To solve the imbalanced classes problem, we can use compute_class_weight() to calculate the class weights of multiclass, but it doesn't work for multilabel.\nIn this case, 'compute_sample_weight()' is a good choice.\n\nfrom sklearn.utils.class_weight import compute_sample_weight\ny_train_weight = compute_sample_weight(\"balanced\", y_train)\n
The blood vessels usually have high contrast in the green channel, anyone tried to train the model by the single green channel?
", "messages": [ { "id": "588630", "postDate": "07/30/2019 20:49:07", "content": "To solve the imbalanced classes problem, we can use compute_class_weight() to calculate the class weights of multiclass, but it doesn't work for multilabel.\nIn this case, 'compute_sample_weight()' is a good choice.\n\nfrom sklearn.utils.class_weight import compute_sample_weight\ny_train_weight = compute_sample_weight(\"balanced\", y_train)\n
The blood vessels usually have high contrast in the green channel, anyone tried to train the model by the single green channel?
", "rawMarkdown": "## compute class weights\nTo solve the imbalanced classes problem, we can use `compute_class_weight()` to calculate the class weights of multiclass, but it doesn't work for multilabel.\nIn this case, 'compute_sample_weight()' is a good choice.\n```\nfrom sklearn.utils.class_weight import compute_sample_weight\ny_train_weight = compute_sample_weight(\"balanced\", y_train)\n```\n\n# Single channel training\nThe blood vessels usually have high contrast in the green channel, anyone tried to train the model by the single green channel?", "votes": null }, { "id": "588890", "postDate": "07/31/2019 07:45:19", "content": "Yes, I am training on single green channel, it \n- reduces memory consumption a lot\n- permits to load all of test data in kernel memory while submission, eradicating data generators and speeding kernel rerun time\n- does not lose any vital information in data\n- performs as well as 3 channel if not well. (Hint: it can perform better than 3 channel with proper preprocessing, provided all other factors remain same)\n- Allows for comparatively less parameters in model ( my current score is with model having ~1.1 million params, with no TTA, no optimized threshold and without pre-trained model.)
", "rawMarkdown": "Yes, I am training on single green channel, it \n- reduces memory consumption a lot\n- permits to load all of test data in kernel memory while submission, eradicating data generators and speeding kernel rerun time\n- does not lose any vital information in data\n- performs as well as 3 channel if not well. (Hint: it can perform better than 3 channel with proper preprocessing, provided all other factors remain same)\n- Allows for comparatively less parameters in model ( my current score is with model having ~1.1 million params, with no TTA, no optimized threshold and without pre-trained model.)", "votes": null }, { "id": "588971", "postDate": "07/31/2019 09:38:07", "content": "Agree! But, the loss reduced slowly. So I tried to histequalize green channel first and then copyed it to three channels, it was a little bit faster. At last, I got CV:0.9126 LB:0.757 (The efficientNetB3 with ImageNet weights)
", "rawMarkdown": "Agree! But, the loss reduced slowly. So I tried to histequalize green channel first and then copyed it to three channels, it was a little bit faster. At last, I got CV:0.9126 LB:0.757 (The efficientNetB3 with ImageNet weights)", "votes": null }, { "id": "588983", "postDate": "07/31/2019 10:07:09", "content": "I am also doing histequalization of green channel. But I think that copying is bit redundant, also it makes image bit darker, making it difficult to detect small microaneurysms, which are present in mild DR. I strongly believe that we can get > 0.75(LB) with much smaller single model (in terms of parameters) and some training tricks.
", "rawMarkdown": "I am also doing histequalization of green channel. But I think that copying is bit redundant, also it makes image bit darker, making it difficult to detect small microaneurysms, which are present in mild DR. I strongly believe that we can get > 0.75(LB) with much smaller single model (in terms of parameters) and some training tricks.", "votes": null }, { "id": "588992", "postDate": "07/31/2019 10:19:06", "content": "It makes sense. I will try to train the single-channel and then we can discuss it. Currently, I am using a part of the old dataset (2015) to pre-train the model, it may help you as well.\nDataset Link:\nhttps://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resized
", "rawMarkdown": "It makes sense. I will try to train the single-channel and then we can discuss it. Currently, I am using a part of the old dataset (2015) to pre-train the model, it may help you as well.\nDataset Link:\nhttps://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resized", "votes": null }, { "id": "589003", "postDate": "07/31/2019 10:32:24", "content": "Thanks!
", "rawMarkdown": "Thanks!", "votes": null }, { "id": "594925", "postDate": "08/08/2019 16:36:15", "content": "Hey can you explain this a bit more. the 5 values this function outputs,how are we supposed to use this?Insert it in our loss function?
", "rawMarkdown": "Hey can you explain this a bit more. the 5 values this function outputs,how are we supposed to use this?Insert it in our loss function?", "votes": null } ], "comments": [ { "id": 588890, "author_name": "ashwan1", "author_url": "", "post_date": "07/31/2019 07:45:19", "content": "Yes, I am training on single green channel, it \n- reduces memory consumption a lot\n- permits to load all of test data in kernel memory while submission, eradicating data generators and speeding kernel rerun time\n- does not lose any vital information in data\n- performs as well as 3 channel if not well. (Hint: it can perform better than 3 channel with proper preprocessing, provided all other factors remain same)\n- Allows for comparatively less parameters in model ( my current score is with model having ~1.1 million params, with no TTA, no optimized threshold and without pre-trained model.)
", "votes": null, "replies": [ { "id": 588971, "author_name": "xiewen112", "author_url": "", "post_date": "07/31/2019 09:38:07", "content": "Agree! But, the loss reduced slowly. So I tried to histequalize green channel first and then copyed it to three channels, it was a little bit faster. At last, I got CV:0.9126 LB:0.757 (The efficientNetB3 with ImageNet weights)
", "votes": null, "replies": [] }, { "id": 588983, "author_name": "ashwan1", "author_url": "", "post_date": "07/31/2019 10:07:09", "content": "I am also doing histequalization of green channel. But I think that copying is bit redundant, also it makes image bit darker, making it difficult to detect small microaneurysms, which are present in mild DR. I strongly believe that we can get > 0.75(LB) with much smaller single model (in terms of parameters) and some training tricks.
", "votes": null, "replies": [] }, { "id": 588992, "author_name": "xiewen112", "author_url": "", "post_date": "07/31/2019 10:19:06", "content": "It makes sense. I will try to train the single-channel and then we can discuss it. Currently, I am using a part of the old dataset (2015) to pre-train the model, it may help you as well.\nDataset Link:\nhttps://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resized
", "votes": null, "replies": [] }, { "id": 589003, "author_name": "ashwan1", "author_url": "", "post_date": "07/31/2019 10:32:24", "content": "Thanks!
", "votes": null, "replies": [] } ] }, { "id": 594925, "author_name": "decentmakeover", "author_url": "", "post_date": "08/08/2019 16:36:15", "content": "Hey can you explain this a bit more. the 5 values this function outputs,how are we supposed to use this?Insert it in our loss function?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "588630": "## compute class weights\nTo solve the imbalanced classes problem, we can use `compute_class_weight()` to calculate the class weights of multiclass, but it doesn't work for multilabel.\nIn this case, 'compute_sample_weight()' is a good choice.\n```\nfrom sklearn.utils.class_weight import compute_sample_weight\ny_train_weight = compute_sample_weight(\"balanced\", y_train)\n```\n\n# Single channel training\nThe blood vessels usually have high contrast in the green channel, anyone tried to train the model by the single green channel?", "588890": "Yes, I am training on single green channel, it \n- reduces memory consumption a lot\n- permits to load all of test data in kernel memory while submission, eradicating data generators and speeding kernel rerun time\n- does not lose any vital information in data\n- performs as well as 3 channel if not well. (Hint: it can perform better than 3 channel with proper preprocessing, provided all other factors remain same)\n- Allows for comparatively less parameters in model ( my current score is with model having ~1.1 million params, with no TTA, no optimized threshold and without pre-trained model.)", "588971": "Agree! But, the loss reduced slowly. So I tried to histequalize green channel first and then copyed it to three channels, it was a little bit faster. At last, I got CV:0.9126 LB:0.757 (The efficientNetB3 with ImageNet weights)", "588983": "I am also doing histequalization of green channel. But I think that copying is bit redundant, also it makes image bit darker, making it difficult to detect small microaneurysms, which are present in mild DR. I strongly believe that we can get > 0.75(LB) with much smaller single model (in terms of parameters) and some training tricks.", "588992": "It makes sense. I will try to train the single-channel and then we can discuss it. Currently, I am using a part of the old dataset (2015) to pre-train the model, it may help you as well.\nDataset Link:\nhttps://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resized", "589003": "Thanks!", "594925": "Hey can you explain this a bit more. the 5 values this function outputs,how are we supposed to use this?Insert it in our loss function?" }, "source": "meta" }