Hi, i am using triplet loss but its not getting converge and stops at 0.7 or around margin. has anyone come across this problem? how to overcome it. I have tried different lr but no use?
", "messages": [ { "id": "939055", "postDate": "07/22/2020 02:14:09", "content": "Hi, i am using triplet loss but its not getting converge and stops at 0.7 or around margin. has anyone come across this problem? how to overcome it. I have tried different lr but no use?
", "rawMarkdown": "Hi, i am using triplet loss but its not getting converge and stops at 0.7 or around margin. has anyone come across this problem? how to overcome it. I have tried different lr but no use?", "votes": null }, { "id": "939067", "postDate": "07/22/2020 02:30:07", "content": "Do you have more info like learning rate, batch size, your architecture (e.g. what kind of pooling layer)? To debug, I would say you can always start from something simple and then gradually add complexity to your network…….
", "rawMarkdown": "Do you have more info like learning rate, batch size, your architecture (e.g. what kind of pooling layer)? To debug, I would say you can always start from something simple and then gradually add complexity to your network.......", "votes": null }, { "id": "939075", "postDate": "07/22/2020 02:40:04", "content": "First, congratulations for going to be master with this competition. here are the details\nlr=1e-5, batch 30(10X3), its VGG16 Arch with changed top to get the descriptor. no other changes.
", "rawMarkdown": "First, congratulations for going to be master with this competition. here are the details\nlr=1e-5, batch 30(10X3), its VGG16 Arch with changed top to get the descriptor. no other changes.", "votes": null }, { "id": "939120", "postDate": "07/22/2020 03:42:23", "content": "It sounds alright. Perhaps you can create a small dataset like 1000 images and debug from there.
", "rawMarkdown": "It sounds alright. Perhaps you can create a small dataset like 1000 images and debug from there.", "votes": null } ], "comments": [ { "id": 939067, "author_name": "pukkinming", "author_url": "", "post_date": "07/22/2020 02:30:07", "content": "Do you have more info like learning rate, batch size, your architecture (e.g. what kind of pooling layer)? To debug, I would say you can always start from something simple and then gradually add complexity to your network…….
", "votes": null, "replies": [ { "id": 939075, "author_name": "udaygurugubelli", "author_url": "", "post_date": "07/22/2020 02:40:04", "content": "First, congratulations for going to be master with this competition. here are the details\nlr=1e-5, batch 30(10X3), its VGG16 Arch with changed top to get the descriptor. no other changes.
", "votes": null, "replies": [] }, { "id": 939120, "author_name": "pukkinming", "author_url": "", "post_date": "07/22/2020 03:42:23", "content": "It sounds alright. Perhaps you can create a small dataset like 1000 images and debug from there.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "939055": "Hi, i am using triplet loss but its not getting converge and stops at 0.7 or around margin. has anyone come across this problem? how to overcome it. I have tried different lr but no use?", "939067": "Do you have more info like learning rate, batch size, your architecture (e.g. what kind of pooling layer)? To debug, I would say you can always start from something simple and then gradually add complexity to your network.......", "939075": "First, congratulations for going to be master with this competition. here are the details\nlr=1e-5, batch 30(10X3), its VGG16 Arch with changed top to get the descriptor. no other changes.", "939120": "It sounds alright. Perhaps you can create a small dataset like 1000 images and debug from there." }, "source": "meta" }