Has anyone tried SMOTE on the data , since using class_weights and kfolds is not helping much.
", "messages": [ { "id": "1099156", "postDate": "12/02/2020 06:01:23", "content": "Has anyone tried SMOTE on the data , since using class_weights and kfolds is not helping much.
", "rawMarkdown": "Has anyone tried SMOTE on the data , since using class_weights and kfolds is not helping much.", "votes": null }, { "id": "1101431", "postDate": "12/03/2020 22:28:19", "content": "Paper if anyone is interested: https://arxiv.org/pdf/1106.1813.pdf
", "rawMarkdown": "Paper if anyone is interested: https://arxiv.org/pdf/1106.1813.pdf", "votes": null }, { "id": "1101786", "postDate": "12/04/2020 09:01:19", "content": "I only know some structured data could use SMOTE, like in some data mining competition.
\nIn CV field, how to use SMOTE is a question, maybe by some data augmentation?
\nHope someone else know that.
Since the distribution of the minority categories is similar, I plan to enhance the minority categories uniformly and ignore the majority categories.
\nI think it can be used in CV , I don't know how. This is an interesting tutorial I came across https://medium.com/swlh/how-to-use-smote-for-dealing-with-imbalanced-image-dataset-for-solving-classification-problems-3aba7d2b9cad
", "rawMarkdown": "I think it can be used in CV , I don't know how. This is an interesting tutorial I came across https://medium.com/swlh/how-to-use-smote-for-dealing-with-imbalanced-image-dataset-for-solving-classification-problems-3aba7d2b9cad", "votes": null }, { "id": "1101870", "postDate": "12/04/2020 10:53:55", "content": "Very important information, learned
", "rawMarkdown": "Very important information, learned", "votes": null }, { "id": "1101963", "postDate": "12/04/2020 12:59:44", "content": "I admit I haven't tested it thoroughly yet. My guess is that upsample of minority classes plus augementations may help.
\n train_images_ = df_train.loc[df_train['label']!=3, ['image_id', 'label']]\n train_images_m = [os.path.join(training_data_path, i) for i in train_images_.image_id.values.tolist()]\n train_targets_m = train_images_['label'] \n\n for i in range(N_UPSAMPLE):\n train_images = train_images + train_images_m\n train_targets = np.concatenate([train_targets, train_targets_m])\nSo we add minority images to the train dataset and then apply augmentations to the whole train dataset, mimicing SMOTE in a way.
", "rawMarkdown": "I admit I haven't tested it thoroughly yet. My guess is that upsample of minority classes plus augementations may help.\n\n\n train_images_ = df_train.loc[df_train['label']!=3, ['image_id', 'label']]\n train_images_m = [os.path.join(training_data_path, i) for i in train_images_.image_id.values.tolist()]\n train_targets_m = train_images_['label'] \n\n for i in range(N_UPSAMPLE):\n train_images = train_images + train_images_m\n train_targets = np.concatenate([train_targets, train_targets_m])\n\n\n\nSo we add minority images to the train dataset and then apply augmentations to the whole train dataset, mimicing SMOTE in a way.", "votes": null }, { "id": "1101981", "postDate": "12/04/2020 13:12:40", "content": "Yes, this might work 👍
", "rawMarkdown": "Yes, this might work 👍", "votes": null }, { "id": "1102023", "postDate": "12/04/2020 13:46:51", "content": "I mean something like this. In one of earlier versions I tried to turn on upsample (UPSAMPLE=True, N_UPSAMPLE = 1) but results got worse (higher overfitting). I assume there was an issue with a model itself and upsample just made it worse. As soon as I stabilize my model, I want to try UPSAMPLE=True again. My gut feeling tells me it won't help and if I want to upsample minority, I need high-quality synthetic images (GAN etc) anyway. But who knows.
", "rawMarkdown": "I mean something like [this](https://www.kaggle.com/dunklerwald/pytorch-efficientnet-with-tta-training). In one of earlier versions I tried to turn on upsample (UPSAMPLE=True, N_UPSAMPLE = 1) but results got worse (higher overfitting). I assume there was an issue with a model itself and upsample just made it worse. As soon as I stabilize my model, I want to try UPSAMPLE=True again. My gut feeling tells me it won't help and if I want to upsample minority, I need high-quality synthetic images (GAN etc) anyway. But who knows.", "votes": null } ], "comments": [ { "id": 1101431, "author_name": "aliabdin1", "author_url": "", "post_date": "12/03/2020 22:28:19", "content": "Paper if anyone is interested: https://arxiv.org/pdf/1106.1813.pdf
", "votes": null, "replies": [] }, { "id": 1101786, "author_name": "fire15", "author_url": "", "post_date": "12/04/2020 09:01:19", "content": "I only know some structured data could use SMOTE, like in some data mining competition.
\nIn CV field, how to use SMOTE is a question, maybe by some data augmentation?
\nHope someone else know that.
I think it can be used in CV , I don't know how. This is an interesting tutorial I came across https://medium.com/swlh/how-to-use-smote-for-dealing-with-imbalanced-image-dataset-for-solving-classification-problems-3aba7d2b9cad
", "votes": null, "replies": [] }, { "id": 1101870, "author_name": "zhangeng", "author_url": "", "post_date": "12/04/2020 10:53:55", "content": "Very important information, learned
", "votes": null, "replies": [] } ] }, { "id": 1101856, "author_name": "zhangeng", "author_url": "", "post_date": "12/04/2020 10:37:07", "content": "Since the distribution of the minority categories is similar, I plan to enhance the minority categories uniformly and ignore the majority categories.
\nI admit I haven't tested it thoroughly yet. My guess is that upsample of minority classes plus augementations may help.
\n train_images_ = df_train.loc[df_train['label']!=3, ['image_id', 'label']]\n train_images_m = [os.path.join(training_data_path, i) for i in train_images_.image_id.values.tolist()]\n train_targets_m = train_images_['label'] \n\n for i in range(N_UPSAMPLE):\n train_images = train_images + train_images_m\n train_targets = np.concatenate([train_targets, train_targets_m])\nSo we add minority images to the train dataset and then apply augmentations to the whole train dataset, mimicing SMOTE in a way.
", "votes": null, "replies": [ { "id": 1101981, "author_name": "akshatdevve", "author_url": "", "post_date": "12/04/2020 13:12:40", "content": "Yes, this might work 👍
", "votes": null, "replies": [] }, { "id": 1102023, "author_name": "dunklerwald", "author_url": "", "post_date": "12/04/2020 13:46:51", "content": "I mean something like this. In one of earlier versions I tried to turn on upsample (UPSAMPLE=True, N_UPSAMPLE = 1) but results got worse (higher overfitting). I assume there was an issue with a model itself and upsample just made it worse. As soon as I stabilize my model, I want to try UPSAMPLE=True again. My gut feeling tells me it won't help and if I want to upsample minority, I need high-quality synthetic images (GAN etc) anyway. But who knows.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1099156": "Has anyone tried SMOTE on the data , since using class_weights and kfolds is not helping much.", "1101431": "Paper if anyone is interested: https://arxiv.org/pdf/1106.1813.pdf", "1101786": "I only know some structured data could use SMOTE, like in some data mining competition.\nIn CV field, how to use SMOTE is a question, maybe by some data augmentation? \nHope someone else know that.", "1101856": "Since the distribution of the minority categories is similar, I plan to enhance the minority categories uniformly and ignore the majority categories.\n\n1. First classify the 5 categories of image data into 5 categories\n2. Images in the most frequent category are not considered for enhancement\n3. Unified enhancement of images in a few categories (because the number of other categories is similar)\ncode show as below:\nif class==3:\n pass\nelse:\n Data enhancement\nThis is my initial thoughts.", "1101864": "I think it can be used in CV , I don't know how. This is an interesting tutorial I came across https://medium.com/swlh/how-to-use-smote-for-dealing-with-imbalanced-image-dataset-for-solving-classification-problems-3aba7d2b9cad", "1101870": "Very important information, learned", "1101963": "I admit I haven't tested it thoroughly yet. My guess is that upsample of minority classes plus augementations may help.\n\n\n train_images_ = df_train.loc[df_train['label']!=3, ['image_id', 'label']]\n train_images_m = [os.path.join(training_data_path, i) for i in train_images_.image_id.values.tolist()]\n train_targets_m = train_images_['label'] \n\n for i in range(N_UPSAMPLE):\n train_images = train_images + train_images_m\n train_targets = np.concatenate([train_targets, train_targets_m])\n\n\n\nSo we add minority images to the train dataset and then apply augmentations to the whole train dataset, mimicing SMOTE in a way.", "1101981": "Yes, this might work 👍", "1102023": "I mean something like [this](https://www.kaggle.com/dunklerwald/pytorch-efficientnet-with-tta-training). In one of earlier versions I tried to turn on upsample (UPSAMPLE=True, N_UPSAMPLE = 1) but results got worse (higher overfitting). I assume there was an issue with a model itself and upsample just made it worse. As soon as I stabilize my model, I want to try UPSAMPLE=True again. My gut feeling tells me it won't help and if I want to upsample minority, I need high-quality synthetic images (GAN etc) anyway. But who knows." }, "source": "meta" }