tiff files have 3 different resolution : highest (1*1), median (1/4 * 1/4) and lowest (1/16, 1/16)\nIn my experiment, I can get 0.90-0.91 CV and 0.86-0.87 LB (single fold with TTA) using median resolution (on one GTX1080). It is necessary or possible to use the highest resolution to improve the score? Could you share your experiment result?
", "messages": [ { "id": 882719, "postDate": "2020-06-12T05:00:10.007Z", "content": "tiff files have 3 different resolution : highest (1*1), median (1/4 * 1/4) and lowest (1/16, 1/16)\nIn my experiment, I can get 0.90-0.91 CV and 0.86-0.87 LB (single fold with TTA) using median resolution (on one GTX1080). It is necessary or possible to use the highest resolution to improve the score? Could you share your experiment result?
", "rawMarkdown": "tiff files have 3 different resolution : highest (1*1), median (1/4 * 1/4) and lowest (1/16, 1/16)\nIn my experiment, I can get 0.90-0.91 CV and 0.86-0.87 LB (single fold with TTA) using median resolution (on one GTX1080). It is necessary or possible to use the highest resolution to improve the score? Could you share your experiment result?", "votes": 15 }, { "id": 883737, "postDate": "2020-06-12T21:37:59.483Z", "content": "Highest and median resolutions are about the same for me. My current LB (0.91) is using median resolution only. I actually gave up on using highest resolution because it takes longer to do experiments, and it didn't seem like there was huge benefit over median resolution.
", "rawMarkdown": "Highest and median resolutions are about the same for me. My current LB (0.91) is using median resolution only. I actually gave up on using highest resolution because it takes longer to do experiments, and it didn't seem like there was huge benefit over median resolution.", "votes": 13, "replies": [ { "id": 914971, "postDate": "2020-07-04T11:52:31.277Z", "content": "Could I ask how many models you use to get 0.93 LB？ All the models use the median resolutions？
", "rawMarkdown": "Could I ask how many models you use to get 0.93 LB？ All the models use the median resolutions？", "votes": 1 } ] }, { "id": 884806, "postDate": "2020-06-13T16:28:39.920Z", "content": "I am using median resolution, in my experiment, I got 0.88x CV and 0.90 LB, also single fold with TTA. I haven't tried high resolution. I think it's necessary to use high resolution, but it takes too long to process the images.
", "rawMarkdown": "I am using median resolution, in my experiment, I got 0.88x CV and 0.90 LB, also single fold with TTA. I haven't tried high resolution. I think it's necessary to use high resolution, but it takes too long to process the images.", "votes": 5, "replies": [ { "id": 888858, "postDate": "2020-06-16T15:46:11.520Z", "content": "Can you please share number of patches and their resolutions?
", "rawMarkdown": "Can you please share number of patches and their resolutions? " }, { "id": 891414, "postDate": "2020-06-18T06:59:47.830Z", "content": "I used 32 patches of median resolution.
", "rawMarkdown": "I used 32 patches of median resolution." } ] }, { "id": 883134, "postDate": "2020-06-12T11:53:07.943Z", "content": "I was able to get on lowest resolution CV/LB ~ around 0.86, medium resolution CV/LB ~0.90.e I did one experiment with high resolution and got local CV of ~0.88. I think it will be important to use high resolution since it contain more details on cell morphology, but again question remains how to effectively utilize it =)
Can you explain how to get the resolutions resolutions briefly?
", "rawMarkdown": "Can you explain how to get the resolutions resolutions briefly?", "votes": 1 }, { "id": 883447, "postDate": "2020-06-12T16:30:32.780Z", "content": "Nice job, didn't know it's possible to get 0.90 LB with medium resolution.
", "rawMarkdown": "Nice job, didn't know it's possible to get 0.90 LB with medium resolution.", "votes": 1 }, { "id": 883806, "postDate": "2020-06-13T00:37:08.093Z", "content": "@drhabib are you using full images or tiles?
", "rawMarkdown": "@drhabib are you using full images or tiles?", "votes": 1 }, { "id": 884920, "postDate": "2020-06-13T18:19:10.010Z", "content": "tiles
", "rawMarkdown": "tiles", "votes": 1 }, { "id": 890835, "postDate": "2020-06-17T18:09:25.227Z", "content": "Is it possible for me to know the batch size and the graphics card you are using?
", "rawMarkdown": "Is it possible for me to know the batch size and the graphics card you are using?" } ] }, { "id": 887278, "postDate": "2020-06-15T15:34:52.640Z", "content": "I only use medium resolution. \nHigh resolution, it takes a long time to process, but I couldn't confirm the effectiveness.
", "rawMarkdown": "I only use medium resolution. \nHigh resolution, it takes a long time to process, but I couldn't confirm the effectiveness. ", "votes": 1 }, { "id": 883491, "postDate": "2020-06-12T17:21:29.820Z", "content": "I used median layer to get lb of 0.90 while cv is 0.885. It seems like your model is overfitting to CV.
", "rawMarkdown": "I used median layer to get lb of 0.90 while cv is 0.885. It seems like your model is overfitting to CV.", "votes": 1 }, { "id": 883228, "postDate": "2020-06-12T13:16:41.863Z", "content": "@seefun My situation is pretty much the same as yours. I was having around 0.82 with the lowest resolution and got around 0.86 to 0.88 with the median resolution (with one GTX1080ti). My CV is a bit lower than yours usually in the 0.88 range. Worth noting that from one fold to another I easily have 0.02 differences.
\n\nI tried same number of tiles/size on the highest resolution downscaled to 1/3 and got a slightly lower 0.85 result from it.
", "rawMarkdown": "@seefun My situation is pretty much the same as yours. I was having around 0.82 with the lowest resolution and got around 0.86 to 0.88 with the median resolution (with one GTX1080ti). My CV is a bit lower than yours usually in the 0.88 range. Worth noting that from one fold to another I easily have 0.02 differences.\n\nI tried same number of tiles/size on the highest resolution downscaled to 1/3 and got a slightly lower 0.85 result from it.", "votes": 1 }, { "id": 882893, "postDate": "2020-06-12T08:09:52.653Z", "content": "the tiff files are MultImage array, are the resolutions stored in the array elements?
", "rawMarkdown": "the tiff files are MultImage array, are the resolutions stored in the array elements?", "replies": [ { "id": 887737, "postDate": "2020-06-15T20:45:37.173Z", "content": "img = skimage.io.MultiImage(os.path.join(TRAIN, name+'.tiff'))[level]
\n\nwhere level = 0, 1 or 2. Take 0 for the highest resolution.
", "rawMarkdown": "img = skimage.io.MultiImage(os.path.join(TRAIN, name+'.tiff'))[level]\n\nwhere level = 0, 1 or 2. Take 0 for the highest resolution. ", "votes": 3 } ] }, { "id": 891235, "postDate": "2020-06-18T03:24:23.913Z", "content": "Increasing Patch Size from 256 to 512 on median layer increase my LOSS. \nKind of disturbing because I don't have resources so have to manage all this huge data on colab...
", "rawMarkdown": "Increasing Patch Size from 256 to 512 on median layer increase my LOSS. \nKind of disturbing because I don't have resources so have to manage all this huge data on colab...", "replies": [ { "id": 904227, "postDate": "2020-06-27T13:00:19.820Z", "content": "hi @micheomaano are you using ColabPro?
", "rawMarkdown": "hi @micheomaano are you using ColabPro?\n" } ] } ], "comments": [ { "id": 883737, "author_name": "Ian Pan", "author_url": "", "post_date": "2020-06-12T21:37:59.483000", "content": "Highest and median resolutions are about the same for me. My current LB (0.91) is using median resolution only. I actually gave up on using highest resolution because it takes longer to do experiments, and it didn't seem like there was huge benefit over median resolution.
", "votes": 13, "replies": [ { "id": 914971, "author_name": "seefun", "author_url": "", "post_date": "2020-07-04T11:52:31.277000", "content": "Could I ask how many models you use to get 0.93 LB？ All the models use the median resolutions？
", "votes": 1, "replies": [] } ] }, { "id": 884806, "author_name": "", "author_url": "", "post_date": "2020-06-13T16:28:39.920000", "content": "I am using median resolution, in my experiment, I got 0.88x CV and 0.90 LB, also single fold with TTA. I haven't tried high resolution. I think it's necessary to use high resolution, but it takes too long to process the images.
", "votes": 5, "replies": [ { "id": 888858, "author_name": "Salman", "author_url": "", "post_date": "2020-06-16T15:46:11.520000", "content": "Can you please share number of patches and their resolutions?
", "votes": 0, "replies": [] }, { "id": 891414, "author_name": "", "author_url": "", "post_date": "2020-06-18T06:59:47.830000", "content": "I used 32 patches of median resolution.
", "votes": 0, "replies": [] } ] }, { "id": 883134, "author_name": "DrHB", "author_url": "", "post_date": "2020-06-12T11:53:07.943000", "content": "I was able to get on lowest resolution CV/LB ~ around 0.86, medium resolution CV/LB ~0.90.e I did one experiment with high resolution and got local CV of ~0.88. I think it will be important to use high resolution since it contain more details on cell morphology, but again question remains how to effectively utilize it =)
Can you explain how to get the resolutions resolutions briefly?
", "votes": 1, "replies": [] }, { "id": 883447, "author_name": "Konstantin Lopukhin", "author_url": "", "post_date": "2020-06-12T16:30:32.780000", "content": "Nice job, didn't know it's possible to get 0.90 LB with medium resolution.
", "votes": 1, "replies": [] }, { "id": 883806, "author_name": "Dracarys", "author_url": "", "post_date": "2020-06-13T00:37:08.093000", "content": "@drhabib are you using full images or tiles?
", "votes": 1, "replies": [] }, { "id": 884920, "author_name": "DrHB", "author_url": "", "post_date": "2020-06-13T18:19:10.010000", "content": "tiles
", "votes": 1, "replies": [] }, { "id": 890835, "author_name": "DeepLearner", "author_url": "", "post_date": "2020-06-17T18:09:25.227000", "content": "Is it possible for me to know the batch size and the graphics card you are using?
", "votes": 0, "replies": [] } ] }, { "id": 887278, "author_name": "fam_taro", "author_url": "", "post_date": "2020-06-15T15:34:52.640000", "content": "I only use medium resolution. \nHigh resolution, it takes a long time to process, but I couldn't confirm the effectiveness.
", "votes": 1, "replies": [] }, { "id": 883491, "author_name": "Shujun", "author_url": "", "post_date": "2020-06-12T17:21:29.820000", "content": "I used median layer to get lb of 0.90 while cv is 0.885. It seems like your model is overfitting to CV.
", "votes": 1, "replies": [] }, { "id": 883228, "author_name": "Arnaud Roussel", "author_url": "", "post_date": "2020-06-12T13:16:41.863000", "content": "@seefun My situation is pretty much the same as yours. I was having around 0.82 with the lowest resolution and got around 0.86 to 0.88 with the median resolution (with one GTX1080ti). My CV is a bit lower than yours usually in the 0.88 range. Worth noting that from one fold to another I easily have 0.02 differences.
\n\nI tried same number of tiles/size on the highest resolution downscaled to 1/3 and got a slightly lower 0.85 result from it.
", "votes": 1, "replies": [] }, { "id": 882893, "author_name": "Tasnim Nishat Islam", "author_url": "", "post_date": "2020-06-12T08:09:52.653000", "content": "the tiff files are MultImage array, are the resolutions stored in the array elements?
", "votes": 0, "replies": [ { "id": 887737, "author_name": "Maxime Markov", "author_url": "", "post_date": "2020-06-15T20:45:37.173000", "content": "img = skimage.io.MultiImage(os.path.join(TRAIN, name+'.tiff'))[level]
\n\nwhere level = 0, 1 or 2. Take 0 for the highest resolution.
", "votes": 3, "replies": [] } ] }, { "id": 891235, "author_name": "Salman", "author_url": "", "post_date": "2020-06-18T03:24:23.913000", "content": "Increasing Patch Size from 256 to 512 on median layer increase my LOSS. \nKind of disturbing because I don't have resources so have to manage all this huge data on colab...
", "votes": 0, "replies": [ { "id": 904227, "author_name": "", "author_url": "", "post_date": "2020-06-27T13:00:19.820000", "content": "hi @micheomaano are you using ColabPro?
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "882719": "tiff files have 3 different resolution : highest (1*1), median (1/4 * 1/4) and lowest (1/16, 1/16)\nIn my experiment, I can get 0.90-0.91 CV and 0.86-0.87 LB (single fold with TTA) using median resolution (on one GTX1080). It is necessary or possible to use the highest resolution to improve the score? Could you share your experiment result?", "883737": "Highest and median resolutions are about the same for me. My current LB (0.91) is using median resolution only. I actually gave up on using highest resolution because it takes longer to do experiments, and it didn't seem like there was huge benefit over median resolution.", "884806": "I am using median resolution, in my experiment, I got 0.88x CV and 0.90 LB, also single fold with TTA. I haven't tried high resolution. I think it's necessary to use high resolution, but it takes too long to process the images.", "883134": "I was able to get on lowest resolution `CV/LB` ~ around `0.86`, medium resolution `CV/LB` ~` 0.90`.e I did one experiment with high resolution and got local `CV` of ~0.88. I think it will be important to use `high resolution` since it contain more details on cell morphology, but again question remains how to effectively utilize it =) ", "887278": "I only use medium resolution. \nHigh resolution, it takes a long time to process, but I couldn't confirm the effectiveness. ", "883491": "I used median layer to get lb of 0.90 while cv is 0.885. It seems like your model is overfitting to CV.", "883228": "@seefun My situation is pretty much the same as yours. I was having around 0.82 with the lowest resolution and got around 0.86 to 0.88 with the median resolution (with one GTX1080ti). My CV is a bit lower than yours usually in the 0.88 range. Worth noting that from one fold to another I easily have 0.02 differences.\n\nI tried same number of tiles/size on the highest resolution downscaled to 1/3 and got a slightly lower 0.85 result from it.", "882893": "the tiff files are MultImage array, are the resolutions stored in the array elements?", "891235": "Increasing Patch Size from 256 to 512 on median layer increase my LOSS. \nKind of disturbing because I don't have resources so have to manage all this huge data on colab..." } }