I believe we r in a stage where everybody has started ensembling their models. Here are some ways of ensembling that may help u achieve a better score.
\n\nLet's say u have two models: one for classification and another for regression\nThe simplest way to ensemble them would be to average their predictions
python\nmodel_preds = []\nfor model in [regression_model, classification_model]:\n pred = model.predict(test_image)\n model_preds.append(pred)\nfinal_pred = take_average(model_preds)\n
Another popular way would be combine predictions of different version(augmentations) of the same image; popularly known as TTA \npython\nmodel = regression or classification \naug_preds = []\nfor augmentations in [augmentations_1, augmentations_2]:\n augment_image = do_augment(test_image, augmentations)\n pred = model.predict(augment_image)\n aug_preds.append(pred)\nfinal_pred = take_average(aug_preds)\n
This is something that I recently learnt: make predictions on different scales of the same image\npython\nmodel = regression or classification \nscale_preds = []\nfor img_size in [img_size_1, img_size_2]:\n resized_image = do_scale(test_image, img_size)\n pred = model.predict(resized_image)\n scale_preds.append(pred)\nfinal_pred = take_average(scale_preds)\n
You can be creative with ur approach of ensembling by combing different ensembling methods.\npython\ncombine_preds = []\nfor model in [model_1, model_2]:\n for augmentation in [augmentations_1, augmentations_2]:\n for img_size in [img_size_1, img_size_2]:\n sc_aug_image = scale_n_augment(test_image, img_size, augmentation)\n pred = model.predict(sc_aug_image)\n combine_preds.append(pred)\nfinal_pred = take_average(combine_preds)\n\nThese are some of the ensembling approaches that I can think of. Please share if u have some other interesting ways to ensemble.
I believe we r in a stage where everybody has started ensembling their models. Here are some ways of ensembling that may help u achieve a better score.
\n\nLet's say u have two models: one for classification and another for regression\nThe simplest way to ensemble them would be to average their predictions
python\nmodel_preds = []\nfor model in [regression_model, classification_model]:\n pred = model.predict(test_image)\n model_preds.append(pred)\nfinal_pred = take_average(model_preds)\n
Another popular way would be combine predictions of different version(augmentations) of the same image; popularly known as TTA \npython\nmodel = regression or classification \naug_preds = []\nfor augmentations in [augmentations_1, augmentations_2]:\n augment_image = do_augment(test_image, augmentations)\n pred = model.predict(augment_image)\n aug_preds.append(pred)\nfinal_pred = take_average(aug_preds)\n
This is something that I recently learnt: make predictions on different scales of the same image\npython\nmodel = regression or classification \nscale_preds = []\nfor img_size in [img_size_1, img_size_2]:\n resized_image = do_scale(test_image, img_size)\n pred = model.predict(resized_image)\n scale_preds.append(pred)\nfinal_pred = take_average(scale_preds)\n
You can be creative with ur approach of ensembling by combing different ensembling methods.\npython\ncombine_preds = []\nfor model in [model_1, model_2]:\n for augmentation in [augmentations_1, augmentations_2]:\n for img_size in [img_size_1, img_size_2]:\n sc_aug_image = scale_n_augment(test_image, img_size, augmentation)\n pred = model.predict(sc_aug_image)\n combine_preds.append(pred)\nfinal_pred = take_average(combine_preds)\n\nThese are some of the ensembling approaches that I can think of. Please share if u have some other interesting ways to ensemble.
One thing I don't understand with ensembling is if you are doing a classification task and you are using model ensembling , if one model predicts a image belongs to class 3 and another as class 4 if we take average of their predictions won't it be wrong??? can anyone provide more insight on this.
", "rawMarkdown": "One thing I don't understand with ensembling is if you are doing a classification task and you are using model ensembling , if one model predicts a image belongs to class 3 and another as class 4 if we take average of their predictions won't it be wrong??? can anyone provide more insight on this. ", "replies": [ { "id": 605844, "postDate": "2019-08-22T20:08:05.063Z", "content": "Then take an odd number, and if all the predictions are unique you can take the mean or the prediction of your best single model.
", "rawMarkdown": "Then take an odd number, and if all the predictions are unique you can take the mean or the prediction of your best single model." }, { "id": 615644, "postDate": "2019-09-02T07:49:00.190Z", "content": "@seif95 say I have 2 regression models I want to ensemble.. I don't understand how averaging would work here. Say one model predicts 1 and the other 4. Ensembling gives 2.5? Even if I average the numbers before thresholding, 1.2 versus 3.8 gives 2.5? How would that work?
", "rawMarkdown": "@seif95 say I have 2 regression models I want to ensemble.. I don't understand how averaging would work here. Say one model predicts 1 and the other 4. Ensembling gives 2.5? Even if I average the numbers before thresholding, 1.2 versus 3.8 gives 2.5? How would that work?" } ] }, { "id": 603803, "postDate": "2019-08-20T17:18:21.050Z", "content": "Thanks for sharing..!! Very helpful...!!!
", "rawMarkdown": "Thanks for sharing..!! Very helpful...!!!" }, { "id": 603626, "postDate": "2019-08-20T13:59:58.543Z", "content": "Thanks ! Will other mechanisms like majority_vote/mode of models work other than average? What is the preferred one usually?
", "rawMarkdown": "Thanks ! Will other mechanisms like majority_vote/mode of models work other than average? What is the preferred one usually?" }, { "id": 621675, "postDate": "2019-09-08T20:28:48.460Z", "rawMarkdown": "", "isDeleted": true }, { "id": 604344, "postDate": "2019-08-21T09:34:43.533Z", "rawMarkdown": "", "isDeleted": true }, { "id": 603562, "postDate": "2019-08-20T12:51:48.507Z", "content": "thanks a lot very informative
", "rawMarkdown": "thanks a lot very informative", "votes": 1 }, { "id": 603591, "postDate": "2019-08-20T13:25:35.487Z", "content": "thanks @bibek777 \nhelps lot to me..
", "rawMarkdown": "thanks @bibek777 \nhelps lot to me.." }, { "id": 606331, "postDate": "2019-08-23T12:54:00.683Z", "content": "Thanks a lot!
", "rawMarkdown": "Thanks a lot!", "isDeleted": true } ], "comments": [ { "id": 605737, "author_name": "Shivu", "author_url": "", "post_date": "2019-08-22T17:10:52.070000", "content": "One thing I don't understand with ensembling is if you are doing a classification task and you are using model ensembling , if one model predicts a image belongs to class 3 and another as class 4 if we take average of their predictions won't it be wrong??? can anyone provide more insight on this.
", "votes": 0, "replies": [ { "id": 605844, "author_name": "Seifeddine Fezzani", "author_url": "", "post_date": "2019-08-22T20:08:05.063000", "content": "Then take an odd number, and if all the predictions are unique you can take the mean or the prediction of your best single model.
", "votes": 0, "replies": [] }, { "id": 615644, "author_name": "Yousef Rabi", "author_url": "", "post_date": "2019-09-02T07:49:00.190000", "content": "@seif95 say I have 2 regression models I want to ensemble.. I don't understand how averaging would work here. Say one model predicts 1 and the other 4. Ensembling gives 2.5? Even if I average the numbers before thresholding, 1.2 versus 3.8 gives 2.5? How would that work?
", "votes": 0, "replies": [] } ] }, { "id": 603803, "author_name": "Saikat Biswas", "author_url": "", "post_date": "2019-08-20T17:18:21.050000", "content": "Thanks for sharing..!! Very helpful...!!!
", "votes": 0, "replies": [] }, { "id": 603626, "author_name": "Kaushik Ramachandran", "author_url": "", "post_date": "2019-08-20T13:59:58.543000", "content": "Thanks ! Will other mechanisms like majority_vote/mode of models work other than average? What is the preferred one usually?
", "votes": 0, "replies": [] }, { "id": 621675, "author_name": "", "author_url": "", "post_date": "2019-09-08T20:28:48.460000", "content": "", "votes": 0, "replies": [] }, { "id": 604344, "author_name": "", "author_url": "", "post_date": "2019-08-21T09:34:43.533000", "content": "", "votes": 0, "replies": [] }, { "id": 603562, "author_name": "leixiang@AInnovation", "author_url": "", "post_date": "2019-08-20T12:51:48.507000", "content": "thanks a lot very informative
", "votes": 1, "replies": [] }, { "id": 603591, "author_name": "Nitesh Yadav", "author_url": "", "post_date": "2019-08-20T13:25:35.487000", "content": "thanks @bibek777 \nhelps lot to me..
", "votes": 0, "replies": [] }, { "id": 606331, "author_name": "", "author_url": "", "post_date": "2019-08-23T12:54:00.683000", "content": "Thanks a lot!
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "603561": "I believe we r in a stage where everybody has started ensembling their models. Here are some ways of ensembling that may help u achieve a better score.\n### 1. Model Ensemble\nLet's say u have two models: one for `classification` and another for `regression`\nThe simplest way to ensemble them would be to average their predictions\n \n```python\nmodel_preds = []\nfor model in [regression_model, classification_model]:\n pred = model.predict(test_image)\n model_preds.append(pred)\nfinal_pred = take_average(model_preds)\n```\n### 2. Augmentation Ensemble(TTA: Test Time Augmentation)\nAnother popular way would be combine predictions of different version(augmentations) of the same image; popularly known as `TTA` \n```python\nmodel = regression or classification \naug_preds = []\nfor augmentations in [augmentations_1, augmentations_2]:\n augment_image = do_augment(test_image, augmentations)\n pred = model.predict(augment_image)\n aug_preds.append(pred)\nfinal_pred = take_average(aug_preds)\n```\n### 3. Scale(Image Size) Ensemble\nThis is something that I recently learnt: make predictions on different scales of the same image\n```python\nmodel = regression or classification \nscale_preds = []\nfor img_size in [img_size_1, img_size_2]:\n resized_image = do_scale(test_image, img_size)\n pred = model.predict(resized_image)\n scale_preds.append(pred)\nfinal_pred = take_average(scale_preds)\n```\n### 4. Be Creative(Combine everything?)\nYou can be creative with ur approach of ensembling by combing different ensembling methods.\n```python\ncombine_preds = []\nfor model in [model_1, model_2]:\n for augmentation in [augmentations_1, augmentations_2]:\n for img_size in [img_size_1, img_size_2]:\n sc_aug_image = scale_n_augment(test_image, img_size, augmentation)\n pred = model.predict(sc_aug_image)\n combine_preds.append(pred)\nfinal_pred = take_average(combine_preds)\n```\nThese are some of the ensembling approaches that I can think of. Please share if u have some other interesting ways to ensemble.\n\n ", "605737": "One thing I don't understand with ensembling is if you are doing a classification task and you are using model ensembling , if one model predicts a image belongs to class 3 and another as class 4 if we take average of their predictions won't it be wrong??? can anyone provide more insight on this. ", "603803": "Thanks for sharing..!! Very helpful...!!!", "603626": "Thanks ! Will other mechanisms like majority_vote/mode of models work other than average? What is the preferred one usually?", "621675": "", "604344": "", "603562": "thanks a lot very informative", "603591": "thanks @bibek777 \nhelps lot to me..", "606331": "Thanks a lot!" } }