I can see in this this discussion that some teams were using EfficientNet. Since one of the top 3 said he used this model in ods slack I tried the b3 one (pretrained) and it gave me 0.591 in a simple fold(5folds) which is low compared to the other models I tried.\nWhat were your scores with this new model?
", "messages": [ { "id": "545861", "postDate": "06/06/2019 01:14:05", "content": "I can see in this this discussion that some teams were using EfficientNet. Since one of the top 3 said he used this model in ods slack I tried the b3 one (pretrained) and it gave me 0.591 in a simple fold(5folds) which is low compared to the other models I tried.\nWhat were your scores with this new model?
", "rawMarkdown": "I can see in this [this discussion](https://www.kaggle.com/c/imet-2019-fgvc6/discussion/87074#latest-545856) that some teams were using EfficientNet. Since one of the top 3 said he used this model in ods slack I tried the b3 one (pretrained) and it gave me 0.591 in a simple fold(5folds) which is low compared to the other models I tried.\nWhat were your scores with this new model?", "votes": null }, { "id": "545864", "postDate": "06/06/2019 01:23:52", "content": "5fold cv: 0.595
", "rawMarkdown": "5fold cv: 0.595", "votes": null }, { "id": "545870", "postDate": "06/06/2019 01:40:06", "content": "What was the LB score?
", "rawMarkdown": "What was the LB score?", "votes": null }, { "id": "545888", "postDate": "06/06/2019 02:23:03", "content": "We did not submit this model
", "rawMarkdown": "We did not submit this model", "votes": null }, { "id": "546204", "postDate": "06/06/2019 11:02:48", "content": ":My opinion of EfficientNet (pytorch, e3): fast, but takes larger batchsize compared with other models that have similar weights size. By the way, the performance is pretty bad... Maybe we have to use NAS for each dataset?
", "rawMarkdown": ":My opinion of EfficientNet (pytorch, e3): fast, but takes larger batchsize compared with other models that have similar weights size. By the way, the performance is pretty bad... Maybe we have to use NAS for each dataset?", "votes": null }, { "id": "546382", "postDate": "06/06/2019 14:31:00", "content": "you're right the b3 converge fast in only some iterations (10-20 batch's iterations to break 0.02 BCE) compared to our serenext101(80-100).\nBut if we take the results mentioned in the paper in consideration, b3 should break 0.6 easily...\nWe're missing something with this new architecture
", "rawMarkdown": "you're right the b3 converge fast in only some iterations (10-20 batch's iterations to break 0.02 BCE) compared to our serenext101(80-100).\nBut if we take the results mentioned in the paper in consideration, b3 should break 0.6 easily...\nWe're missing something with this new architecture", "votes": null }, { "id": "546687", "postDate": "06/06/2019 20:09:13", "content": "B3 with pretrained weights is pretty bad for my poor pipeline.
", "rawMarkdown": "B3 with pretrained weights is pretty bad for my poor pipeline.", "votes": null }, { "id": "546721", "postDate": "06/06/2019 20:54:27", "content": "I cannot break 580CV. I think one thing is the NN with more parameters are generalized better, so effientnet with small number of parameters may not generalize well as other structure
", "rawMarkdown": "I cannot break 580CV. I think one thing is the NN with more parameters are generalized better, so effientnet with small number of parameters may not generalize well as other structure", "votes": null }, { "id": "546903", "postDate": "06/07/2019 02:18:58", "content": "You're right ! deep models perform better with this dataset. Maybe B7 would give better results but unfortunately there are no pretrained weights until now.
", "rawMarkdown": "You're right ! deep models perform better with this dataset. Maybe B7 would give better results but unfortunately there are no pretrained weights until now.", "votes": null } ], "comments": [ { "id": 545864, "author_name": "wuyhbb", "author_url": "", "post_date": "06/06/2019 01:23:52", "content": "5fold cv: 0.595
", "votes": null, "replies": [ { "id": 545870, "author_name": "rinnqd", "author_url": "", "post_date": "06/06/2019 01:40:06", "content": "What was the LB score?
", "votes": null, "replies": [] }, { "id": 545888, "author_name": "wuyhbb", "author_url": "", "post_date": "06/06/2019 02:23:03", "content": "We did not submit this model
", "votes": null, "replies": [] } ] }, { "id": 546204, "author_name": "yiheng", "author_url": "", "post_date": "06/06/2019 11:02:48", "content": ":My opinion of EfficientNet (pytorch, e3): fast, but takes larger batchsize compared with other models that have similar weights size. By the way, the performance is pretty bad... Maybe we have to use NAS for each dataset?
", "votes": null, "replies": [ { "id": 546382, "author_name": "rinnqd", "author_url": "", "post_date": "06/06/2019 14:31:00", "content": "you're right the b3 converge fast in only some iterations (10-20 batch's iterations to break 0.02 BCE) compared to our serenext101(80-100).\nBut if we take the results mentioned in the paper in consideration, b3 should break 0.6 easily...\nWe're missing something with this new architecture
", "votes": null, "replies": [] } ] }, { "id": 546687, "author_name": "jionie", "author_url": "", "post_date": "06/06/2019 20:09:13", "content": "B3 with pretrained weights is pretty bad for my poor pipeline.
", "votes": null, "replies": [] }, { "id": 546721, "author_name": "strideradu", "author_url": "", "post_date": "06/06/2019 20:54:27", "content": "I cannot break 580CV. I think one thing is the NN with more parameters are generalized better, so effientnet with small number of parameters may not generalize well as other structure
", "votes": null, "replies": [ { "id": 546903, "author_name": "rinnqd", "author_url": "", "post_date": "06/07/2019 02:18:58", "content": "You're right ! deep models perform better with this dataset. Maybe B7 would give better results but unfortunately there are no pretrained weights until now.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "545861": "I can see in this [this discussion](https://www.kaggle.com/c/imet-2019-fgvc6/discussion/87074#latest-545856) that some teams were using EfficientNet. Since one of the top 3 said he used this model in ods slack I tried the b3 one (pretrained) and it gave me 0.591 in a simple fold(5folds) which is low compared to the other models I tried.\nWhat were your scores with this new model?", "545864": "5fold cv: 0.595", "545870": "What was the LB score?", "545888": "We did not submit this model", "546204": ":My opinion of EfficientNet (pytorch, e3): fast, but takes larger batchsize compared with other models that have similar weights size. By the way, the performance is pretty bad... Maybe we have to use NAS for each dataset?", "546382": "you're right the b3 converge fast in only some iterations (10-20 batch's iterations to break 0.02 BCE) compared to our serenext101(80-100).\nBut if we take the results mentioned in the paper in consideration, b3 should break 0.6 easily...\nWe're missing something with this new architecture", "546687": "B3 with pretrained weights is pretty bad for my poor pipeline.", "546721": "I cannot break 580CV. I think one thing is the NN with more parameters are generalized better, so effientnet with small number of parameters may not generalize well as other structure", "546903": "You're right ! deep models perform better with this dataset. Maybe B7 would give better results but unfortunately there are no pretrained weights until now." }, "source": "meta" }