Hi, I am using 128 * 128 image size and got 0.933 LB. I noticed that turning to 256*256 is a way to improve. I am just wondering how can we transfer 128*128 parameters to 256*256 model(by changing the image size, the number of parameter has changed)? just want to learn this knowledge, Thx.
", "messages": [ { "id": 431959, "postDate": "2018-12-03T06:54:32.140Z", "content": "If you have done print(model.summary) function several times in previous, you probably notice that the structure of model (Let's say, resnet50) is independent with respect to your inputs.\nSo, you can use model.save_weights('resnet50_128_version1.h5') at the end of your model. \nThen change the input shape to 256*256 use ResNet50(input_shape = (256,256,1or3).......)\nThrid, use model.load_weights('resnet50_128_version1.h5')\nThen all things will be done.
", "rawMarkdown": "If you have done print(model.summary) function several times in previous, you probably notice that the structure of model (Let's say, resnet50) is independent with respect to your inputs.\nSo, you can use model.save_weights('resnet50_128_version1.h5') at the end of your model. \nThen change the input shape to 256*256 use ResNet50(input_shape = (256,256,1or3).......)\nThrid, use model.load_weights('resnet50_128_version1.h5')\nThen all things will be done.\n", "votes": 2, "replies": [ { "id": 431972, "postDate": "2018-12-03T07:55:32.137Z", "content": "Thx a lot!
", "rawMarkdown": "Thx a lot!" } ] }, { "id": 432131, "postDate": "2018-12-03T12:51:39.800Z", "content": "Hi MarcYueZhao, Have you finetuned on larger image size (256*256)? Just curious to know if that improves the model!
", "rawMarkdown": "Hi MarcYueZhao, Have you finetuned on larger image size (256*256)? Just curious to know if that improves the model!", "replies": [ { "id": 432581, "postDate": "2018-12-04T04:20:00.213Z", "content": "Hi Shivam, I didnt' do that because of time limitation. I just lower the lr and try to get higher score.
", "rawMarkdown": "Hi Shivam, I didnt' do that because of time limitation. I just lower the lr and try to get higher score." }, { "id": 432628, "postDate": "2018-12-04T05:42:43.907Z", "content": "Thanks for the reply.. you can check this discussion which is pretty good.. some people tried larger size to finetune the model. https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/72697
", "rawMarkdown": "Thanks for the reply.. you can check this discussion which is pretty good.. some people tried larger size to finetune the model. https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/72697" } ] }, { "id": 431915, "postDate": "2018-12-03T04:52:38.123Z", "content": "Hi, I am using 128 * 128 image size and got 0.933 LB. I noticed that turning to 256*256 is a way to improve. I am just wondering how can we transfer 128*128 parameters to 256*256 model(by changing the image size, the number of parameter has changed)? just want to learn this knowledge, Thx.
", "rawMarkdown": "Hi, I am using 128 * 128 image size and got 0.933 LB. I noticed that turning to 256*256 is a way to improve. I am just wondering how can we transfer 128*128 parameters to 256*256 model(by changing the image size, the number of parameter has changed)? just want to learn this knowledge, Thx." } ], "comments": [ { "id": 431959, "author_name": "Dilapsky Lee", "author_url": "", "post_date": "2018-12-03T06:54:32.140000", "content": "If you have done print(model.summary) function several times in previous, you probably notice that the structure of model (Let's say, resnet50) is independent with respect to your inputs.\nSo, you can use model.save_weights('resnet50_128_version1.h5') at the end of your model. \nThen change the input shape to 256*256 use ResNet50(input_shape = (256,256,1or3).......)\nThrid, use model.load_weights('resnet50_128_version1.h5')\nThen all things will be done.
", "votes": 2, "replies": [ { "id": 431972, "author_name": "MarcYueZhao", "author_url": "", "post_date": "2018-12-03T07:55:32.137000", "content": "Thx a lot!
", "votes": 0, "replies": [] } ] }, { "id": 432131, "author_name": "Shivam Gupta", "author_url": "", "post_date": "2018-12-03T12:51:39.800000", "content": "Hi MarcYueZhao, Have you finetuned on larger image size (256*256)? Just curious to know if that improves the model!
", "votes": 0, "replies": [ { "id": 432581, "author_name": "MarcYueZhao", "author_url": "", "post_date": "2018-12-04T04:20:00.213000", "content": "Hi Shivam, I didnt' do that because of time limitation. I just lower the lr and try to get higher score.
", "votes": 0, "replies": [] }, { "id": 432628, "author_name": "Shivam Gupta", "author_url": "", "post_date": "2018-12-04T05:42:43.907000", "content": "Thanks for the reply.. you can check this discussion which is pretty good.. some people tried larger size to finetune the model. https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/72697
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "431959": "If you have done print(model.summary) function several times in previous, you probably notice that the structure of model (Let's say, resnet50) is independent with respect to your inputs.\nSo, you can use model.save_weights('resnet50_128_version1.h5') at the end of your model. \nThen change the input shape to 256*256 use ResNet50(input_shape = (256,256,1or3).......)\nThrid, use model.load_weights('resnet50_128_version1.h5')\nThen all things will be done.\n", "432131": "Hi MarcYueZhao, Have you finetuned on larger image size (256*256)? Just curious to know if that improves the model!", "431915": "Hi, I am using 128 * 128 image size and got 0.933 LB. I noticed that turning to 256*256 is a way to improve. I am just wondering how can we transfer 128*128 parameters to 256*256 model(by changing the image size, the number of parameter has changed)? just want to learn this knowledge, Thx." } }