Hi guys,
\nCurious to know if anyone has managed to match the performance of the public notebooks using smaller images and/or smaller models? I see the >0.60 notebooks are using images >500x500 and some pretty large models.
\nSomeone give me some confidence that this is possible with my lowly 2080ti 😓
", "messages": [ { "id": "1697625", "postDate": "02/19/2022 18:12:33", "content": "Hi guys,
\nCurious to know if anyone has managed to match the performance of the public notebooks using smaller images and/or smaller models? I see the >0.60 notebooks are using images >500x500 and some pretty large models.
\nSomeone give me some confidence that this is possible with my lowly 2080ti 😓
", "rawMarkdown": "Hi guys,\n\nCurious to know if anyone has managed to match the performance of the public notebooks using smaller images and/or smaller models? I see the >0.60 notebooks are using images >500x500 and some pretty large models. \n\nSomeone give me some confidence that this is possible with my lowly 2080ti 😓", "votes": null }, { "id": "1697651", "postDate": "02/19/2022 18:39:20", "content": "The best LB score I've managed to get with 256x256 image and efficientnet_b0 backbone is 0.322 (trained for 20 epochs, ArcFaceLoss)
", "rawMarkdown": "The best LB score I've managed to get with 256x256 image and efficientnet_b0 backbone is 0.322 (trained for 20 epochs, ArcFaceLoss)", "votes": null }, { "id": "1697683", "postDate": "02/19/2022 19:17:38", "content": "I personally think that crucial think is to prepare data for this task. We have different kind of photos in dataset (problems and potential solution were described in different topics) . I assume that small images (even smaller then 256x256) are enough to score high. Second assumption after my quick research is that data preparation here will be 80% of work here 😄
", "rawMarkdown": "I personally think that crucial think is to prepare data for this task. We have different kind of photos in dataset (problems and potential solution were described in different topics) . I assume that small images (even smaller then 256x256) are enough to score high. Second assumption after my quick research is that data preparation here will be 80% of work here 😄", "votes": null }, { "id": "1697709", "postDate": "02/19/2022 19:46:00", "content": "Similar story, 0.38 with effnet b3 + arcface and 256x.
", "rawMarkdown": "Similar story, 0.38 with effnet b3 + arcface and 256x.", "votes": null }, { "id": "1697711", "postDate": "02/19/2022 19:47:20", "content": "Hope this is true and that the performance of models with larger images might be matched with better preprocessing!
", "rawMarkdown": "Hope this is true and that the performance of models with larger images might be matched with better preprocessing!", "votes": null }, { "id": "1697723", "postDate": "02/19/2022 19:51:56", "content": "Current score in my opinion is rather good but not excellent:) I suppose that in this completion we cross at least 0.85 … even 0.9 is possible.
", "rawMarkdown": "Current score in my opinion is rather good but not excellent:) I suppose that in this completion we cross at least 0.85 … even 0.9 is possible.", "votes": null }, { "id": "1697941", "postDate": "02/20/2022 01:40:32", "content": "Hi, I wonder a single model or k-folds ?
", "rawMarkdown": "Hi, I wonder a single model or k-folds ?", "votes": null }, { "id": "1698069", "postDate": "02/20/2022 05:50:57", "content": "cv 0.382
\nLB 0.355
\nSwin_224 + arcface
\n224 x 224 images
@mrinath How do you split the data, if you don't mind me asking?
", "rawMarkdown": "mrinath How do you split the data, if you don't mind me asking?", "votes": null }, { "id": "1698091", "postDate": "02/20/2022 06:11:55", "content": "I am using stratified kfold on individual id, I am considering those individual_ids as \"new individuals\" who are not in the training fold
", "rawMarkdown": "I am using stratified kfold on individual id, I am considering those individual_ids as \"new individuals\" who are not in the training fold", "votes": null }, { "id": "1698102", "postDate": "02/20/2022 06:22:16", "content": "If you're asking me, the score is with single model.
", "rawMarkdown": "If you're asking me, the score is with single model.", "votes": null }, { "id": "1698104", "postDate": "02/20/2022 06:24:28", "content": "Hi @mrinath, is your CV-LB correlated? Some of the experiments didn't correlate for me and there's a pretty significant gap between CV and LB scores for me (around 0.06 to 0.08)
", "rawMarkdown": "Hi @mrinath, is your CV-LB correlated? Some of the experiments didn't correlate for me and there's a pretty significant gap between CV and LB scores for me (around 0.06 to 0.08)", "votes": null }, { "id": "1698115", "postDate": "02/20/2022 06:44:44", "content": "till now it is correlating,
\nhmm about your gap, how are you validating?
StratifiedKfold on individual_id
swin_384+arcface
\nLB 459!
I think image size will matter here in this comp,I guess better for the models to distinguish
", "rawMarkdown": "I think image size will matter here in this comp,I guess better for the models to distinguish", "votes": null }, { "id": "1699368", "postDate": "02/21/2022 06:20:16", "content": "I am seeing swin models training faster than effnets, even with same config, and swin being a larger model.
\nIs anyone seeing the same thing? (using pytorch)
@mrinath Are you using models from timm?
", "rawMarkdown": "mrinath Are you using models from timm?", "votes": null }, { "id": "1699756", "postDate": "02/21/2022 12:12:25", "content": "yes, swin 224 and effnet_b4
", "rawMarkdown": "yes, swin 224 and effnet_b4", "votes": null } ], "comments": [ { "id": 1697651, "author_name": "atharvaingle", "author_url": "", "post_date": "02/19/2022 18:39:20", "content": "The best LB score I've managed to get with 256x256 image and efficientnet_b0 backbone is 0.322 (trained for 20 epochs, ArcFaceLoss)
", "votes": null, "replies": [ { "id": 1697709, "author_name": "taindow", "author_url": "", "post_date": "02/19/2022 19:46:00", "content": "Similar story, 0.38 with effnet b3 + arcface and 256x.
", "votes": null, "replies": [] }, { "id": 1697941, "author_name": "rainfalllove", "author_url": "", "post_date": "02/20/2022 01:40:32", "content": "Hi, I wonder a single model or k-folds ?
", "votes": null, "replies": [] }, { "id": 1698102, "author_name": "atharvaingle", "author_url": "", "post_date": "02/20/2022 06:22:16", "content": "If you're asking me, the score is with single model.
", "votes": null, "replies": [] } ] }, { "id": 1697683, "author_name": "remekkinas", "author_url": "", "post_date": "02/19/2022 19:17:38", "content": "I personally think that crucial think is to prepare data for this task. We have different kind of photos in dataset (problems and potential solution were described in different topics) . I assume that small images (even smaller then 256x256) are enough to score high. Second assumption after my quick research is that data preparation here will be 80% of work here 😄
", "votes": null, "replies": [ { "id": 1697711, "author_name": "taindow", "author_url": "", "post_date": "02/19/2022 19:47:20", "content": "Hope this is true and that the performance of models with larger images might be matched with better preprocessing!
", "votes": null, "replies": [] }, { "id": 1697723, "author_name": "remekkinas", "author_url": "", "post_date": "02/19/2022 19:51:56", "content": "Current score in my opinion is rather good but not excellent:) I suppose that in this completion we cross at least 0.85 … even 0.9 is possible.
", "votes": null, "replies": [] } ] }, { "id": 1698069, "author_name": "mrinath", "author_url": "", "post_date": "02/20/2022 05:50:57", "content": "cv 0.382
\nLB 0.355
\nSwin_224 + arcface
\n224 x 224 images
@mrinath How do you split the data, if you don't mind me asking?
", "votes": null, "replies": [] }, { "id": 1698091, "author_name": "mrinath", "author_url": "", "post_date": "02/20/2022 06:11:55", "content": "I am using stratified kfold on individual id, I am considering those individual_ids as \"new individuals\" who are not in the training fold
", "votes": null, "replies": [] }, { "id": 1698104, "author_name": "atharvaingle", "author_url": "", "post_date": "02/20/2022 06:24:28", "content": "Hi @mrinath, is your CV-LB correlated? Some of the experiments didn't correlate for me and there's a pretty significant gap between CV and LB scores for me (around 0.06 to 0.08)
", "votes": null, "replies": [] }, { "id": 1698115, "author_name": "mrinath", "author_url": "", "post_date": "02/20/2022 06:44:44", "content": "till now it is correlating,
\nhmm about your gap, how are you validating?
StratifiedKfold on individual_id
swin_384+arcface
\nLB 459!
I think image size will matter here in this comp,I guess better for the models to distinguish
", "votes": null, "replies": [] } ] }, { "id": 1699368, "author_name": "mrinath", "author_url": "", "post_date": "02/21/2022 06:20:16", "content": "I am seeing swin models training faster than effnets, even with same config, and swin being a larger model.
\nIs anyone seeing the same thing? (using pytorch)
@mrinath Are you using models from timm?
", "votes": null, "replies": [] }, { "id": 1699756, "author_name": "mrinath", "author_url": "", "post_date": "02/21/2022 12:12:25", "content": "yes, swin 224 and effnet_b4
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1697625": "Hi guys,\n\nCurious to know if anyone has managed to match the performance of the public notebooks using smaller images and/or smaller models? I see the >0.60 notebooks are using images >500x500 and some pretty large models. \n\nSomeone give me some confidence that this is possible with my lowly 2080ti 😓", "1697651": "The best LB score I've managed to get with 256x256 image and efficientnet_b0 backbone is 0.322 (trained for 20 epochs, ArcFaceLoss)", "1697683": "I personally think that crucial think is to prepare data for this task. We have different kind of photos in dataset (problems and potential solution were described in different topics) . I assume that small images (even smaller then 256x256) are enough to score high. Second assumption after my quick research is that data preparation here will be 80% of work here 😄", "1697709": "Similar story, 0.38 with effnet b3 + arcface and 256x.", "1697711": "Hope this is true and that the performance of models with larger images might be matched with better preprocessing!", "1697723": "Current score in my opinion is rather good but not excellent:) I suppose that in this completion we cross at least 0.85 … even 0.9 is possible.", "1697941": "Hi, I wonder a single model or k-folds ?", "1698069": "cv 0.382\nLB 0.355\nSwin_224 + arcface\n224 x 224 images", "1698088": "mrinath How do you split the data, if you don't mind me asking?", "1698091": "I am using stratified kfold on individual id, I am considering those individual_ids as \"new individuals\" who are not in the training fold", "1698102": "If you're asking me, the score is with single model.", "1698104": "Hi @mrinath, is your CV-LB correlated? Some of the experiments didn't correlate for me and there's a pretty significant gap between CV and LB scores for me (around 0.06 to 0.08)", "1698115": "till now it is correlating, \nhmm about your gap, how are you validating?", "1698161": "StratifiedKfold on `individual_id`", "1698499": "swin_384+arcface\nLB 459!", "1698615": "I think image size will matter here in this comp,I guess better for the models to distinguish", "1699368": "I am seeing swin models training faster than effnets, even with same config, and swin being a larger model.\nIs anyone seeing the same thing? (using pytorch)", "1699744": "mrinath Are you using models from timm?", "1699756": "yes, swin 224 and effnet_b4" }, "source": "meta" }