rate in the effnet / resnet, especially when lr is low. In Heng's kit, they are both 0.2, but for big backbone I set drop_rate = 0.5 to combat overfitting, but model will still overfit when mAP is close to 0.385. This is the limit of simple aux loss strategy.
", "rawMarkdown": "@artnotintelligence you probably need to specify high drop_rate and drop_path_rate in the effnet / resnet, especially when lr is low. In Heng's kit, they are both 0.2, but for big backbone I set drop_rate = 0.5 to combat overfitting, but model will still overfit when mAP is close to 0.385. This is the limit of simple aux loss strategy.", "votes": 1 }, { "id": 1351808, "postDate": "2021-06-16T16:07:19.397Z", "content": "I add dropout rate up to 0.5 along with l2 regularization loss, which doesn't help avoid overfitting.
", "rawMarkdown": "I add dropout rate up to 0.5 along with l2 regularization loss, which doesn't help avoid overfitting. " }, { "id": 1351850, "postDate": "2021-06-16T16:40:34.700Z", "content": "\"This is the limit of simple aux loss strategy.\"
\nyou can predict the lung segmentation or pixel values of the related masked region.
\n![\"\"](\"https://i.ibb.co/ZTbK99K/Selection-265.png\")
\n![\"\"](\"https://i.ibb.co/fk6LsBX/Selection-264.png\")
![\"\"](\"https://i.ibb.co/RNM9cFV/Selection-266.png\")
input to the green \"predict the value of cross entropy loss\" model could be array of pixel differences of the predicted and truth of mask region.
\nwe bypass the step to obtain pseudo label … we just need the \"pseudo label loss\" for backprop
", "rawMarkdown": "\n\ninput to the green \"predict the value of cross entropy loss\" model could be array of pixel differences of the predicted and truth of mask region.\n\nwe bypass the step to obtain pseudo label ... we just need the \"pseudo label loss\" for backprop", "votes": 2 }, { "id": 1353291, "postDate": "2021-06-17T03:15:37.277Z", "content": "I don't quite understand the last part, are you saying that , for example, if softmax(logits) = [0.1, 0.1, 0.2, 0.6], then we use pseudo label = [0, 0, 0, 1]
\nas gt label to produce CE loss, then further regularize CE loss with predicted CE loss?
no. you don't predict the pseudo label.
\nyou can predict the CE loss value (of pseudo label) directly
@houndcl i think @phalanx Unet aux loss is better than mine. Maybe it can hit 0.395
", "rawMarkdown": "@houndcl i think @phalanx Unet aux loss is better than mine. Maybe it can hit 0.395", "votes": 1 } ] }, { "id": 1347120, "postDate": "2021-06-13T02:24:02.733Z", "content": "archive for old post
\ni am preparing starter kit over the next few weeks.
\nIf you have recommended papers or methods, please let me know. I will include it in the starter kit.
What is desired:
\nold methods that have proved to work on xray images, etc
\nnew methods that may show interesting results (e.g. transformer, weak/semi-supervised learning)
\npaper that you want to understand learn, but there is no open source to verify if your understanding is correct
\ntraining/inference speedup
\ne.g. …
\nOnce i select the method, the starter kits will include:
reference code for training and inference
\nintermediate trained model
\nlogfile of training for reference and checking implementation, etc
\nexperiment results
\nsome training/hyperparameters tricks. etc
\nNote: i am currently busy with kaggle BMS molecule image-text translation. I will be more active here after BMS is over
the starter kit post will be something like this: https://www.kaggle.com/c/bms-molecular-translation/discussion/231190
", "rawMarkdown": "archive for old post \n\n----\ni am preparing starter kit over the next few weeks.\nIf you have recommended papers or methods, please let me know. I will include it in the starter kit.\n\nWhat is desired:\n\nold methods that have proved to work on xray images, etc\nnew methods that may show interesting results (e.g. transformer, weak/semi-supervised learning)\npaper that you want to understand learn, but there is no open source to verify if your understanding is correct\ntraining/inference speedup\ne.g. …\nOnce i select the method, the starter kits will include:\n\nreference code for training and inference\nintermediate trained model\nlogfile of training for reference and checking implementation, etc\nexperiment results\nsome training/hyperparameters tricks. etc\nNote: i am currently busy with kaggle BMS molecule image-text translation. I will be more active here after BMS is over\n\nthe starter kit post will be something like this: https://www.kaggle.com/c/bms-molecular-translation/discussion/231190", "votes": 4 }, { "id": 1367336, "postDate": "2021-06-27T15:49:09.230Z", "content": "![\"https://i.ibb.co/Y7BbmCN/Selection-425.png\"](\"https://i.ibb.co/Y7BbmCN/Selection-425.png\")
![\"https://i.ibb.co/42YxkwH/Selection-450.png\"](\"https://i.ibb.co/42YxkwH/Selection-450.png\")
![\"https://i.ibb.co/5R3BkWv/Selection-449.png\"](\"https://i.ibb.co/5R3BkWv/Selection-449.png\")
![\"https://i.ibb.co/KbQsYMV/Selection-448.png\"](\"https://i.ibb.co/KbQsYMV/Selection-448.png\")
![\"https://i.ibb.co/55XgDND/Selection-454.png\"](\"https://i.ibb.co/55XgDND/Selection-454.png\")
![\"\"](\"https://i.ibb.co/xgh6zvn/agree.png\")
\nThe real world confusion matrix from RICORD dataset. The \"typical appearance\" data seems cleaner in this competition. The accurate prediction of \"atypical\" and \"indeterminate\" appearances remain challenging. Atypical class may be slightly easier to improve, by integrating external dataset such as MIMIC and NIH. There are plenty of images with mass | nodule | effusion | pneumothorax | edema etc labels.
What is rescore?
", "rawMarkdown": "What is rescore?", "votes": 1, "replies": [ { "id": 1380875, "postDate": "2021-07-08T11:54:31.467Z", "content": "At the beginning of the competition there was a bug in the leaderboard and once the bug was fixed all submissions were run again and \"rescored\" so they got a new score without the bug
", "rawMarkdown": "At the beginning of the competition there was a bug in the leaderboard and once the bug was fixed all submissions were run again and \"rescored\" so they got a new score without the bug" } ] }, { "id": 1374603, "postDate": "2021-07-03T12:53:52.260Z", "content": "i suddenly has an idea:
\nIn this research, we proposed an ACGAN based model
\ncalled CovidGAN that generates synthetic CXR images to
\nenlarge the dataset and to improve the performance of CNN
\nin COVID-19 detection. The research is implemented on a
\ndataset with 403 COVID-CXR images and 721 Normal-CXR
\nimages
https://arxiv.org/pdf/2103.05094.pdf
", "rawMarkdown": "In this research, we proposed an ACGAN based model\ncalled CovidGAN that generates synthetic CXR images to\nenlarge the dataset and to improve the performance of CNN\nin COVID-19 detection. The research is implemented on a\ndataset with 403 COVID-CXR images and 721 Normal-CXR\nimages\n\nhttps://arxiv.org/pdf/2103.05094.pdf", "votes": 1 }, { "id": 1366297, "postDate": "2021-06-26T16:22:40.367Z", "content": "![\"https://i.ibb.co/xfvbwKF/Selection-393.png\"](\"https://i.ibb.co/xfvbwKF/Selection-393.png\")
\nhttps://www.youtube.com/channel/UCOkkljs06NPPkjNysCdQV4w/videos
channel to replace patch as token in trasnformer
\nCross-Covariance Image Transformer (XCiT)
\nhttps://www.youtube.com/watch?v=g08NkNWmZTA
\nhttps://github.com/facebookresearch/xcit
https://arxiv.org/pdf/2106.13112.pdf
\nVolo+yolo?
![\"https://i.ibb.co/ngYDtJM/Selection-352.png\"](\"https://i.ibb.co/ngYDtJM/Selection-352.png\")
\nhttps://arxiv.org/pdf/2104.10858.pdf
\nAll Tokens Matter: Token Labeling for Training Better Vision Transformers
![\"\"](\"https://i.ibb.co/s11Vc7Q/Selection-316.png\")
\nhttps://www.youtube.com/watch?v=0TwZfRcqhbI&list=PL6liSIqFR4BUr7F4pL6aoucjkV3HYZ48J
\nhttps://github.com/facebookresearch/unbiased-teacher
I tried similiar setup but my mAP at study level was lower with segmentation head. What's your score without segmentation loss?
", "rawMarkdown": "I tried similiar setup but my mAP at study level was lower with segmentation head. What's your score without segmentation loss?", "votes": 1 }, { "id": 1347883, "postDate": "2021-06-13T14:45:53.973Z", "content": "Hi @hengck23, what was your score without the aux loss ?
", "rawMarkdown": "Hi @hengck23, what was your score without the **aux loss** ?", "votes": 1, "replies": [ { "id": 1350072, "postDate": "2021-06-15T08:45:30.530Z", "content": "What is aux loss ? I'm googling it and not finding anything!
", "rawMarkdown": "What is aux loss ? I'm googling it and not finding anything!" }, { "id": 1350176, "postDate": "2021-06-15T10:09:45.767Z", "content": "@josephamigo I believe aux loss is auxiliary loss. It helps to reduce the vanishing gradient problem for earlier layers, stabilizes the training and is used as regularization. It's only used for training and not for inference.
\nReference: https://stats.stackexchange.com/questions/304699/what-is-auxiliary-loss-as-mentioned-in-pspnet-paper
", "rawMarkdown": "@josephamigo I believe aux loss is auxiliary loss. It helps to reduce the vanishing gradient problem for earlier layers, stabilizes the training and is used as regularization. It's only used for training and not for inference.\n\nReference: https://stats.stackexchange.com/questions/304699/what-is-auxiliary-loss-as-mentioned-in-pspnet-paper", "votes": 5 }, { "id": 1350199, "postDate": "2021-06-15T10:32:51.100Z", "content": "Thank you very much!!
", "rawMarkdown": "Thank you very much!!", "votes": 1 } ] }, { "id": 1334679, "postDate": "2021-06-03T17:27:06.223Z", "content": "cool just waiting for your discussion and ideas @hengck23 and starting baseline of @yasufuminakama and saturday.
", "rawMarkdown": "cool just waiting for your discussion and ideas @hengck23 and starting baseline of @yasufuminakama and saturday.", "votes": 1 }, { "id": 1319215, "postDate": "2021-05-23T03:11:15.323Z", "content": "Not really a method, but a good summary. Also, it mentions a way to speed up a training procedure.
\nCheXtransfer: Performance and Parameter Efficiency of ImageNet Models for Chest X-Ray Interpretation
\nhttps://arxiv.org/abs/2101.06871
Welcome! I will learn from you again.
", "rawMarkdown": "Welcome! I will learn from you again.", "votes": 1 }, { "id": 1359001, "postDate": "2021-06-21T02:02:50.340Z", "content": "lung mask:
\nhttps://github.com/v7labs/covid-19-xray-dataset
\nhttps://github.com/ieee8023/covid-chestxray-dataset
https://github.com/raghavian/lungVAE
\nhttps://arxiv.org/pdf/2104.05892.pdf
lung opacity:
\nhttps://arxiv.org/pdf/2002.02497.pdf
\nhttps://github.com/mlmed/torchxrayvision
\nhttps://openaccess.thecvf.com/content_CVPRW_2020/papers/w22/Gabruseva_Deep_Learning_for_Automatic_Pneumonia_Detection_CVPRW_2020_paper.pdf
![\"\"](\"https://i.ibb.co/Y2SycDk/Selection-322.png\")
Hello,How can I get this picture?
", "rawMarkdown": "Hello,How can I get this picture?" }, { "id": 1358760, "postDate": "2021-06-20T17:52:43.727Z", "content": "Vision Transformer using Low-level Chest X-ray Feature Corpus for COVID-19 Diagnosis and Severity Quantification
", "rawMarkdown": "\n\nVision Transformer using Low-level Chest X-ray Feature Corpus for COVID-19 Diagnosis and Severity Quantification", "votes": 2 }, { "id": 1358997, "postDate": "2021-06-21T01:46:33.963Z", "content": "Thanks for your ideas!
", "rawMarkdown": "Thanks for your ideas!" }, { "id": 1359004, "postDate": "2021-06-21T02:09:05.740Z", "content": "PCAM pooling
\n![\"\"](\"https://github.com/jfhealthcare/Chexpert/raw/master/PCAM.png\")
\nhttps://github.com/jfhealthcare/Chexpert
In run_train.py(also in run_submit.py), there seems a bug in calculating the topk accuracy,
\npredict = probability.argsort(-1)[::-1] should be predict = probability.argsort(-1)[:, ::-1]?
good catch.
\nthe same occurs in do_valid() function in run_train.py.
Thanks
", "rawMarkdown": "good catch.\nthe same occurs in do\\_valid() function in run\\_train.py.\n\nThanks\n", "votes": 1 }, { "id": 1357541, "postDate": "2021-06-19T18:44:03.557Z", "content": "nice catch, Anii thanks!!
\n1) BTW do you also get NaN in map score ?
2) also a minor point in run_train.py (affects only the logging, code runs fine)
batch_loss = np.array([loss0.item(), loss1.item(), loss2.item()])
loss2 is always 0.0 from initialisation. I guess if you want to correct it we need to add before:
\nloss2 = loss0+loss1
after the topk bug is fixed, you will get something like:
\n![\"https://pbs.twimg.com/media/EzEYd9CW8AEQlLh?format=jpg&name=large\"](\"https://pbs.twimg.com/media/EzEYd9CW8AEQlLh?format=jpg&name=large\")
experiment = ['effb3-full-512-mask-v8', 'run_train_2.py']\n\n |---------------- VALID -------------|---- TRAIN/BATCH -------\n epoch | loss map (map*0.6) top1,2 | loss0 loss1 \n----------------------------------------------------------------------\n 0.00 | 1.412 0.257 (0.171) 0.281 0.425 | 0.000 0.000 0.000 | \n 0.16 | 1.103 0.419 (0.279) 0.597 0.755 | 1.276 0.561 0.000 | \n 0.32 | 0.994 0.464 (0.309) 0.612 0.808 | 1.096 0.396 0.000 | \n 0.48 | 1.073 0.453 (0.302) 0.585 0.779 | 1.044 0.302 0.000 | \n 0.64 | 0.925 0.505 (0.337) 0.638 0.812 | 1.045 0.258 0.000 | \n 0.80 | 0.971 0.475 (0.317) 0.634 0.794 | 1.013 0.246 0.000 | \n 0.96 | 0.940 0.488 (0.326) 0.639 0.812 | 1.055 0.232 0.000 | \n 1.13 | 0.949 0.507 (0.338) 0.639 0.815 | 0.990 0.221 0.000 | \n 1.29 | 1.013 0.502 (0.334) 0.585 0.764 | 1.017 0.224 0.000 | \n 1.45 | 0.970 0.489 (0.326) 0.631 0.813 | 0.991 0.211 0.000 | \nthis reveals a few tricks:
\nmap is worse than top1. this is the effect of imbalance class. map is averaged across all classes, i did not apply that for top1
the probing discussion is in https://www.kaggle.com/c/siim-covid19-detection/discussion/244066. You can obtain the prior class probability
if you know how many test samples there are, you can \"de-rank\" the probabiliy. e.g https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/73738
the cost of mistakes is not the same for each sample. It is more costly to make mistakes for rare classes. some kind of cost-aware loss function or post-processing may be able to improve results?
\ntop2 results are pretty good. if you want to post-process results, you may consider \"moving prediction from top2 to top1\" if the \"cost\" of this action justifies itself.
one can think of more ideas if one plots some graphs to analyze the distribution of relative/absolute score of different class
There is a discussion to replace Attention with MLP. I don't know if it can be a good application here.
", "rawMarkdown": "There is a discussion to replace Attention with MLP. I don't know if it can be a good application here.", "votes": 2 }, { "id": 1314181, "postDate": "2021-05-19T02:36:30.963Z", "content": "It’s an honor to explore the philosophy of artificial intelligence with you on another battlefield
", "rawMarkdown": "It’s an honor to explore the philosophy of artificial intelligence with you on another battlefield" }, { "id": 1462191, "postDate": "2021-08-09T18:11:05.583Z", "content": "Do you think inputting the image with only 1 channel as opposed to 3 could increase the score? maybe decrease it? idk. personally, when i used one channel the accuracy reduced
", "rawMarkdown": "Do you think inputting the image with only 1 channel as opposed to 3 could increase the score? maybe decrease it? idk. personally, when i used one channel the accuracy reduced", "votes": -1 }, { "id": 1464120, "postDate": "2021-08-10T13:04:56.283Z", "content": "@hengck23 Thank you for all information you shared for this competition. As a team we used your code, ideas and implemented them. Looking forward for next \"starter kit\" discussion 👀
", "rawMarkdown": "@hengck23 Thank you for all information you shared for this competition. As a team we used your code, ideas and implemented them. Looking forward for next \"starter kit\" discussion 👀" }, { "id": 1393690, "postDate": "2021-07-19T19:54:31.270Z", "content": "Nice work! I'd like to ask what is the \"00007400_model.path\" set as the initial checkpoint. As I tried the training with your original setting (efficient net_b3a) with none initial checkpoint, it reached the best point at 3200 iterations, and if I try with effnetv2, it converges faster, around 600 iterations.
", "rawMarkdown": "Nice work! I'd like to ask what is the \"00007400_model.path\" set as the initial checkpoint. As I tried the training with your original setting (efficient net_b3a) with none initial checkpoint, it reached the best point at 3200 iterations, and if I try with effnetv2, it converges faster, around 600 iterations. ", "replies": [ { "id": 1393708, "postDate": "2021-07-19T20:12:59.367Z", "content": "Another question, I obtained a CV score 0.374 for training with your original setting (effnet_b3a) for fold 1. What is the expected LB score for this single folder submission (with \"none 1 0 0 1 1\")? Thanks!
", "rawMarkdown": "Another question, I obtained a CV score 0.374 for training with your original setting (effnet_b3a) for fold 1. What is the expected LB score for this single folder submission (with \"none 1 0 0 1 1\")? Thanks!" } ] }, { "id": 1379390, "postDate": "2021-07-07T10:26:39.310Z", "content": "hi
\nwhy use efficientnet first 6 layers out put as mask to calculate aux loss?
\nHow do you know use 6 layer output rather than 7 layer output?
hiii i tried your started kit its so good i was able to train but having problems while submission i mean how to get in on kaggle i trained on my system i need to submit in kaggle can u help please
", "rawMarkdown": "hiii i tried your started kit its so good i was able to train but having problems while submission i mean how to get in on kaggle i trained on my system i need to submit in kaggle can u help please" }, { "id": 1359754, "postDate": "2021-06-21T13:53:56.010Z", "content": "Do you expect the multi-task model to work better than the models that solve each task (OD, clf)?
", "rawMarkdown": "Do you expect the multi-task model to work better than the models that solve each task (OD, clf)?" }, { "id": 1354831, "postDate": "2021-06-17T23:28:42.667Z", "content": "Great work. I'll try the aux loss.
", "rawMarkdown": "Great work. I'll try the aux loss.", "replies": [ { "id": 1356696, "postDate": "2021-06-19T07:13:16.947Z", "content": "Hey I have tried segmentation as aux loss the masks generated by the bounding box seem to be not that accurate especially for the typical label and it has not worked for me so update if you somehow make it work
", "rawMarkdown": "Hey I have tried segmentation as aux loss the masks generated by the bounding box seem to be not that accurate especially for the typical label and it has not worked for me so update if you somehow make it work" } ] }, { "id": 1350860, "postDate": "2021-06-15T21:20:04.390Z", "content": "Very nice work!! Do you think it would be helpful to remove / not calculate loss for the classification head of the EfficientDet since there is basically only one class in the detection part? And did you implement this model in your Google Drive?
", "rawMarkdown": "Very nice work!! Do you think it would be helpful to remove / not calculate loss for the classification head of the EfficientDet since there is basically only one class in the detection part? And did you implement this model in your Google Drive?" }, { "id": 1350206, "postDate": "2021-06-15T10:45:09.393Z", "content": "Hey thank you for your code, could you enlighten me about \"mask\" ? I've seen in your code that your network has a second head for it and that you calculate an auxiliary loss with it, but I don't know what mask is supposed to be? The boxes from the train datas ?
", "rawMarkdown": "Hey thank you for your code, could you enlighten me about \"mask\" ? I've seen in your code that your network has a second head for it and that you calculate an auxiliary loss with it, but I don't know what mask is supposed to be? The boxes from the train datas ?", "replies": [ { "id": 1350219, "postDate": "2021-06-15T11:01:46.540Z", "content": "they are rendered box annotation: 1 for pixel inside box and 0 for pixel outside
", "rawMarkdown": "they are rendered box annotation: 1 for pixel inside box and 0 for pixel outside", "votes": 3, "replies": [ { "id": 1350929, "postDate": "2021-06-15T23:44:19.817Z", "content": "@hengck23 Thanks for this starter!! I'm starting some experiments setup, just to make sure that everything is fine could you please guide us on how we can generate the masks ? eg see dataset.py
\nmask_file = data_dir + '/%s_mask_full_512/%s.png' % (d.set, d.image)
i provided some images at train_mask_full_512 at the google drive.
\nbasically, you can use e.g, opencv rect draw function to draw the bounding box and resize to 512x512
@hengck23 in your model why u name self.mask as one. I though u will output Bbox mask at output, but it give Single output only.
\nIs mask one prepared out of BBOxes ?
How are you generating the segmentation masks for the dataset ?
", "rawMarkdown": "How are you generating the segmentation masks for the dataset ?" }, { "id": 1347872, "postDate": "2021-06-13T14:35:12.640Z", "content": "Hello,how can I see the picture?
", "rawMarkdown": "Hello,how can I see the picture?" }, { "id": 1334609, "postDate": "2021-06-03T16:10:30.030Z", "content": "Cool. Expect👀
", "rawMarkdown": "Cool. Expect👀" }, { "id": 1329685, "postDate": "2021-05-31T09:52:57.157Z", "content": "SWALP sounds interesting but most materials are quite old(correct me if I was wrong), I can hardly find any codebase that uses this technique in object detection. For users with limited GPU resource people like me, this is the first thing comes out of my mind ;)
", "rawMarkdown": "[SWALP](https://pytorch.org/blog/stochastic-weight-averaging-in-pytorch/#low-precision-training) sounds interesting but most materials are quite old(correct me if I was wrong), I can hardly find any codebase that uses this technique in object detection. For users with limited GPU resource people like me, this is the first thing comes out of my mind ;) " }, { "id": 1320282, "postDate": "2021-05-24T01:09:51.920Z", "content": "🐸迟但到……..
", "rawMarkdown": "🐸迟但到........" }, { "id": 1319330, "postDate": "2021-05-23T05:49:00.847Z", "content": "Wow, Can't wait to learn more!
", "rawMarkdown": "Wow, Can't wait to learn more!" }, { "id": 1375776, "postDate": "2021-07-04T13:33:27.403Z", "rawMarkdown": "", "votes": -1, "isDeleted": true }, { "id": 1335208, "postDate": "2021-06-04T04:52:43.490Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1360192, "postDate": "2021-06-21T22:19:24.737Z", "content": "thanks for sharing
", "rawMarkdown": "thanks for sharing" }, { "id": 1359366, "postDate": "2021-06-21T08:41:59.303Z", "content": "so cool thank u
", "rawMarkdown": "so cool thank u" } ], "comments": [ { "id": 1941002, "author_name": "Nanduvardhanreddy", "author_url": "", "post_date": "2022-09-15T17:39:42.567000", "content": "I really appreciate your work and thanks for sharing information
", "votes": 5, "replies": [] }, { "id": 1363385, "author_name": "hengck23", "author_url": "", "post_date": "2021-06-24T06:11:58.033000", "content": "after numerous experiments:
\nconfirmed that aux loss using the mask can improve results significantly.
\nI can get local CV up to 0.39+ for single fold using different network (e.g. transformer), different pooling (attention-based, global average), different loss (lovasz, sigmoid, softamx, more weighing in aux loss, hard mining top +ve/-ve pixels etc)
freezing at fine tuning
increasing difficulty at training and fine-tuning
but it is sometimes unstable (i.e. different methods improve results for different folds/subsets. I can't find a method that improves for all cases)
this is because the mask is \"approximate\". we need a way to better refine the labeling of the pixel to mark them as indicating lung opacity/non-opacity or other.
you can think of pixels inside the box as \"likely to contain lung opacity\", but \"not all\" pixels are like that (e.g. box contains non-lung region). There is a similar argument for pixels that is outside the box. you need a pooling loss.
the number of train images is too small. Strong network overfits easily
I stacked top 50 object detections boxes with confidence to generate the \"finer\" mask. Regions with high-confidence opacity will be brighter whereas negative sample will be darker but not completely blank. Almost perfectly filtered distractive objects such as letter and metals. But got same result for appearance classification. Without external data, I doubt whether multitasking model will result in an improvement.
", "votes": 0, "replies": [] }, { "id": 1365025, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2021-06-25T11:55:16.643000", "content": "\n\nbut it is sometimes unstable (i.e. different methods improve results for different folds/subsets. I can't find a method that improves for all cases)
\n
Exactly! In my case too even average of folds is not the ideal improvement on leaderboard, I select my final best model as average of all my CV folds and LB. Even changing the seed changes the scores a lot (made it only for worse in my trails).
\n\n\nyou can think of pixels inside the box as \"likely to contain lung opacity\", but \"not all\" pixels are like that (e.g. box contains non-lung region). There is a similar argument for pixels that is outside the box. you need a pooling loss.
\n
I am using something like pooling loss working out great for me!
\n\n\nthe number of train images is too small. Strong network overfits easily
\n
I wonder if the extra data pretraining might solve that for us although that might mean even stronger augmentations or higher drop rates to make it not overfit.
", "votes": 0, "replies": [] }, { "id": 1365783, "author_name": "Anii", "author_url": "", "post_date": "2021-06-26T05:56:54.817000", "content": "I trained a MoCo model on CheXpert dataset(license issue, only for experiments), MoCo models was trained well with top 1 accuracy over 90%(k=65536 and 4096). But finetuned on this dataset doesn't improve the performance compare to imagenet pretrained weights. Naive knn on MoCo representations doesn't work as well, as expected.
", "votes": 0, "replies": [] } ] }, { "id": 1462936, "author_name": "hengck23", "author_url": "", "post_date": "2021-08-10T03:39:04.480000", "content": "to make this thread complete, i have updated:
\n[1] notebook to show interference code and results
\n(My submission couldn't complete in time and hence I missed the submission deadline)
\nhttps://www.kaggle.com/hengck23/final-v-00-private-public-score-0-615-0-628?scriptVersionId=70933214
this is bare minium model (only one k-fold model for study classification and one k-fold detection model)
\nwith decent results
Private Score 0.615
\nPublic Score 0.628
[2] code and trained model for effcienetDet (see folder \"2021-08-11-effdet\")
\nhttps://drive.google.com/drive/folders/16JzkuBpW5cwEexM5awq0wOK6t9gYLeKc
[3] more results on effcienetDet
\n
voc12 mAP computation, python code: https://github.com/Cartucho/mAP/blob/master/main.py
", "votes": 3, "replies": [ { "id": 1362347, "author_name": "hengck23", "author_url": "", "post_date": "2021-06-23T11:46:44.013000", "content": "
Constrained Optimization for Training Deep Neural Networks Under Class Imbalance
\nhttps://arxiv.org/pdf/2102.12894v1.pdf
we love neighbors:
\nhttps://www.kaggle.com/c/siim-covid19-detection/discussion/240250
\n\"FIGURE 7: Lung zones: proposed methodology for dividing a frontal CXR into 3 zones per lung (total of 6 zones). The upper zone extends from the apices to the superior portion of the hilum. The mid zone spans the space between the superior and inferior hilar margins. The lower zone extends from the inferior hilar margins to the costophrenic sulci.\"
\n
see also:
\n
Note: bbox aux loss help image classification.
\nconversely, 4-class image label aux loss should help opacity localization/detection.
\n(you need an aux-loss yolo or efficientDet)
modified meta pesudo label
\n
insprations from recent cvpr 2021
\n[CVPR 2021] Background Splitting: Finding Rare Categories in a Sea of Background
Thanks this is interesting, however in this case the background images are pseudo-labeled by an imagenet-pretrained model. Can this apply to our case.. ?
", "votes": 0, "replies": [] }, { "id": 1361158, "author_name": "human intelligence", "author_url": "", "post_date": "2021-06-22T16:00:35.597000", "content": "
\nget your best object detection model trained with default background box 0,0,0,0 -> label \"negative\" samples and make the predicted opacity box as background -> train again. From the image above (covid negative), you can see that vanilla model does makes some \"reasonable\" but wrong predictions on the negative image, but it also looks at the foreign objects such as the \"BIPE\" and spine. [0,0,0,0] fake background box tells the model \"your prediction is very wrong!\", but doesn't tell how & where to improve. Without a proper regularization, model will be easily overfitted to looking at irrelevant objects, like metal, heart, patient pose, etc. Pseudolabelling background with good boxes instead of 0,0,0,0 fake box (should) help model understand the corresponding region is not opacity and makes the loss smoother.
\nBuilding High Performance Chest X-Ray Classification Models | Talk 1 | CVPR 2021 | WIANLP Tutorial
Topic VI: Data Efficient Deep Learning for Medical Image Analysis | CVPR 2021 | VITA Workshop
\nhttps://www.youtube.com/watch?v=HnBvpwP0Wp4
success in experiments in transformer opens up a world of new possibilities …
\ncode coming up ….
\n
\n
\n
Have you share the code with transformer?
", "votes": 0, "replies": [] } ] }, { "id": 1349610, "author_name": "hengck23", "author_url": "", "post_date": "2021-06-15T00:50:44.503000", "content": "important conclusion from version.1 2020-14-jun:
\n(do not take any words for granted. i encourage you to verify the claims below and discuss here)
Confirmed. Different backbones of EffNet with aux loss consistently resulted in CV mAP 0.38x. Model get overfitted easily without aux loss.
", "votes": 0, "replies": [] }, { "id": 1350981, "author_name": "hengck23", "author_url": "", "post_date": "2021-06-16T01:50:15.457000", "content": "you may also want to check that at CV mAP 0.38x. , the validation and training loss gap is not large (much better than the case without aux loss). this indicates aux loss improve generalisation.
\nbut the loss curves are bumpy and fluctuating. there is still some room for improvement if you can smooth and stabilise the curve
", "votes": 1, "replies": [] }, { "id": 1351581, "author_name": "Anii", "author_url": "", "post_date": "2021-06-16T12:32:59.537000", "content": "I can confirm that augmentation helps a lot. But regard aux loss, my resnet101 model(224x224) started to overfit heavily when I add aux loss. The training accuracy is nearly 95% with aux loss and 70% without aux loss, validation accuracy is about 67% percent in both cases. I tried using all images and ignore all images(except negative class) without boxes, still no luck ;(
\nI'll try to use your input pipeline to investigate this further, maybe segmentation head only works better only if your augmentation is strong enough?
resnet perform worse than efficient net in my experiments
", "votes": 1, "replies": [] }, { "id": 1351728, "author_name": "human intelligence", "author_url": "", "post_date": "2021-06-16T14:51:38.603000", "content": "@artnotintelligence you probably need to specify high drop_rate and drop_path_rate in the effnet / resnet, especially when lr is low. In Heng's kit, they are both 0.2, but for big backbone I set drop_rate = 0.5 to combat overfitting, but model will still overfit when mAP is close to 0.385. This is the limit of simple aux loss strategy.
", "votes": 1, "replies": [] }, { "id": 1351808, "author_name": "Anii", "author_url": "", "post_date": "2021-06-16T16:07:19.397000", "content": "I add dropout rate up to 0.5 along with l2 regularization loss, which doesn't help avoid overfitting.
", "votes": 0, "replies": [] }, { "id": 1351850, "author_name": "hengck23", "author_url": "", "post_date": "2021-06-16T16:40:34.700000", "content": "\"This is the limit of simple aux loss strategy.\"
\nyou can predict the lung segmentation or pixel values of the related masked region.
\n
\n
input to the green \"predict the value of cross entropy loss\" model could be array of pixel differences of the predicted and truth of mask region.
\nwe bypass the step to obtain pseudo label … we just need the \"pseudo label loss\" for backprop
", "votes": 2, "replies": [] }, { "id": 1353291, "author_name": "Anii", "author_url": "", "post_date": "2021-06-17T03:15:37.277000", "content": "I don't quite understand the last part, are you saying that , for example, if softmax(logits) = [0.1, 0.1, 0.2, 0.6], then we use pseudo label = [0, 0, 0, 1]
\nas gt label to produce CE loss, then further regularize CE loss with predicted CE loss?
no. you don't predict the pseudo label.
\nyou can predict the CE loss value (of pseudo label) directly
@houndcl i think @phalanx Unet aux loss is better than mine. Maybe it can hit 0.395
", "votes": 1, "replies": [] } ] }, { "id": 1347120, "author_name": "hengck23", "author_url": "", "post_date": "2021-06-13T02:24:02.733000", "content": "archive for old post
\ni am preparing starter kit over the next few weeks.
\nIf you have recommended papers or methods, please let me know. I will include it in the starter kit.
What is desired:
\nold methods that have proved to work on xray images, etc
\nnew methods that may show interesting results (e.g. transformer, weak/semi-supervised learning)
\npaper that you want to understand learn, but there is no open source to verify if your understanding is correct
\ntraining/inference speedup
\ne.g. …
\nOnce i select the method, the starter kits will include:
reference code for training and inference
\nintermediate trained model
\nlogfile of training for reference and checking implementation, etc
\nexperiment results
\nsome training/hyperparameters tricks. etc
\nNote: i am currently busy with kaggle BMS molecule image-text translation. I will be more active here after BMS is over
the starter kit post will be something like this: https://www.kaggle.com/c/bms-molecular-translation/discussion/231190
", "votes": 4, "replies": [] }, { "id": 1367336, "author_name": "hengck23", "author_url": "", "post_date": "2021-06-27T15:49:09.230000", "content": "
\nThe real world confusion matrix from RICORD dataset. The \"typical appearance\" data seems cleaner in this competition. The accurate prediction of \"atypical\" and \"indeterminate\" appearances remain challenging. Atypical class may be slightly easier to improve, by integrating external dataset such as MIMIC and NIH. There are plenty of images with mass | nodule | effusion | pneumothorax | edema etc labels.
What is rescore?
", "votes": 1, "replies": [ { "id": 1380875, "author_name": "Simon", "author_url": "", "post_date": "2021-07-08T11:54:31.467000", "content": "At the beginning of the competition there was a bug in the leaderboard and once the bug was fixed all submissions were run again and \"rescored\" so they got a new score without the bug
", "votes": 0, "replies": [] } ] }, { "id": 1374603, "author_name": "hengck23", "author_url": "", "post_date": "2021-07-03T12:53:52.260000", "content": "i suddenly has an idea:
\nIn this research, we proposed an ACGAN based model
\ncalled CovidGAN that generates synthetic CXR images to
\nenlarge the dataset and to improve the performance of CNN
\nin COVID-19 detection. The research is implemented on a
\ndataset with 403 COVID-CXR images and 721 Normal-CXR
\nimages
https://arxiv.org/pdf/2103.05094.pdf
", "votes": 1, "replies": [] }, { "id": 1366297, "author_name": "hengck23", "author_url": "", "post_date": "2021-06-26T16:22:40.367000", "content": "
\nhttps://www.youtube.com/channel/UCOkkljs06NPPkjNysCdQV4w/videos
channel to replace patch as token in trasnformer
\nCross-Covariance Image Transformer (XCiT)
\nhttps://www.youtube.com/watch?v=g08NkNWmZTA
\nhttps://github.com/facebookresearch/xcit
https://arxiv.org/pdf/2106.13112.pdf
\nVolo+yolo?
\nhttps://arxiv.org/pdf/2104.10858.pdf
\nAll Tokens Matter: Token Labeling for Training Better Vision Transformers
\nhttps://www.youtube.com/watch?v=0TwZfRcqhbI&list=PL6liSIqFR4BUr7F4pL6aoucjkV3HYZ48J
\nhttps://github.com/facebookresearch/unbiased-teacher
I tried similiar setup but my mAP at study level was lower with segmentation head. What's your score without segmentation loss?
", "votes": 1, "replies": [] }, { "id": 1347883, "author_name": "Awsaf", "author_url": "", "post_date": "2021-06-13T14:45:53.973000", "content": "Hi @hengck23, what was your score without the aux loss ?
", "votes": 1, "replies": [ { "id": 1350072, "author_name": "Joseph AMIGO", "author_url": "", "post_date": "2021-06-15T08:45:30.530000", "content": "What is aux loss ? I'm googling it and not finding anything!
", "votes": 0, "replies": [] }, { "id": 1350176, "author_name": "Ashutosh Behera", "author_url": "", "post_date": "2021-06-15T10:09:45.767000", "content": "@josephamigo I believe aux loss is auxiliary loss. It helps to reduce the vanishing gradient problem for earlier layers, stabilizes the training and is used as regularization. It's only used for training and not for inference.
\nReference: https://stats.stackexchange.com/questions/304699/what-is-auxiliary-loss-as-mentioned-in-pspnet-paper
", "votes": 5, "replies": [] }, { "id": 1350199, "author_name": "Joseph AMIGO", "author_url": "", "post_date": "2021-06-15T10:32:51.100000", "content": "Thank you very much!!
", "votes": 1, "replies": [] } ] }, { "id": 1334679, "author_name": "Aman Deep Gupta", "author_url": "", "post_date": "2021-06-03T17:27:06.223000", "content": "cool just waiting for your discussion and ideas @hengck23 and starting baseline of @yasufuminakama and saturday.
", "votes": 1, "replies": [] }, { "id": 1319215, "author_name": "Cyr1ll", "author_url": "", "post_date": "2021-05-23T03:11:15.323000", "content": "Not really a method, but a good summary. Also, it mentions a way to speed up a training procedure.
\nCheXtransfer: Performance and Parameter Efficiency of ImageNet Models for Chest X-Ray Interpretation
\nhttps://arxiv.org/abs/2101.06871
Welcome! I will learn from you again.
", "votes": 1, "replies": [] }, { "id": 1359001, "author_name": "hengck23", "author_url": "", "post_date": "2021-06-21T02:02:50.340000", "content": "lung mask:
\nhttps://github.com/v7labs/covid-19-xray-dataset
\nhttps://github.com/ieee8023/covid-chestxray-dataset
https://github.com/raghavian/lungVAE
\nhttps://arxiv.org/pdf/2104.05892.pdf
lung opacity:
\nhttps://arxiv.org/pdf/2002.02497.pdf
\nhttps://github.com/mlmed/torchxrayvision
\nhttps://openaccess.thecvf.com/content_CVPRW_2020/papers/w22/Gabruseva_Deep_Learning_for_Automatic_Pneumonia_Detection_CVPRW_2020_paper.pdf
Hello,How can I get this picture?
", "votes": 0, "replies": [] }, { "id": 1358760, "author_name": "hengck23", "author_url": "", "post_date": "2021-06-20T17:52:43.727000", "content": "Vision Transformer using Low-level Chest X-ray Feature Corpus for COVID-19 Diagnosis and Severity Quantification
", "votes": 2, "replies": [] }, { "id": 1358997, "author_name": "Zekun", "author_url": "", "post_date": "2021-06-21T01:46:33.963000", "content": "Thanks for your ideas!
", "votes": 0, "replies": [] }, { "id": 1359004, "author_name": "hengck23", "author_url": "", "post_date": "2021-06-21T02:09:05.740000", "content": "PCAM pooling
\n
\nhttps://github.com/jfhealthcare/Chexpert
In run_train.py(also in run_submit.py), there seems a bug in calculating the topk accuracy,
\npredict = probability.argsort(-1)[::-1] should be predict = probability.argsort(-1)[:, ::-1]?
good catch.
\nthe same occurs in do_valid() function in run_train.py.
Thanks
", "votes": 1, "replies": [] }, { "id": 1357541, "author_name": "Ioannis M", "author_url": "", "post_date": "2021-06-19T18:44:03.557000", "content": "nice catch, Anii thanks!!
\n1) BTW do you also get NaN in map score ?
2) also a minor point in run_train.py (affects only the logging, code runs fine)
batch_loss = np.array([loss0.item(), loss1.item(), loss2.item()])
loss2 is always 0.0 from initialisation. I guess if you want to correct it we need to add before:
\nloss2 = loss0+loss1
after the topk bug is fixed, you will get something like:
\n experiment = ['effb3-full-512-mask-v8', 'run_train_2.py']\n\n |---------------- VALID -------------|---- TRAIN/BATCH -------\n epoch | loss map (map*0.6) top1,2 | loss0 loss1 \n----------------------------------------------------------------------\n 0.00 | 1.412 0.257 (0.171) 0.281 0.425 | 0.000 0.000 0.000 | \n 0.16 | 1.103 0.419 (0.279) 0.597 0.755 | 1.276 0.561 0.000 | \n 0.32 | 0.994 0.464 (0.309) 0.612 0.808 | 1.096 0.396 0.000 | \n 0.48 | 1.073 0.453 (0.302) 0.585 0.779 | 1.044 0.302 0.000 | \n 0.64 | 0.925 0.505 (0.337) 0.638 0.812 | 1.045 0.258 0.000 | \n 0.80 | 0.971 0.475 (0.317) 0.634 0.794 | 1.013 0.246 0.000 | \n 0.96 | 0.940 0.488 (0.326) 0.639 0.812 | 1.055 0.232 0.000 | \n 1.13 | 0.949 0.507 (0.338) 0.639 0.815 | 0.990 0.221 0.000 | \n 1.29 | 1.013 0.502 (0.334) 0.585 0.764 | 1.017 0.224 0.000 | \n 1.45 | 0.970 0.489 (0.326) 0.631 0.813 | 0.991 0.211 0.000 | \nthis reveals a few tricks:
\nmap is worse than top1. this is the effect of imbalance class. map is averaged across all classes, i did not apply that for top1
the probing discussion is in https://www.kaggle.com/c/siim-covid19-detection/discussion/244066. You can obtain the prior class probability
if you know how many test samples there are, you can \"de-rank\" the probabiliy. e.g https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/73738
the cost of mistakes is not the same for each sample. It is more costly to make mistakes for rare classes. some kind of cost-aware loss function or post-processing may be able to improve results?
\ntop2 results are pretty good. if you want to post-process results, you may consider \"moving prediction from top2 to top1\" if the \"cost\" of this action justifies itself.
one can think of more ideas if one plots some graphs to analyze the distribution of relative/absolute score of different class
There is a discussion to replace Attention with MLP. I don't know if it can be a good application here.
", "votes": 2, "replies": [] }, { "id": 1314181, "author_name": "", "author_url": "", "post_date": "2021-05-19T02:36:30.963000", "content": "", "votes": 0, "replies": [] }, { "id": 1462191, "author_name": "", "author_url": "", "post_date": "2021-08-09T18:11:05.583000", "content": "", "votes": -1, "replies": [] }, { "id": 1464120, "author_name": "", "author_url": "", "post_date": "2021-08-10T13:04:56.283000", "content": "", "votes": 0, "replies": [] }, { "id": 1393690, "author_name": "", "author_url": "", "post_date": "2021-07-19T19:54:31.270000", "content": "", "votes": 0, "replies": [ { "id": 1393708, "author_name": "", "author_url": "", "post_date": "2021-07-19T20:12:59.367000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 1379390, "author_name": "", "author_url": "", "post_date": "2021-07-07T10:26:39.310000", "content": "", "votes": 0, "replies": [] }, { "id": 1375162, "author_name": "", "author_url": "", "post_date": "2021-07-04T00:00:17.667000", "content": "", "votes": 0, "replies": [] }, { "id": 1359754, "author_name": "", "author_url": "", "post_date": "2021-06-21T13:53:56.010000", "content": "", "votes": 0, "replies": [] }, { "id": 1354831, "author_name": "", "author_url": "", "post_date": "2021-06-17T23:28:42.667000", "content": "", "votes": 0, "replies": [ { "id": 1356696, "author_name": "", "author_url": "", "post_date": "2021-06-19T07:13:16.947000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 1350860, "author_name": "", "author_url": "", "post_date": "2021-06-15T21:20:04.390000", "content": "", "votes": 0, "replies": [] }, { "id": 1350206, "author_name": "", "author_url": "", "post_date": "2021-06-15T10:45:09.393000", "content": "", "votes": 0, "replies": [ { "id": 1350219, "author_name": "", "author_url": "", "post_date": "2021-06-15T11:01:46.540000", "content": "", "votes": 3, "replies": [ { "id": 1350929, "author_name": "", "author_url": "", "post_date": "2021-06-15T23:44:19.817000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 1350931, "author_name": "", "author_url": "", "post_date": "2021-06-15T23:56:14.127000", "content": "", "votes": 3, "replies": [] }, { "id": 1354133, "author_name": "", "author_url": "", "post_date": "2021-06-17T11:57:39.610000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 1349430, "author_name": "", "author_url": "", "post_date": "2021-06-14T18:52:42.410000", "content": "", "votes": 0, "replies": [] }, { "id": 1347872, "author_name": "", "author_url": "", "post_date": "2021-06-13T14:35:12.640000", "content": "", "votes": 0, "replies": [] }, { "id": 1334609, "author_name": "", "author_url": "", "post_date": "2021-06-03T16:10:30.030000", "content": "", "votes": 0, "replies": [] }, { "id": 1329685, "author_name": "", "author_url": "", "post_date": "2021-05-31T09:52:57.157000", "content": "", "votes": 0, "replies": [] }, { "id": 1320282, "author_name": "", "author_url": "", "post_date": "2021-05-24T01:09:51.920000", "content": "", "votes": 0, "replies": [] }, { "id": 1319330, "author_name": "", "author_url": "", "post_date": "2021-05-23T05:49:00.847000", "content": "", "votes": 0, "replies": [] }, { "id": 1375776, "author_name": "", "author_url": "", "post_date": "2021-07-04T13:33:27.403000", "content": "", "votes": -1, "replies": [] }, { "id": 1335208, "author_name": "", "author_url": "", "post_date": "2021-06-04T04:52:43.490000", "content": "", "votes": 0, "replies": [] }, { "id": 1360192, "author_name": "", "author_url": "", "post_date": "2021-06-21T22:19:24.737000", "content": "", "votes": 0, "replies": [] }, { "id": 1359366, "author_name": "", "author_url": "", "post_date": "2021-06-21T08:41:59.303000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1314110": "Let's make this a gold starter kit!\n---\n\ni am preparing starter kit over the next few weeks.\nIf you have recommended papers or methods, please let me know. I will include it in the starter kit.\n \ngoogle drive link for code, model, train, log, etc:\nhttps://drive.google.com/drive/folders/1oXfTayH6eRG36DAKmvwVyFE5cmPR6K6W?usp=sharing\n\n---\n\nversion.1 2020-14-jun: completed and ready for download!\n- image classification only\n- compare effects of \"segmentation as aux loss\"\n- ~~ lb-0.389 for efficienet-b3 at image size=512 (5-fold ensmeble, original+flipTTA)~~\n- ~~lb-0.396 for efficienet-b5 at image size=640 (5-fold ensmeble, original+flipTTA)~~\n\nafter rescore\n- lb-0.444 for efficienet-b3 at image size=512 (5-fold ensmeble, original+flipTTA)\n- lb-0.450 for efficienet-b5 at image size=640 (5-fold ensmeble, original+flipTTA)\n \n\n\n---\n\nversion.2: ... in preparation ...\n- due date: 21-jun\n- ~~\"Be Your Own Teacher: Improve the Performance of Convolutional Neural Networks via Self Distillation\" - iccv 2019~~ (let me try object detection first and decide if i should focus more on image classification or object detection)\n\n\n- customize bifpn from https://github.com/rwightman/efficientdet-pytorch\n- use efficientNet v2 as backbone: https://arxiv.org/pdf/2104.00298.pdf\n- use adversial-prop-det for training: https://arxiv.org/pdf/2103.13886.pdf\n\nreference: https://amaarora.github.io/2021/01/13/efficientdet-pytorch.html\n\n\nversion.1 shows that pure classification does not work better than one with joint aux loss.\nthe reason is overfitting\n\nhence we need to implement something like this .... \n\n\n---\n\nlater version\nyolo-2021 : https://twitter.com/alexeyab84/status/1398443022619189248?s=20\n\nreference: https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229615\n\nit would be interesting to use multi-task yolo by extracting extra labels from radiologist report in external data ...\n\n\n\n---\n\netc (random notes for myself):\n\nFB MoCo self-supervised contrastive learning?\n\ntransformers ????\n\nmeta-pseudo label adapted for detection, transMIL (transformer MIL, ranking as a \"sorted\" sequence prediction)\n\n- mAP for imbalanced class? \nRobust Deep AUC Maximization: A New Surrogate Loss and Empirical Studies on Medical Image Classification\nhttps://libauc.org/\n\nSemi-Supervised AUC Optimization without Guessing Labels of Unlabeled Data\n\n- change yolo objectiveness score to iou score\nsee https://www.kaggle.com/c/global-wheat-detection/discussion/172436\n\ntrasnformer yolo https://www.gitmemory.com/issue/ultralytics/yolov5/2329/808852867\n\nre-ranking trasnformer for ensemble\n\nresize artificats, adversial noise \n\nConstrained Optimization to Train Neural Networks\non Critical and Under-Represented Classes\n\ntransformer hybrid: \nhttps://www.kaggle.com/bamps53/lb0-87-part-of-3rd-place-solution?scriptVersionId=65265405\nhttps://www.kaggle.com/c/bms-molecular-translation/discussion/243932\n\nhttps://brixia.github.io/\n\nwhy to convert weak segmentation ground truth (bbox) into a stronger one? ... pool most probable pixels? lung prior? read papers on the weak labels for segmentation.\n\n\n---\n\nkeys to winning\n\n- a better loss function for mAP image classification and localization (same or different one?) ...\n- someone can recommend a ranking-based loss or large margin function for mAP?\n\n- use of external data (especially MIDRC-RICORD Data, Kaggle previous data etc)\n- small data size is one of the main problems in this challenge. Your results may be unstable (overfitting, etc) ... you need to think of a way to solve this, e.g. deep/extra supervision, augmentation \n", "1941002": "I really appreciate your work and thanks for sharing information", "1363385": "after numerous experiments:\n\n- confirmed that aux loss using the mask can improve results significantly. \nI can get local CV up to 0.39+ for single fold using different network (e.g. transformer), different pooling (attention-based, global average), different loss (lovasz, sigmoid, softamx, more weighing in aux loss, hard mining top +ve/-ve pixels etc)\n\n- freezing at fine tuning\n\n- increasing difficulty at training and fine-tuning\n\n- but it is sometimes unstable (i.e. different methods improve results for different folds/subsets. I can't find a method that improves for all cases)\n\n- this is because the mask is \"approximate\". we need a way to better refine the labeling of the pixel to mark them as indicating lung opacity/non-opacity or other.\n\n- you can think of pixels inside the box as \"likely to contain lung opacity\", but \"not all\" pixels are like that (e.g. box contains non-lung region). There is a similar argument for pixels that is outside the box. you need a pooling loss.\n\n- the number of train images is too small. Strong network overfits easily", "1462936": "to make this thread complete, i have updated:\n\n[1] notebook to show interference code and results\n(My submission couldn't complete in time and hence I missed the submission deadline)\nhttps://www.kaggle.com/hengck23/final-v-00-private-public-score-0-615-0-628?scriptVersionId=70933214\n\nthis is bare minium model (only one k-fold model for study classification and one k-fold detection model)\nwith decent results\n\nPrivate Score 0.615\nPublic Score 0.628\n\n\n\n---\n\n[2] code and trained model for effcienetDet (see folder \"2021-08-11-effdet\")\nhttps://drive.google.com/drive/folders/16JzkuBpW5cwEexM5awq0wOK6t9gYLeKc\n\n\n---\n\n\n[3] more results on effcienetDet\n\n\n\n", "1361974": "voc12 mAP computation, python code: https://github.com/Cartucho/mAP/blob/master/main.py", "1359014": "we love neighbors:\n- Semi-Supervised Learning of Visual Features by Non-Parametrically\nPredicting View Assignments with Support Samples\n- Many-to-One Distribution Learning and K-Nearest Neighbor Smoothing for\nThoracic Disease Identification", "1358761": "https://www.kaggle.com/c/siim-covid19-detection/discussion/240250\n\n\"FIGURE 7: Lung zones: proposed methodology for dividing a frontal CXR into 3 zones per lung (total of 6 zones). The upper zone extends from the apices to the superior portion of the hilum. The mid zone spans the space between the superior and inferior hilar margins. The lower zone extends from the inferior hilar margins to the costophrenic sulci.\"\n\n\n\n\nsee also:\n- Vision Transformer using Low-level Chest X-ray Feature Corpus for COVID-19 Diagnosis and Severity Quantification\n- Vision Transformer for COVID-19 CXR Diagnosis using Chest X-ray Feature Corpus\n\n\n\n\n- Radiologic diagnosis of patients with COVID-19\n", "1356561": "Note: bbox aux loss help image classification.\n\n**conversely,** 4-class image label aux loss should help opacity localization/detection.\n(you need an aux-loss yolo or efficientDet)", "1360751": "insprations from recent cvpr 2021\n[CVPR 2021] Background Splitting: Finding Rare Categories in a Sea of Background\n\n", "1360311": "success in experiments in transformer opens up a world of new possibilities ...\ncode coming up ....", "1349610": "important conclusion from version.1 2020-14-jun: \n(do not take any words for granted. i encourage you to verify the claims below and discuss here)\n- baseline performance: no augmentation: CV/LB 0.34/?, add augmentation 0.36/0.38, add aux loss 0.38/0.395\n(this is based on old scoring system, where sample submission scored 0.050)\n", "1347120": "archive for old post \n\n----\ni am preparing starter kit over the next few weeks.\nIf you have recommended papers or methods, please let me know. I will include it in the starter kit.\n\nWhat is desired:\n\nold methods that have proved to work on xray images, etc\nnew methods that may show interesting results (e.g. transformer, weak/semi-supervised learning)\npaper that you want to understand learn, but there is no open source to verify if your understanding is correct\ntraining/inference speedup\ne.g. …\nOnce i select the method, the starter kits will include:\n\nreference code for training and inference\nintermediate trained model\nlogfile of training for reference and checking implementation, etc\nexperiment results\nsome training/hyperparameters tricks. etc\nNote: i am currently busy with kaggle BMS molecule image-text translation. I will be more active here after BMS is over\n\nthe starter kit post will be something like this: https://www.kaggle.com/c/bms-molecular-translation/discussion/231190", "1367336": "", "1380600": "What is rescore?", "1374603": "i suddenly has an idea:\n- super-resolution as self supervsied loss for unlabelled data", "1366470": "In this research, we proposed an ACGAN based model\ncalled CovidGAN that generates synthetic CXR images to\nenlarge the dataset and to improve the performance of CNN\nin COVID-19 detection. The research is implemented on a\ndataset with 403 COVID-CXR images and 721 Normal-CXR\nimages\n\nhttps://arxiv.org/pdf/2103.05094.pdf", "1366297": "\nhttps://www.youtube.com/channel/UCOkkljs06NPPkjNysCdQV4w/videos", "1366209": "channel to replace patch as token in trasnformer\nCross-Covariance Image Transformer (XCiT)\nhttps://www.youtube.com/watch?v=g08NkNWmZTA\nhttps://github.com/facebookresearch/xcit", "1364534": "https://arxiv.org/pdf/2106.13112.pdf\nVolo+yolo?", "1363356": "\nhttps://arxiv.org/pdf/2104.10858.pdf\nAll Tokens Matter: Token Labeling for Training Better Vision Transformers", "1356552": "\nhttps://www.youtube.com/watch?v=0TwZfRcqhbI&list=PL6liSIqFR4BUr7F4pL6aoucjkV3HYZ48J\nhttps://github.com/facebookresearch/unbiased-teacher\n", "1348118": "I tried similiar setup but my mAP at study level was lower with segmentation head. What's your score without segmentation loss?", "1347883": "Hi @hengck23, what was your score without the **aux loss** ?", "1334679": "cool just waiting for your discussion and ideas @hengck23 and starting baseline of @yasufuminakama and saturday.", "1319215": "Not really a method, but a good summary. Also, it mentions a way to speed up a training procedure.\n\n**CheXtransfer: Performance and Parameter Efficiency of ImageNet Models for Chest X-Ray Interpretation**\nhttps://arxiv.org/abs/2101.06871", "1314643": "Welcome! I will learn from you again.", "1359001": "lung mask:\nhttps://github.com/v7labs/covid-19-xray-dataset\nhttps://github.com/ieee8023/covid-chestxray-dataset\n\nhttps://github.com/raghavian/lungVAE\nhttps://arxiv.org/pdf/2104.05892.pdf\n\nlung opacity:\nhttps://arxiv.org/pdf/2002.02497.pdf\nhttps://github.com/mlmed/torchxrayvision\nhttps://openaccess.thecvf.com/content_CVPRW_2020/papers/w22/Gabruseva_Deep_Learning_for_Automatic_Pneumonia_Detection_CVPRW_2020_paper.pdf", "1358302": "", "1356643": "In run_train.py(also in run_submit.py), there seems a bug in calculating the topk accuracy,\n`predict = probability.argsort(-1)[::-1]` should be `predict = probability.argsort(-1)[:, ::-1]`? ", "1336213": "There is a discussion to replace Attention with MLP. I don't know if it can be a good application here.", "1314181": "It’s an honor to explore the philosophy of artificial intelligence with you on another battlefield", "1462191": "Do you think inputting the image with only 1 channel as opposed to 3 could increase the score? maybe decrease it? idk. personally, when i used one channel the accuracy reduced", "1464120": "@hengck23 Thank you for all information you shared for this competition. As a team we used your code, ideas and implemented them. Looking forward for next \"starter kit\" discussion 👀", "1393690": "Nice work! I'd like to ask what is the \"00007400_model.path\" set as the initial checkpoint. As I tried the training with your original setting (efficient net_b3a) with none initial checkpoint, it reached the best point at 3200 iterations, and if I try with effnetv2, it converges faster, around 600 iterations. ", "1379390": "hi \nwhy use efficientnet first 6 layers out put as mask to calculate aux loss?\nHow do you know use 6 layer output rather than 7 layer output?\n", "1375162": "hiii i tried your started kit its so good i was able to train but having problems while submission i mean how to get in on kaggle i trained on my system i need to submit in kaggle can u help please", "1359754": "Do you expect the multi-task model to work better than the models that solve each task (OD, clf)?", "1354831": "Great work. I'll try the aux loss.", "1350860": "Very nice work!! Do you think it would be helpful to remove / not calculate loss for the classification head of the EfficientDet since there is basically only one class in the detection part? And did you implement this model in your Google Drive?", "1350206": "Hey thank you for your code, could you enlighten me about \"mask\" ? I've seen in your code that your network has a second head for it and that you calculate an auxiliary loss with it, but I don't know what mask is supposed to be? The boxes from the train datas ?", "1349430": "How are you generating the segmentation masks for the dataset ?", "1347872": "Hello,how can I see the picture?", "1334609": "Cool. Expect👀", "1329685": "[SWALP](https://pytorch.org/blog/stochastic-weight-averaging-in-pytorch/#low-precision-training) sounds interesting but most materials are quite old(correct me if I was wrong), I can hardly find any codebase that uses this technique in object detection. For users with limited GPU resource people like me, this is the first thing comes out of my mind ;) ", "1320282": "🐸迟但到........", "1319330": "Wow, Can't wait to learn more!", "1375776": "", "1335208": "", "1360192": "thanks for sharing", "1359366": "so cool thank u" } }