Sorry, I took a while to write this. I was struggling with my course exam 👀
\n\nCongratulations to all the participants and winners! This competition was a little bit tricky and different, especially due to the very noisy label in the dataset. I am going to share my solution which outperformed all of my other networks and ensembles.
\n\nPrivate score: 0.66758 \nPublic score: 0.66654 (Yes! The network gave higher score in private!)\nNetwork backbone: InceptionResNetV2\nSegmentation Model: FPN
\n\nTrain height: 384\nTrain weight: 512\nLoss: Lovasz\nOptimizer: Radam\nLearning rate: 1e-4\nPretrained weight: Imagenet
\n\nI used albumentations.
\n\nTTA:\n1. HorizontalFlip, VerticalFlip\n2. ShiftScaleRotate \n3. Blur\n4. ToGray\n4. RandomBrightnessContrast, RandomGamma, RGBShift, ElasticTransform (One of them)
\n\nPost-processing (For Prediction) Augmentation: \n- HorizontalFlip, VerticalFlip, ToGray, RandomGamma, Normalize.
\n\nFor all classes:\nUpper: 0.6\nLower: 0.4\nArea: 100\nMin area: 200
\n\nI have used a total of 4 folds and 4 weights from each fold. So there was a total of 16 weights from 4 folds and the prediction was the simple average of 16 folds.
\n\nThere was no other ensemble or use of any extra classifier.
\n\nI didn't have any good GPU locally. So I used the Google cloud in the last few days. Google gave $300 dollar free trial credits which were good for around 120 hours VM instance.
\n\nI build a VM there with V100 GPU.
\n\nSorry, I took a while to write this. I was struggling with my course exam 👀
\n\nCongratulations to all the participants and winners! This competition was a little bit tricky and different, especially due to the very noisy label in the dataset. I am going to share my solution which outperformed all of my other networks and ensembles.
\n\nPrivate score: 0.66758 \nPublic score: 0.66654 (Yes! The network gave higher score in private!)\nNetwork backbone: InceptionResNetV2\nSegmentation Model: FPN
\n\nTrain height: 384\nTrain weight: 512\nLoss: Lovasz\nOptimizer: Radam\nLearning rate: 1e-4\nPretrained weight: Imagenet
\n\nI used albumentations.
\n\nTTA:\n1. HorizontalFlip, VerticalFlip\n2. ShiftScaleRotate \n3. Blur\n4. ToGray\n4. RandomBrightnessContrast, RandomGamma, RGBShift, ElasticTransform (One of them)
\n\nPost-processing (For Prediction) Augmentation: \n- HorizontalFlip, VerticalFlip, ToGray, RandomGamma, Normalize.
\n\nFor all classes:\nUpper: 0.6\nLower: 0.4\nArea: 100\nMin area: 200
\n\nI have used a total of 4 folds and 4 weights from each fold. So there was a total of 16 weights from 4 folds and the prediction was the simple average of 16 folds.
\n\nThere was no other ensemble or use of any extra classifier.
\n\nI didn't have any good GPU locally. So I used the Google cloud in the last few days. Google gave $300 dollar free trial credits which were good for around 120 hours VM instance.
\n\nI build a VM there with V100 GPU.
\n\nYour minimum area is quite small. With single model and small min area, getting rid of false positives is very impressive!!!
", "rawMarkdown": "Your minimum area is quite small. With single model and small min area, getting rid of false positives is very impressive!!!", "votes": null }, { "id": "680616", "postDate": "11/25/2019 01:09:19", "content": "What is Area=100, less then min area?
", "rawMarkdown": "What is Area=100, less then min area?", "votes": null } ], "comments": [ { "id": 680614, "author_name": "raghaw", "author_url": "", "post_date": "11/25/2019 01:08:07", "content": "Your minimum area is quite small. With single model and small min area, getting rid of false positives is very impressive!!!
", "votes": null, "replies": [] }, { "id": 680616, "author_name": "raghaw", "author_url": "", "post_date": "11/25/2019 01:09:19", "content": "What is Area=100, less then min area?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "680489": "Sorry, I took a while to write this. I was struggling with my course exam 👀 \n\nCongratulations to all the participants and winners! This competition was a little bit tricky and different, especially due to the very noisy label in the dataset. I am going to share my solution which outperformed all of my other networks and ensembles. \n\n## Score and Network Summary\n**Private score:** 0.66758 \n**Public score:** 0.66654 (*Yes! The network gave higher score in private!*)\n**Network backbone:** InceptionResNetV2\n**Segmentation Model:** FPN\n\n**Train height:** 384\n**Train weight:** 512\n**Loss:** Lovasz\n**Optimizer:** Radam\n**Learning rate:** 1e-4\n**Pretrained weight:** Imagenet\n\n### Augmentations:\n I used albumentations. \n\n**TTA:**\n1. HorizontalFlip, VerticalFlip\n2. ShiftScaleRotate \n3. Blur\n4. ToGray\n4. RandomBrightnessContrast, RandomGamma, RGBShift, ElasticTransform (One of them)\n\n**Post-processing (For Prediction) Augmentation:** \n- HorizontalFlip, VerticalFlip, ToGray, RandomGamma, Normalize. \n\n### Post-Processing confidence density:\nFor all classes:\n**Upper**: 0.6\n**Lower**: 0.4\n**Area**: 100\n**Min area**: 200\n\n### Folds and weights\nI have used a total of 4 folds and 4 weights from each fold. So there was a total of 16 weights from 4 folds and the prediction was the simple average of 16 folds. \n\n**There was no other ensemble or use of any extra classifier.**\n\n### GPU\nI didn't have any good GPU locally. So I used the Google cloud in the last few days. Google gave $300 dollar free trial credits which were good for around 120 hours VM instance.\n\nI build a VM there with V100 GPU. \n\n### Some thoughts\n1. I believe this network performs well due to the InceptionResNetV2 only. Because I have output from a similar setup with other backbones which wasn't anywhere near to this score. \n2. I should have trusted the validation score. This model produced the highest Kaggle dice score at validation compared to other models.", "680614": "Your minimum area is quite small. With single model and small min area, getting rid of false positives is very impressive!!!", "680616": "What is Area=100, less then min area?" }, "source": "meta" }