It has been proven that the DenseNet is performing much better than ResNet. Assuming our GPU is supportive, why there's nobody implement the model with DenseNet? If anyone has the same thoughts as me, I would like to discuss more details.
", "messages": [ { "id": "1006383", "postDate": "09/11/2020 08:09:33", "content": "It has been proven that the DenseNet is performing much better than ResNet. Assuming our GPU is supportive, why there's nobody implement the model with DenseNet? If anyone has the same thoughts as me, I would like to discuss more details.
", "rawMarkdown": "It has been proven that the DenseNet is performing much better than ResNet. Assuming our GPU is supportive, why there's nobody implement the model with DenseNet? If anyone has the same thoughts as me, I would like to discuss more details.", "votes": null }, { "id": "1009774", "postDate": "09/14/2020 08:18:41", "content": "That's not a straightforward competition for using CNN with pictures and target for every picture. We have CT scans only for the first measurement and in my opinion one way to use CNN is as a part of autoencoders to extend tabular data. An in this case I would try both DenseNet and ResNet and choose the one with the best result in the particular data. If recourses are restricted I would choose top models with the high metric values in papers and with appropriate computational complexity (chances are, DenseNet is in the top) + some classic models (definitely ResNet).
", "rawMarkdown": "That's not a straightforward competition for using CNN with pictures and target for every picture. We have CT scans only for the first measurement and in my opinion one way to use CNN is as a part of autoencoders to extend tabular data. An in this case I would try both DenseNet and ResNet and choose the one with the best result in the particular data. If recourses are restricted I would choose top models with the high metric values in papers and with appropriate computational complexity (chances are, DenseNet is in the top) + some classic models (definitely ResNet).", "votes": null }, { "id": "1009899", "postDate": "09/14/2020 10:56:33", "content": "Thank you for sharing! Actually I have tried DenseNet to deal with image data but given the limited GPU, the batch_size is restricted to 32 and the image_size is no more than 100*100. It turns out that it is better than ResNet but worse than EfficientNet. So in my opinion, I am afraid the endeavour of using two models would be a little bit difficult with regards to limited GPU. As for your advice, could you please elaborate more on how to choose between DenseNet and ResNet models?
", "rawMarkdown": "Thank you for sharing! Actually I have tried DenseNet to deal with image data but given the limited GPU, the batch_size is restricted to 32 and the image_size is no more than 100*100. It turns out that it is better than ResNet but worse than EfficientNet. So in my opinion, I am afraid the endeavour of using two models would be a little bit difficult with regards to limited GPU. As for your advice, could you please elaborate more on how to choose between DenseNet and ResNet models?", "votes": null } ], "comments": [ { "id": 1009774, "author_name": "koza4ukdmitrij", "author_url": "", "post_date": "09/14/2020 08:18:41", "content": "That's not a straightforward competition for using CNN with pictures and target for every picture. We have CT scans only for the first measurement and in my opinion one way to use CNN is as a part of autoencoders to extend tabular data. An in this case I would try both DenseNet and ResNet and choose the one with the best result in the particular data. If recourses are restricted I would choose top models with the high metric values in papers and with appropriate computational complexity (chances are, DenseNet is in the top) + some classic models (definitely ResNet).
", "votes": null, "replies": [ { "id": 1009899, "author_name": "joanhou", "author_url": "", "post_date": "09/14/2020 10:56:33", "content": "Thank you for sharing! Actually I have tried DenseNet to deal with image data but given the limited GPU, the batch_size is restricted to 32 and the image_size is no more than 100*100. It turns out that it is better than ResNet but worse than EfficientNet. So in my opinion, I am afraid the endeavour of using two models would be a little bit difficult with regards to limited GPU. As for your advice, could you please elaborate more on how to choose between DenseNet and ResNet models?
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1006383": "It has been proven that the DenseNet is performing much better than ResNet. Assuming our GPU is supportive, why there's nobody implement the model with DenseNet? If anyone has the same thoughts as me, I would like to discuss more details.", "1009774": "That's not a straightforward competition for using CNN with pictures and target for every picture. We have CT scans only for the first measurement and in my opinion one way to use CNN is as a part of autoencoders to extend tabular data. An in this case I would try both DenseNet and ResNet and choose the one with the best result in the particular data. If recourses are restricted I would choose top models with the high metric values in papers and with appropriate computational complexity (chances are, DenseNet is in the top) + some classic models (definitely ResNet).", "1009899": "Thank you for sharing! Actually I have tried DenseNet to deal with image data but given the limited GPU, the batch_size is restricted to 32 and the image_size is no more than 100*100. It turns out that it is better than ResNet but worse than EfficientNet. So in my opinion, I am afraid the endeavour of using two models would be a little bit difficult with regards to limited GPU. As for your advice, could you please elaborate more on how to choose between DenseNet and ResNet models?" }, "source": "meta" }