I used a convolutional neural network in keras for the challenge, but I haven't really been able to get much out of it at all. I've previously used CNN's for similar tasks, but I did have a lot more data for those. My question is, given our data amount, what would be the model people would ideally use? I've been trying to tweak various parameters but it isn't really going anywhere so i got curious as to other approaches
", "messages": [ { "id": "564522", "postDate": "06/29/2019 14:56:21", "content": "I used a convolutional neural network in keras for the challenge, but I haven't really been able to get much out of it at all. I've previously used CNN's for similar tasks, but I did have a lot more data for those. My question is, given our data amount, what would be the model people would ideally use? I've been trying to tweak various parameters but it isn't really going anywhere so i got curious as to other approaches
", "rawMarkdown": "I used a convolutional neural network in keras for the challenge, but I haven't really been able to get much out of it at all. I've previously used CNN's for similar tasks, but I did have a lot more data for those. My question is, given our data amount, what would be the model people would ideally use? I've been trying to tweak various parameters but it isn't really going anywhere so i got curious as to other approaches", "votes": null }, { "id": "564885", "postDate": "06/30/2019 05:55:34", "content": "https://github.com/facebookresearch/SparseConvNet
", "rawMarkdown": "https://github.com/facebookresearch/SparseConvNet", "votes": null }, { "id": "566273", "postDate": "07/02/2019 01:48:53", "content": "It seems that because of the small number of data, it is not possible to train a model from zero, as it will not be able to generalize, I tried and got 0.006 in LB.\nBased on my little knowledge, I say that you should use robust pre-trained models (transfer learning) and use data augmentation techniques. Using these two techniques, I got a better score of 0.36. but I'm still noob :/
", "rawMarkdown": "It seems that because of the small number of data, it is not possible to train a model from zero, as it will not be able to generalize, I tried and got 0.006 in LB.\nBased on my little knowledge, I say that you should use robust pre-trained models (transfer learning) and use data augmentation techniques. Using these two techniques, I got a better score of 0.36. but I'm still noob :/", "votes": null }, { "id": "566274", "postDate": "07/02/2019 01:50:59", "content": "Thanks for the suggestions @mustaffxx .
", "rawMarkdown": "Thanks for the suggestions @mustaffxx .", "votes": null }, { "id": "566315", "postDate": "07/02/2019 03:46:31", "content": "Data augmentation is essential --and transfer learning is also a really good idea. But there is still a question about how the final layers of the pre-trained + un-trained hybrid should be constructed. So it is helpful to have some intuition based on familiarity with a variety of model architectures.
", "rawMarkdown": "Data augmentation is essential --and transfer learning is also a really good idea. But there is still a question about how the final layers of the pre-trained + un-trained hybrid should be constructed. So it is helpful to have some intuition based on familiarity with a variety of model architectures.", "votes": null } ], "comments": [ { "id": 564885, "author_name": "puremath86", "author_url": "", "post_date": "06/30/2019 05:55:34", "content": "https://github.com/facebookresearch/SparseConvNet
", "votes": null, "replies": [] }, { "id": 566273, "author_name": "mustaffxx", "author_url": "", "post_date": "07/02/2019 01:48:53", "content": "It seems that because of the small number of data, it is not possible to train a model from zero, as it will not be able to generalize, I tried and got 0.006 in LB.\nBased on my little knowledge, I say that you should use robust pre-trained models (transfer learning) and use data augmentation techniques. Using these two techniques, I got a better score of 0.36. but I'm still noob :/
", "votes": null, "replies": [ { "id": 566274, "author_name": "gunishj", "author_url": "", "post_date": "07/02/2019 01:50:59", "content": "Thanks for the suggestions @mustaffxx .
", "votes": null, "replies": [] }, { "id": 566315, "author_name": "puremath86", "author_url": "", "post_date": "07/02/2019 03:46:31", "content": "Data augmentation is essential --and transfer learning is also a really good idea. But there is still a question about how the final layers of the pre-trained + un-trained hybrid should be constructed. So it is helpful to have some intuition based on familiarity with a variety of model architectures.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "564522": "I used a convolutional neural network in keras for the challenge, but I haven't really been able to get much out of it at all. I've previously used CNN's for similar tasks, but I did have a lot more data for those. My question is, given our data amount, what would be the model people would ideally use? I've been trying to tweak various parameters but it isn't really going anywhere so i got curious as to other approaches", "564885": "https://github.com/facebookresearch/SparseConvNet", "566273": "It seems that because of the small number of data, it is not possible to train a model from zero, as it will not be able to generalize, I tried and got 0.006 in LB.\nBased on my little knowledge, I say that you should use robust pre-trained models (transfer learning) and use data augmentation techniques. Using these two techniques, I got a better score of 0.36. but I'm still noob :/", "566274": "Thanks for the suggestions @mustaffxx .", "566315": "Data augmentation is essential --and transfer learning is also a really good idea. But there is still a question about how the final layers of the pre-trained + un-trained hybrid should be constructed. So it is helpful to have some intuition based on familiarity with a variety of model architectures." }, "source": "meta" }