I just tried the code based on our code on salt competition, and found the IOU is always 1 when train. Is there something that I should aware of for this competition?
\n\nAnd when use such model to predict, I found the result is always empty mask. Is this because I downsized the image and the mask?
", "messages": [ { "id": "412292", "postDate": "10/29/2018 23:47:31", "content": "I just tried the code based on our code on salt competition, and found the IOU is always 1 when train. Is there something that I should aware of for this competition?
\n\nAnd when use such model to predict, I found the result is always empty mask. Is this because I downsized the image and the mask?
", "rawMarkdown": "I just tried the code based on our code on salt competition, and found the IOU is always 1 when train. Is there something that I should aware of for this competition?\n\nAnd when use such model to predict, I found the result is always empty mask. Is this because I downsized the image and the mask?", "votes": null }, { "id": "412298", "postDate": "10/30/2018 00:15:26", "content": "Can it be that there is a problem in data loader, and all masks are just empty?
", "rawMarkdown": "Can it be that there is a problem in data loader, and all masks are just empty?", "votes": null }, { "id": "412303", "postDate": "10/30/2018 00:27:11", "content": "I am wondering when you train your model, what's the metrics you were using?
", "rawMarkdown": "I am wondering when you train your model, what's the metrics you were using?", "votes": null }, { "id": "412311", "postDate": "10/30/2018 00:57:54", "content": "You can check my kernel for more details, but during training I look mostly at dice (and IoU that is related to dice): it is good enough to see the relative performance of the model when I adjust training parameters. I check F2 metric, one used in the competition, only occupationally, once per 8-16 epochs and when I look into post-processing procedure.
", "rawMarkdown": "You can check [my kernel][1] for more details, but during training I look mostly at dice (and IoU that is related to dice): it is good enough to see the relative performance of the model when I adjust training parameters. I check F2 metric, one used in the competition, only occupationally, once per 8-16 epochs and when I look into post-processing procedure.\n\n\n [1]: https://www.kaggle.com/iafoss/unet34-dice-0-87", "votes": null }, { "id": "412931", "postDate": "10/31/2018 02:44:09", "content": "Not sure what your IOU code is - but I used my salt code and have been using mean_iou for a metric in model.fit. It does report values that look reasonable. I think my mean_iou script was taken from a early kernel and was common to many of the kernels in the Salt challenge.
", "rawMarkdown": "Not sure what your IOU code is - but I used my salt code and have been using mean_iou for a metric in model.fit. It does report values that look reasonable. I think my mean_iou script was taken from a early kernel and was common to many of the kernels in the Salt challenge.", "votes": null }, { "id": "412972", "postDate": "10/31/2018 04:20:24", "content": "Thanks for your reply. I found a bug when I load my masks. But still has some issue for my training process as IOU still can go to 1= =
", "rawMarkdown": "Thanks for your reply. I found a bug when I load my masks. But still has some issue for my training process as IOU still can go to 1= =", "votes": null }, { "id": "413154", "postDate": "10/31/2018 11:32:59", "content": "@Strideradu remember that in the TGS Salt there was just a binary 0-1 mask that you would run RLE on and submit. There was just one mask per image, even if disconnected.
\n\nIn this competition, however, you need to submit a mask per ship, so the mask on which your run RLE is 0-background 1-k ships.
\n\nIf that is the case all you need to do is run ndi.label on the binarized output from your network and you should be fine.
Can it be that there is a problem in data loader, and all masks are just empty?
", "votes": null, "replies": [ { "id": 412303, "author_name": "strideradu", "author_url": "", "post_date": "10/30/2018 00:27:11", "content": "I am wondering when you train your model, what's the metrics you were using?
", "votes": null, "replies": [] }, { "id": 412311, "author_name": "iafoss", "author_url": "", "post_date": "10/30/2018 00:57:54", "content": "You can check my kernel for more details, but during training I look mostly at dice (and IoU that is related to dice): it is good enough to see the relative performance of the model when I adjust training parameters. I check F2 metric, one used in the competition, only occupationally, once per 8-16 epochs and when I look into post-processing procedure.
", "votes": null, "replies": [] } ] }, { "id": 412931, "author_name": "pcjimmmy", "author_url": "", "post_date": "10/31/2018 02:44:09", "content": "Not sure what your IOU code is - but I used my salt code and have been using mean_iou for a metric in model.fit. It does report values that look reasonable. I think my mean_iou script was taken from a early kernel and was common to many of the kernels in the Salt challenge.
", "votes": null, "replies": [ { "id": 412972, "author_name": "strideradu", "author_url": "", "post_date": "10/31/2018 04:20:24", "content": "Thanks for your reply. I found a bug when I load my masks. But still has some issue for my training process as IOU still can go to 1= =
", "votes": null, "replies": [] } ] }, { "id": 413154, "author_name": "jakubczakon", "author_url": "", "post_date": "10/31/2018 11:32:59", "content": "@Strideradu remember that in the TGS Salt there was just a binary 0-1 mask that you would run RLE on and submit. There was just one mask per image, even if disconnected.
\n\nIn this competition, however, you need to submit a mask per ship, so the mask on which your run RLE is 0-background 1-k ships.
\n\nIf that is the case all you need to do is run ndi.label on the binarized output from your network and you should be fine.