In previous similar competition https://www.kaggle.com/c/dogs-vs-cats-redux-kernels-edition I have trained several models https://github.com/mrgloom/kaggle-dogs-vs-cats-solution and by learning model from scratch I get competitive results comparing to finetuning imagenet pretrained model, seems this competition is harder, have anyone trained successful CNN model from scratch?
", "messages": [ { "id": "241011", "postDate": "11/07/2017 21:02:59", "content": "In previous similar competition https://www.kaggle.com/c/dogs-vs-cats-redux-kernels-edition I have trained several models https://github.com/mrgloom/kaggle-dogs-vs-cats-solution and by learning model from scratch I get competitive results comparing to finetuning imagenet pretrained model, seems this competition is harder, have anyone trained successful CNN model from scratch?
", "rawMarkdown": "In previous similar competition https://www.kaggle.com/c/dogs-vs-cats-redux-kernels-edition I have trained several models https://github.com/mrgloom/kaggle-dogs-vs-cats-solution and by learning model from scratch I get competitive results comparing to finetuning imagenet pretrained model, seems this competition is harder, have anyone trained successful CNN model from scratch?", "votes": null }, { "id": "246081", "postDate": "11/20/2017 14:29:52", "content": "The issue here is the memory : I am unable to use the entire set of images to train the CNN because of which I am getting jusdt 10% validation accuracy. Thus I feel, a pre-trained model is a must unless you have a high end machine/server. Bdw I have 256GB RAM with GPU and still ResourceAllocation Error.
", "rawMarkdown": "The issue here is the memory : I am unable to use the entire set of images to train the CNN because of which I am getting jusdt 10% validation accuracy. Thus I feel, a pre-trained model is a must unless you have a high end machine/server. Bdw I have 256GB RAM with GPU and still ResourceAllocation Error.", "votes": null }, { "id": "248398", "postDate": "11/25/2017 23:34:44", "content": "I began to try a little bit of this with resnet50 on p2 and p3 and it wasn't feasible for me. With that in mind I eventually concluded that imagenet weights were pretty good at this problem. So that's good for our hardware.
", "rawMarkdown": "I began to try a little bit of this with resnet50 on p2 and p3 and it wasn't feasible for me. With that in mind I eventually concluded that imagenet weights were pretty good at this problem. So that's good for our hardware.", "votes": null }, { "id": "255117", "postDate": "12/08/2017 10:55:01", "content": "I'm also interested in building a model from scratch, I will report If I manage to have a somewhat working solution.
", "rawMarkdown": "I'm also interested in building a model from scratch, I will report If I manage to have a somewhat working solution.", "votes": null }, { "id": "256114", "postDate": "12/11/2017 07:13:37", "content": "I tried to fit resnet50 to solve this problem, but it didn't converge, training batch size is set to be 32, lr = 0.01 first, and I tried to change it from 0.01, 0.001 to 0.00001, but none of them work. Can somebody tell me how to solve this problem?
", "rawMarkdown": "I tried to fit resnet50 to solve this problem, but it didn't converge, training batch size is set to be 32, lr = 0.01 first, and I tried to change it from 0.01, 0.001 to 0.00001, but none of them work. Can somebody tell me how to solve this problem?", "votes": null } ], "comments": [ { "id": 246081, "author_name": "vasaibaby", "author_url": "", "post_date": "11/20/2017 14:29:52", "content": "The issue here is the memory : I am unable to use the entire set of images to train the CNN because of which I am getting jusdt 10% validation accuracy. Thus I feel, a pre-trained model is a must unless you have a high end machine/server. Bdw I have 256GB RAM with GPU and still ResourceAllocation Error.
", "votes": null, "replies": [] }, { "id": 248398, "author_name": "jnblanchard", "author_url": "", "post_date": "11/25/2017 23:34:44", "content": "I began to try a little bit of this with resnet50 on p2 and p3 and it wasn't feasible for me. With that in mind I eventually concluded that imagenet weights were pretty good at this problem. So that's good for our hardware.
", "votes": null, "replies": [] }, { "id": 255117, "author_name": "jairsan", "author_url": "", "post_date": "12/08/2017 10:55:01", "content": "I'm also interested in building a model from scratch, I will report If I manage to have a somewhat working solution.
", "votes": null, "replies": [] }, { "id": 256114, "author_name": "xiaodanli003", "author_url": "", "post_date": "12/11/2017 07:13:37", "content": "I tried to fit resnet50 to solve this problem, but it didn't converge, training batch size is set to be 32, lr = 0.01 first, and I tried to change it from 0.01, 0.001 to 0.00001, but none of them work. Can somebody tell me how to solve this problem?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "241011": "In previous similar competition https://www.kaggle.com/c/dogs-vs-cats-redux-kernels-edition I have trained several models https://github.com/mrgloom/kaggle-dogs-vs-cats-solution and by learning model from scratch I get competitive results comparing to finetuning imagenet pretrained model, seems this competition is harder, have anyone trained successful CNN model from scratch?", "246081": "The issue here is the memory : I am unable to use the entire set of images to train the CNN because of which I am getting jusdt 10% validation accuracy. Thus I feel, a pre-trained model is a must unless you have a high end machine/server. Bdw I have 256GB RAM with GPU and still ResourceAllocation Error.", "248398": "I began to try a little bit of this with resnet50 on p2 and p3 and it wasn't feasible for me. With that in mind I eventually concluded that imagenet weights were pretty good at this problem. So that's good for our hardware.", "255117": "I'm also interested in building a model from scratch, I will report If I manage to have a somewhat working solution.", "256114": "I tried to fit resnet50 to solve this problem, but it didn't converge, training batch size is set to be 32, lr = 0.01 first, and I tried to change it from 0.01, 0.001 to 0.00001, but none of them work. Can somebody tell me how to solve this problem?" }, "source": "meta" }