Hello everyone.
\n\nI have created a dataset which (partially) solves a class imbalance of the provided data.\nAs you know the original competition data has ~98% of benign cases and only ~2% of malignant.
\n\nSo I have scrapped only malignant melanoma images from the 2019 ISIC competition and added them to original dataset, which leaves us with ~86% benign and ~14% malignant cases.
\n\nThen all images were center cropped and resized to 256x256.\nI also have a kernel which uses this dataset and score 0.925 on the LB.
", "messages": [ { "id": "888591", "postDate": "06/16/2020 12:58:05", "content": "Hello everyone.
\n\nI have created a dataset which (partially) solves a class imbalance of the provided data.\nAs you know the original competition data has ~98% of benign cases and only ~2% of malignant.
\n\nSo I have scrapped only malignant melanoma images from the 2019 ISIC competition and added them to original dataset, which leaves us with ~86% benign and ~14% malignant cases.
\n\nThen all images were center cropped and resized to 256x256.\nI also have a kernel which uses this dataset and score 0.925 on the LB.
", "rawMarkdown": "Hello everyone.\n\nI have [created a dataset](https://www.kaggle.com/nroman/melanoma-external-malignant-256) which (partially) solves a class imbalance of the provided data.\nAs you know the original competition data has ~98% of benign cases and only ~2% of malignant.\n\nSo I have scrapped only malignant melanoma images from the [2019 ISIC competition](https://challenge2019.isic-archive.com/data.html) and added them to original dataset, which leaves us with ~86% benign and ~14% malignant cases.\n\nThen all images were center cropped and resized to 256x256.\nI also have a [kernel](https://www.kaggle.com/nroman/melanoma-pytorch-starter-efficientnet) which uses this dataset and score 0.925 on the LB.", "votes": null }, { "id": "889334", "postDate": "06/16/2020 23:13:35", "content": "I downloaded the data and used it in colab but many images are missing from the folder but are there in the csv
", "rawMarkdown": "I downloaded the data and used it in colab but many images are missing from the folder but are there in the csv", "votes": null }, { "id": "891702", "postDate": "06/18/2020 11:54:24", "content": "@nroman i have used the above dataset that you mentioned without TTA. I have trained a efficientnet-b1. The validation score is around is quite good around 0.95 on 5 fold. But the LB score is not at all good.
", "rawMarkdown": "nroman i have used the above dataset that you mentioned without TTA. I have trained a efficientnet-b1. The validation score is around is quite good around 0.95 on 5 fold. But the LB score is not at all good.", "votes": null }, { "id": "891850", "postDate": "06/18/2020 13:48:16", "content": "Same experience, using EfficientNet B6 with Metadata. Validation 0.9+ but LB 0.85-.87 range.
", "rawMarkdown": "Same experience, using EfficientNet B6 with Metadata. Validation 0.9+ but LB 0.85-.87 range.", "votes": null } ], "comments": [ { "id": 889334, "author_name": "rajprakrit", "author_url": "", "post_date": "06/16/2020 23:13:35", "content": "I downloaded the data and used it in colab but many images are missing from the folder but are there in the csv
", "votes": null, "replies": [] }, { "id": 891702, "author_name": "ayushpatidar", "author_url": "", "post_date": "06/18/2020 11:54:24", "content": "@nroman i have used the above dataset that you mentioned without TTA. I have trained a efficientnet-b1. The validation score is around is quite good around 0.95 on 5 fold. But the LB score is not at all good.
", "votes": null, "replies": [ { "id": 891850, "author_name": "richardepstein", "author_url": "", "post_date": "06/18/2020 13:48:16", "content": "Same experience, using EfficientNet B6 with Metadata. Validation 0.9+ but LB 0.85-.87 range.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "888591": "Hello everyone.\n\nI have [created a dataset](https://www.kaggle.com/nroman/melanoma-external-malignant-256) which (partially) solves a class imbalance of the provided data.\nAs you know the original competition data has ~98% of benign cases and only ~2% of malignant.\n\nSo I have scrapped only malignant melanoma images from the [2019 ISIC competition](https://challenge2019.isic-archive.com/data.html) and added them to original dataset, which leaves us with ~86% benign and ~14% malignant cases.\n\nThen all images were center cropped and resized to 256x256.\nI also have a [kernel](https://www.kaggle.com/nroman/melanoma-pytorch-starter-efficientnet) which uses this dataset and score 0.925 on the LB.", "889334": "I downloaded the data and used it in colab but many images are missing from the folder but are there in the csv", "891702": "nroman i have used the above dataset that you mentioned without TTA. I have trained a efficientnet-b1. The validation score is around is quite good around 0.95 on 5 fold. But the LB score is not at all good.", "891850": "Same experience, using EfficientNet B6 with Metadata. Validation 0.9+ but LB 0.85-.87 range." }, "source": "meta" }