Hi, may I ask how much LB do you boost after using external data? For me:\n1. ResNet18 is about 0.04 (0.45->0.49 single fold)\n2. ResNet50 is about 0.05 (0.48->0.53 single fold)
", "messages": [ { "id": "450878", "postDate": "01/06/2019 01:49:24", "content": "Hi, may I ask how much LB do you boost after using external data? For me:\n1. ResNet18 is about 0.04 (0.45->0.49 single fold)\n2. ResNet50 is about 0.05 (0.48->0.53 single fold)
", "rawMarkdown": "Hi, may I ask how much LB do you boost after using external data? For me:\n1. ResNet18 is about 0.04 (0.45->0.49 single fold)\n2. ResNet50 is about 0.05 (0.48->0.53 single fold)", "votes": null }, { "id": "450889", "postDate": "01/06/2019 02:41:19", "content": "For me, using the external data worsens the result.\nI do have a feeling that I might be doing something wrong though.
", "rawMarkdown": "For me, using the external data worsens the result.\nI do have a feeling that I might be doing something wrong though.", "votes": null }, { "id": "450907", "postDate": "01/06/2019 03:55:06", "content": "I got ~0.05-0.1 gain for models I tested.
", "rawMarkdown": "I got ~0.05-0.1 gain for models I tested.", "votes": null }, { "id": "450911", "postDate": "01/06/2019 04:28:29", "content": "you can try https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/75691#446302
", "rawMarkdown": "you can try https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/75691#446302", "votes": null }, { "id": "450913", "postDate": "01/06/2019 04:30:48", "content": "emm, 0.1 is great. \nMay I ask what is the model and loss function for 0.1? \nCurrently, resnet50+focal loss is best for me.
", "rawMarkdown": "emm, 0.1 is great. \nMay I ask what is the model and loss function for 0.1? \nCurrently, resnet50+focal loss is best for me.", "votes": null }, { "id": "450990", "postDate": "01/06/2019 08:17:46", "content": "It depends on whether or not you exclude the leak from the data.
", "rawMarkdown": "It depends on whether or not you exclude the leak from the data.", "votes": null }, { "id": "451607", "postDate": "01/07/2019 10:47:29", "content": "For me, it's about 0.1. But I don't think I am using the external data in an exactly right way. And my model is not very solid so it's looks like a huge approve. BTW I'm using Se-ResNet50
", "rawMarkdown": "For me, it's about 0.1. But I don't think I am using the external data in an exactly right way. And my model is not very solid so it's looks like a huge approve. BTW I'm using Se-ResNet50", "votes": null }, { "id": "451762", "postDate": "01/07/2019 16:17:46", "content": "ResNeXt50 + focal loss. ResNeXt50 requires about the same resources as ResNet50 but has better performance.
", "rawMarkdown": "ResNeXt50 + focal loss. ResNeXt50 requires about the same resources as ResNet50 but has better performance.", "votes": null }, { "id": "451973", "postDate": "01/08/2019 01:46:52", "content": "Have you tried bce loss? It seems that most people use bce loss instead of focal loss. In my case, bce loss always get worse results than focal loss.
", "rawMarkdown": "Have you tried bce loss? It seems that most people use bce loss instead of focal loss. In my case, bce loss always get worse results than focal loss.", "votes": null }, { "id": "451974", "postDate": "01/08/2019 01:48:27", "content": "Hi, I guess I am also not using external data in the right way. BTW what is the loss function that you are using?
", "rawMarkdown": "Hi, I guess I am also not using external data in the right way. BTW what is the loss function that you are using?", "votes": null }, { "id": "451981", "postDate": "01/08/2019 02:04:19", "content": "When I tested it, I got slightly worse results, but it was quite close. I also tested F1 loss, but it didn't work well on small batches (large images).
", "rawMarkdown": "When I tested it, I got slightly worse results, but it was quite close. I also tested F1 loss, but it didn't work well on small batches (large images).", "votes": null } ], "comments": [ { "id": 450889, "author_name": "ayerajath", "author_url": "", "post_date": "01/06/2019 02:41:19", "content": "For me, using the external data worsens the result.\nI do have a feeling that I might be doing something wrong though.
", "votes": null, "replies": [ { "id": 450911, "author_name": "zjucor", "author_url": "", "post_date": "01/06/2019 04:28:29", "content": "you can try https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/75691#446302
", "votes": null, "replies": [] } ] }, { "id": 450907, "author_name": "iafoss", "author_url": "", "post_date": "01/06/2019 03:55:06", "content": "I got ~0.05-0.1 gain for models I tested.
", "votes": null, "replies": [ { "id": 450913, "author_name": "zjucor", "author_url": "", "post_date": "01/06/2019 04:30:48", "content": "emm, 0.1 is great. \nMay I ask what is the model and loss function for 0.1? \nCurrently, resnet50+focal loss is best for me.
", "votes": null, "replies": [] }, { "id": 451762, "author_name": "iafoss", "author_url": "", "post_date": "01/07/2019 16:17:46", "content": "ResNeXt50 + focal loss. ResNeXt50 requires about the same resources as ResNet50 but has better performance.
", "votes": null, "replies": [] }, { "id": 451973, "author_name": "zjucor", "author_url": "", "post_date": "01/08/2019 01:46:52", "content": "Have you tried bce loss? It seems that most people use bce loss instead of focal loss. In my case, bce loss always get worse results than focal loss.
", "votes": null, "replies": [] }, { "id": 451981, "author_name": "iafoss", "author_url": "", "post_date": "01/08/2019 02:04:19", "content": "When I tested it, I got slightly worse results, but it was quite close. I also tested F1 loss, but it didn't work well on small batches (large images).
", "votes": null, "replies": [] } ] }, { "id": 450990, "author_name": "petewills", "author_url": "", "post_date": "01/06/2019 08:17:46", "content": "It depends on whether or not you exclude the leak from the data.
", "votes": null, "replies": [] }, { "id": 451607, "author_name": "snaker", "author_url": "", "post_date": "01/07/2019 10:47:29", "content": "For me, it's about 0.1. But I don't think I am using the external data in an exactly right way. And my model is not very solid so it's looks like a huge approve. BTW I'm using Se-ResNet50
", "votes": null, "replies": [ { "id": 451974, "author_name": "zjucor", "author_url": "", "post_date": "01/08/2019 01:48:27", "content": "Hi, I guess I am also not using external data in the right way. BTW what is the loss function that you are using?
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "450878": "Hi, may I ask how much LB do you boost after using external data? For me:\n1. ResNet18 is about 0.04 (0.45->0.49 single fold)\n2. ResNet50 is about 0.05 (0.48->0.53 single fold)", "450889": "For me, using the external data worsens the result.\nI do have a feeling that I might be doing something wrong though.", "450907": "I got ~0.05-0.1 gain for models I tested.", "450911": "you can try https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/75691#446302", "450913": "emm, 0.1 is great. \nMay I ask what is the model and loss function for 0.1? \nCurrently, resnet50+focal loss is best for me.", "450990": "It depends on whether or not you exclude the leak from the data.", "451607": "For me, it's about 0.1. But I don't think I am using the external data in an exactly right way. And my model is not very solid so it's looks like a huge approve. BTW I'm using Se-ResNet50", "451762": "ResNeXt50 + focal loss. ResNeXt50 requires about the same resources as ResNet50 but has better performance.", "451973": "Have you tried bce loss? It seems that most people use bce loss instead of focal loss. In my case, bce loss always get worse results than focal loss.", "451974": "Hi, I guess I am also not using external data in the right way. BTW what is the loss function that you are using?", "451981": "When I tested it, I got slightly worse results, but it was quite close. I also tested F1 loss, but it didn't work well on small batches (large images)." }, "source": "meta" }