i started by by tuning the parameters like transformations(added some new with xtra_tfms) and changed the size to 384 and batch size to 10 which gave better accuracy.\ni tried to balance the imbalanced data with SMOTE but it did not work out as disk space was getting overflowed.\ncallback was used to save the best results during learn
\n\n17 models were trained and were checked for accuracy on public leaderbooard. the accuracies were noted. all the submissions of the models can be found at-https://www.kaggle.com/sakshamaggarwal/submissions-csv with numbers corresponding to the models
\n\n4 mixnetxl 42487\n0 resnet152 41942\n7 resnext50d_32x4d 41664\n11 mnasnet_100 40909\n16 tf_mixnet_m 40542\n9 gluon_seresnext101_32x4d 40220\n5 enb2 40097\n13 gluon_resnet34_v1b 39908\n1 resnet50 39475\n6 resnet18 39297\n8 mobilenetv3 38675\n10 mixnet_s 38208\n15 semnasnet_100 38064\n3 resnet34 38030\n2 enb3 36474\n12 mixnet_l 34596\n14 mobilenetv1_100 26050
\n\ntop 5 models were then essembled to and submitted. the kernel regarding the same has been uploaded as kaggle notebook.\nthe colab link to the training of mixnetxl - \nhttps://colab.research.google.com/drive/1mNnEoaMOMm_fW4JBg6rtA6jyoVNKEyAa
", "comments": [], "comment_count": 0, "messages": [ { "id": 746049, "postDate": "2020-02-14T14:46:43.477Z", "content": "i started by by tuning the parameters like transformations(added some new with xtra_tfms) and changed the size to 384 and batch size to 10 which gave better accuracy.\ni tried to balance the imbalanced data with SMOTE but it did not work out as disk space was getting overflowed.\ncallback was used to save the best results during learn
\n\n17 models were trained and were checked for accuracy on public leaderbooard. the accuracies were noted. all the submissions of the models can be found at-https://www.kaggle.com/sakshamaggarwal/submissions-csv with numbers corresponding to the models
\n\n4 mixnetxl 42487\n0 resnet152 41942\n7 resnext50d_32x4d 41664\n11 mnasnet_100 40909\n16 tf_mixnet_m 40542\n9 gluon_seresnext101_32x4d 40220\n5 enb2 40097\n13 gluon_resnet34_v1b 39908\n1 resnet50 39475\n6 resnet18 39297\n8 mobilenetv3 38675\n10 mixnet_s 38208\n15 semnasnet_100 38064\n3 resnet34 38030\n2 enb3 36474\n12 mixnet_l 34596\n14 mobilenetv1_100 26050
\n\ntop 5 models were then essembled to and submitted. the kernel regarding the same has been uploaded as kaggle notebook.\nthe colab link to the training of mixnetxl - \nhttps://colab.research.google.com/drive/1mNnEoaMOMm_fW4JBg6rtA6jyoVNKEyAa
", "rawMarkdown": "i started by by tuning the parameters like transformations(added some new with xtra_tfms) and changed the size to 384 and batch size to 10 which gave better accuracy.\ni tried to balance the imbalanced data with SMOTE but it did not work out as disk space was getting overflowed.\ncallback was used to save the best results during learn\n\n17 models were trained and were checked for accuracy on public leaderbooard. the accuracies were noted. all the submissions of the models can be found at-https://www.kaggle.com/sakshamaggarwal/submissions-csv with numbers corresponding to the models\n\n4\tmixnetxl\t42487\n0\tresnet152\t41942\n7\tresnext50d_32x4d\t41664\n11\tmnasnet_100\t40909\n16\ttf_mixnet_m\t40542\n9\tgluon_seresnext101_32x4d\t40220\n5\tenb2\t40097\n13\tgluon_resnet34_v1b\t39908\n1\tresnet50\t39475\n6\tresnet18\t39297\n8\tmobilenetv3\t38675\n10\tmixnet_s\t38208\n15\tsemnasnet_100\t38064\n3\tresnet34\t38030\n2\tenb3\t36474\n12\tmixnet_l\t34596\n14\tmobilenetv1_100\t26050\n\n top 5 models were then essembled to and submitted. the kernel regarding the same has been uploaded as kaggle notebook.\nthe colab link to the training of mixnetxl - \nhttps://colab.research.google.com/drive/1mNnEoaMOMm_fW4JBg6rtA6jyoVNKEyAa" } ], "complete": true, "source": "live", "fetched_at": "2026-10-20T05:03:59Z", "competition_id": 18191, "competition_slug": "cl-cv2020", "team_id": 4296331, "solution_url": "/competitions/cl-cv2020/writeups/saksham-aggarwal-4th-place-solution" }