![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2605845%2F572574b6e6c9fce1e8bbccee936abf18%2Fimage.png?generation=1571150483941191&alt=media\")
Hi!
\n\nI'm going towards the community because I just can't seem to figure out what is wrong with my pipeline. I'm using a custom data generator, and Keras using the segmentation models library to make a simple unet. I tried making one that would take all 4 labels at first and it didn't learn correctly. So I tried simplifying and making a model that would only take one label at first to try to see if that would still produce some bad results.\nThese are my loss and dice score measurements:\n
.png?generation=1571150497219032&alt=media)
It seems like the model is overfitting like crazy, but to the point where the loss goes to 1 and the dice score to 0, that sounds like a lot! Especially since the train loss and dice score are doing pretty good. I've checked a lot of different things, so I don't really know where to go now, thus turning to the community.
\n\nHere is the kernel I'm currently running.\nIf you have any ides, I'm really open for suggestions, thanks a lot!
\n\nEDIT: I've tried lowering the learning rate, there still is a massive overfit on validation. When plotting the predictions, all the predicted masks look the same. So there definitely is something wrong there, I just don't know what that is.
", "messages": [ { "id": "649581", "postDate": "10/15/2019 14:44:04", "content": "Hi!
\n\nI'm going towards the community because I just can't seem to figure out what is wrong with my pipeline. I'm using a custom data generator, and Keras using the segmentation models library to make a simple unet. I tried making one that would take all 4 labels at first and it didn't learn correctly. So I tried simplifying and making a model that would only take one label at first to try to see if that would still produce some bad results.\nThese are my loss and dice score measurements:\n
.png?generation=1571150497219032&alt=media)
It seems like the model is overfitting like crazy, but to the point where the loss goes to 1 and the dice score to 0, that sounds like a lot! Especially since the train loss and dice score are doing pretty good. I've checked a lot of different things, so I don't really know where to go now, thus turning to the community.
\n\nHere is the kernel I'm currently running.\nIf you have any ides, I'm really open for suggestions, thanks a lot!
\n\nEDIT: I've tried lowering the learning rate, there still is a massive overfit on validation. When plotting the predictions, all the predicted masks look the same. So there definitely is something wrong there, I just don't know what that is.
", "rawMarkdown": "Hi!\n\nI'm going towards the community because I just can't seem to figure out what is wrong with my pipeline. I'm using a custom data generator, and Keras using the [segmentation models library](https://github.com/qubvel/segmentation_models) to make a simple unet. I tried making one that would take all 4 labels at first and it didn't learn correctly. So I tried simplifying and making a model that would only take one label at first to try to see if that would still produce some bad results.\nThese are my loss and dice score measurements:\n\n \n.png?generation=1571150497219032&alt=media)\n\nIt seems like the model is overfitting like crazy, but to the point where the loss goes to 1 and the dice score to 0, that sounds like a lot! Especially since the train loss and dice score are doing pretty good. I've checked a lot of different things, so I don't really know where to go now, thus turning to the community.\n\n[Here is the kernel](https://www.kaggle.com/maxlenormand/single-class-unet-with-datagenerator-class?scriptVersionId=21997720) I'm currently running.\nIf you have any ides, I'm really open for suggestions, thanks a lot!\n\nEDIT: I've tried lowering the learning rate, there still is a massive overfit on validation. When plotting the predictions, all the predicted masks look the same. So there definitely is something wrong there, I just don't know what that is.", "votes": null }, { "id": "651488", "postDate": "10/17/2019 14:36:38", "content": "Problem seems to be solved.\nThe Semantic Segmentation library seems to have losses build-in, which I used, and that seemed to fix my problem, the validation dice loss goes up now. Pretty weird though, not sure what was wrong with my original loss.
", "rawMarkdown": "Problem seems to be solved.\nThe Semantic Segmentation library seems to have losses build-in, which I used, and that seemed to fix my problem, the validation dice loss goes up now. Pretty weird though, not sure what was wrong with my original loss.", "votes": null } ], "comments": [ { "id": 651488, "author_name": "maxlenormand", "author_url": "", "post_date": "10/17/2019 14:36:38", "content": "Problem seems to be solved.\nThe Semantic Segmentation library seems to have losses build-in, which I used, and that seemed to fix my problem, the validation dice loss goes up now. Pretty weird though, not sure what was wrong with my original loss.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "649581": "Hi!\n\nI'm going towards the community because I just can't seem to figure out what is wrong with my pipeline. I'm using a custom data generator, and Keras using the [segmentation models library](https://github.com/qubvel/segmentation_models) to make a simple unet. I tried making one that would take all 4 labels at first and it didn't learn correctly. So I tried simplifying and making a model that would only take one label at first to try to see if that would still produce some bad results.\nThese are my loss and dice score measurements:\n\n \n.png?generation=1571150497219032&alt=media)\n\nIt seems like the model is overfitting like crazy, but to the point where the loss goes to 1 and the dice score to 0, that sounds like a lot! Especially since the train loss and dice score are doing pretty good. I've checked a lot of different things, so I don't really know where to go now, thus turning to the community.\n\n[Here is the kernel](https://www.kaggle.com/maxlenormand/single-class-unet-with-datagenerator-class?scriptVersionId=21997720) I'm currently running.\nIf you have any ides, I'm really open for suggestions, thanks a lot!\n\nEDIT: I've tried lowering the learning rate, there still is a massive overfit on validation. When plotting the predictions, all the predicted masks look the same. So there definitely is something wrong there, I just don't know what that is.", "651488": "Problem seems to be solved.\nThe Semantic Segmentation library seems to have losses build-in, which I used, and that seemed to fix my problem, the validation dice loss goes up now. Pretty weird though, not sure what was wrong with my original loss." }, "source": "meta" }