![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F4256010%2Fb8426fcea87a9a916a275feb4b487015%2FUntitled.png?generation=1589292019787248&alt=media\")
Hi all, I'm wondering if anyone else has noticed their training curves have consistent patterns across variations in learning rate schedule, image size, and data augmentation.\nHere are just three traces, but I have many more with the same dip at 175 iterations and 375 iterations. The 4 individual losses from the training also have similar profiles across runs. Seems very weird to me! Anyone know why?
\n\n
Hi all, I'm wondering if anyone else has noticed their training curves have consistent patterns across variations in learning rate schedule, image size, and data augmentation.\nHere are just three traces, but I have many more with the same dip at 175 iterations and 375 iterations. The 4 individual losses from the training also have similar profiles across runs. Seems very weird to me! Anyone know why?
\n\n
The problem here as I understood from the paper and I might be wrong, that it's very easy to train (and overfit) for a location or multiple locations but then fail to get a good result if you test or try to generalize to other locations ... I believe data augmentation will be critical in this competition ...
", "rawMarkdown": "The problem here as I understood from the paper and I might be wrong, that it's very easy to train (and overfit) for a location or multiple locations but then fail to get a good result if you test or try to generalize to other locations ... I believe data augmentation will be critical in this competition ...", "votes": null }, { "id": "844781", "postDate": "05/12/2020 20:51:20", "content": "That seems very weird to me too. If I saw this, I would be suspicious that there is a bug somewhere in my code. Try to ensure that these changes you are making are actually being applied to the data/network. I'd be interested to hear the reason behind this
", "rawMarkdown": "That seems very weird to me too. If I saw this, I would be suspicious that there is a bug somewhere in my code. Try to ensure that these changes you are making are actually being applied to the data/network. I'd be interested to hear the reason behind this", "votes": null }, { "id": "847179", "postDate": "05/14/2020 08:07:29", "content": "After some more playing around I actually think it's got a lot to do with the data. In my shuffle, I use a seed for reproducibility so it's always in the same order. I suppose the dip has to do with a particularly \"easy\" batch. When I changed the seed the pattern changed.
", "rawMarkdown": "After some more playing around I actually think it's got a lot to do with the data. In my shuffle, I use a seed for reproducibility so it's always in the same order. I suppose the dip has to do with a particularly \"easy\" batch. When I changed the seed the pattern changed.", "votes": null } ], "comments": [ { "id": 844356, "author_name": "ma7moud", "author_url": "", "post_date": "05/12/2020 15:13:02", "content": "The problem here as I understood from the paper and I might be wrong, that it's very easy to train (and overfit) for a location or multiple locations but then fail to get a good result if you test or try to generalize to other locations ... I believe data augmentation will be critical in this competition ...
", "votes": null, "replies": [] }, { "id": 844781, "author_name": "matthewmasters", "author_url": "", "post_date": "05/12/2020 20:51:20", "content": "That seems very weird to me too. If I saw this, I would be suspicious that there is a bug somewhere in my code. Try to ensure that these changes you are making are actually being applied to the data/network. I'd be interested to hear the reason behind this
", "votes": null, "replies": [ { "id": 847179, "author_name": "alexandersoare", "author_url": "", "post_date": "05/14/2020 08:07:29", "content": "After some more playing around I actually think it's got a lot to do with the data. In my shuffle, I use a seed for reproducibility so it's always in the same order. I suppose the dip has to do with a particularly \"easy\" batch. When I changed the seed the pattern changed.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "844200": "Hi all, I'm wondering if anyone else has noticed their training curves have consistent patterns across variations in learning rate schedule, image size, and data augmentation.\nHere are just three traces, but I have many more with the same dip at 175 iterations and 375 iterations. The 4 individual losses from the training also have similar profiles across runs. Seems very weird to me! Anyone know why?\n\n", "844356": "The problem here as I understood from the paper and I might be wrong, that it's very easy to train (and overfit) for a location or multiple locations but then fail to get a good result if you test or try to generalize to other locations ... I believe data augmentation will be critical in this competition ...", "844781": "That seems very weird to me too. If I saw this, I would be suspicious that there is a bug somewhere in my code. Try to ensure that these changes you are making are actually being applied to the data/network. I'd be interested to hear the reason behind this", "847179": "After some more playing around I actually think it's got a lot to do with the data. In my shuffle, I use a seed for reproducibility so it's always in the same order. I suppose the dip has to do with a particularly \"easy\" batch. When I changed the seed the pattern changed." }, "source": "meta" }