As we have seen so far through some incredible public notebooks there are a lot of different ways to successfully tackle this competition so far. Some examples are EfficientNets, Resnets, WaveNets, and even more classical models like Catboost! This makes me particularly excited because of the ensembling possibilities. Just to give an idea of the impact of this I have shared a notebook utilizing the work of @cdeotte and @yunsuxiaozi. In these incredible works, they both achieve a high CV/LB score, but they do so with different models which of course have different strengths. In averaging their solutions we can greatly improve the LB score to 0.39. I have a number of other ideas from this that could be useful. As I am certain there are a few other ways this can be done well. I think this competition will be interesting because the winner will likely have many diverse models rather than just a really good 2 or 3 models. I think creativity may get a leg up here. Also interested to see how ensembling will be affected by the metric as the sum of each row must always add to 1. If you are interested or need help in blending models you can find my example here:
\nhttps://www.kaggle.com/code/cody11null/quick-ensemble
", "messages": [ { "id": 2618600, "postDate": "2024-01-24T20:56:49.893Z", "content": "As we have seen so far through some incredible public notebooks there are a lot of different ways to successfully tackle this competition so far. Some examples are EfficientNets, Resnets, WaveNets, and even more classical models like Catboost! This makes me particularly excited because of the ensembling possibilities. Just to give an idea of the impact of this I have shared a notebook utilizing the work of @cdeotte and @yunsuxiaozi. In these incredible works, they both achieve a high CV/LB score, but they do so with different models which of course have different strengths. In averaging their solutions we can greatly improve the LB score to 0.39. I have a number of other ideas from this that could be useful. As I am certain there are a few other ways this can be done well. I think this competition will be interesting because the winner will likely have many diverse models rather than just a really good 2 or 3 models. I think creativity may get a leg up here. Also interested to see how ensembling will be affected by the metric as the sum of each row must always add to 1. If you are interested or need help in blending models you can find my example here:
\n", "rawMarkdown": "As we have seen so far through some incredible public notebooks there are a lot of different ways to successfully tackle this competition so far. Some examples are EfficientNets, Resnets, WaveNets, and even more classical models like Catboost! This makes me particularly excited because of the ensembling possibilities. Just to give an idea of the impact of this I have shared a notebook utilizing the work of @cdeotte and @yunsuxiaozi. In these incredible works, they both achieve a high CV/LB score, but they do so with different models which of course have different strengths. In averaging their solutions we can greatly improve the LB score to 0.39. I have a number of other ideas from this that could be useful. As I am certain there are a few other ways this can be done well. I think this competition will be interesting because the winner will likely have many diverse models rather than just a really good 2 or 3 models. I think creativity may get a leg up here. Also interested to see how ensembling will be affected by the metric as the sum of each row must always add to 1. If you are interested or need help in blending models you can find my example here: \n\nhttps://www.kaggle.com/code/cody11null/quick-ensemble", "votes": 21 } ], "comments": [], "raw_markdown_by_id": { "2618600": "As we have seen so far through some incredible public notebooks there are a lot of different ways to successfully tackle this competition so far. Some examples are EfficientNets, Resnets, WaveNets, and even more classical models like Catboost! This makes me particularly excited because of the ensembling possibilities. Just to give an idea of the impact of this I have shared a notebook utilizing the work of @cdeotte and @yunsuxiaozi. In these incredible works, they both achieve a high CV/LB score, but they do so with different models which of course have different strengths. In averaging their solutions we can greatly improve the LB score to 0.39. I have a number of other ideas from this that could be useful. As I am certain there are a few other ways this can be done well. I think this competition will be interesting because the winner will likely have many diverse models rather than just a really good 2 or 3 models. I think creativity may get a leg up here. Also interested to see how ensembling will be affected by the metric as the sum of each row must always add to 1. If you are interested or need help in blending models you can find my example here: \n\nhttps://www.kaggle.com/code/cody11null/quick-ensemble" } }