Have fun.
", "messages": [ { "id": "158117", "postDate": "01/25/2017 17:04:56", "content": "\n\nHave fun.
", "rawMarkdown": "**[Kernel][1]** and [Github][2]\r\n\r\nHave fun.\r\n\r\n [1]: https://www.kaggle.com/nothxplz/dogs-vs-cats-redux-kernels-edition/cats-vs-dogs-05-pytorch-example\r\n [2]: https://github.com/desimone/pytorch-cat-vs-dogs", "votes": null }, { "id": "158201", "postDate": "01/26/2017 03:59:53", "content": "it might be nice for you to introduce the torch package to the community and flush out what you are doing in your github (kernel is empty btw).
\n\nYes we can all go read up the docs, but having someone who knows it might be a big push to get people to adopt--it wasn't immediately clear in the code what was happening.
\n\nJust a suggestion
", "rawMarkdown": "it might be nice for you to introduce the torch package to the community and flush out what you are doing in your github (kernel is empty btw).\r\n\r\nYes we can all go read up the docs, but having someone who knows it might be a big push to get people to adopt--it wasn't immediately clear in the code what was happening.\r\n\r\nJust a suggestion", "votes": null }, { "id": "161562", "postDate": "02/14/2017 13:55:25", "content": "Hey bdd, thanks for the code -- great to see pytorch getting some use here already.
\n\nDid you test any of the other models that come included with torchvision or just res101? I was able to get a slight improvement from res152 and a few other tweaks (see here for implementation) so just curious if you had any luck with other model architectures.
", "rawMarkdown": "Hey bdd, thanks for the code -- great to see pytorch getting some use here already. \n\nDid you test any of the other models that come included with torchvision or just res101? I was able to get a slight improvement from res152 and a few other tweaks (see [here][1] for implementation) so just curious if you had any luck with other model architectures. \n\n\n [1]: https://github.com/rdcolema/pytorch-image-classification/blob/master/pytorch_model.ipynb", "votes": null }, { "id": "161601", "postDate": "02/14/2017 19:51:27", "content": "Glad it was useful.
\n\nI did not try other models or data augmentation. I wanted to show a friend how to do model fine-tuning in pytorch, and at the time there were no examples. I considered spending some time trying to get further up the leaderboard, but spot-checked some of data and found many images are mislabeled.
", "rawMarkdown": "Glad it was useful. \n\nI did not try other models or data augmentation. I wanted to show a friend how to do model fine-tuning in pytorch, and at the time there were no examples. I considered spending some time trying to get further up the leaderboard, but spot-checked some of data and found many images are mislabeled.", "votes": null }, { "id": "163727", "postDate": "02/25/2017 13:23:33", "content": "Hello, I'm trying to reproduce your results in Caffe, I have successfully trained ResNet-101 by finetuning and on 80%-20% train-test split I get accuracy 99.4%, howewer in terms of LogLoss I can't get such good score as yours(my best score is ~0.14).
\n\nAs I see here you use CrossEntropyLoss and I'm using SoftmaxWithLoss, seems they are the same, so maybe some other tricks were used?
https://github.com/mrgloom/kaggle-dogs-vs-cats-solution
", "rawMarkdown": "Hello, I'm trying to reproduce your results in Caffe, I have successfully trained ResNet-101 by finetuning and on 80%-20% train-test split I get accuracy 99.4%, howewer in terms of LogLoss I can't get such good score as yours(my best score is ~0.14).\n\nAs I see [here][1] you use `CrossEntropyLoss` and I'm using [SoftmaxWithLoss][2], seems they are the same, so maybe some other tricks were used?\n\nhttps://github.com/mrgloom/kaggle-dogs-vs-cats-solution\n\n\n [1]: https://github.com/desimone/pytorch-cat-vs-dogs/blob/master/main.py#L139\n [2]: http://caffe.berkeleyvision.org/doxygen/classcaffe_1_1SoftmaxWithLossLayer.html", "votes": null } ], "comments": [ { "id": 158201, "author_name": "rdslater", "author_url": "", "post_date": "01/26/2017 03:59:53", "content": "it might be nice for you to introduce the torch package to the community and flush out what you are doing in your github (kernel is empty btw).
\n\nYes we can all go read up the docs, but having someone who knows it might be a big push to get people to adopt--it wasn't immediately clear in the code what was happening.
\n\nJust a suggestion
", "votes": null, "replies": [] }, { "id": 161562, "author_name": "rdcolema", "author_url": "", "post_date": "02/14/2017 13:55:25", "content": "Hey bdd, thanks for the code -- great to see pytorch getting some use here already.
\n\nDid you test any of the other models that come included with torchvision or just res101? I was able to get a slight improvement from res152 and a few other tweaks (see here for implementation) so just curious if you had any luck with other model architectures.
", "votes": null, "replies": [ { "id": 161601, "author_name": "nothxplz", "author_url": "", "post_date": "02/14/2017 19:51:27", "content": "Glad it was useful.
\n\nI did not try other models or data augmentation. I wanted to show a friend how to do model fine-tuning in pytorch, and at the time there were no examples. I considered spending some time trying to get further up the leaderboard, but spot-checked some of data and found many images are mislabeled.
", "votes": null, "replies": [] } ] }, { "id": 163727, "author_name": "mrgloom", "author_url": "", "post_date": "02/25/2017 13:23:33", "content": "Hello, I'm trying to reproduce your results in Caffe, I have successfully trained ResNet-101 by finetuning and on 80%-20% train-test split I get accuracy 99.4%, howewer in terms of LogLoss I can't get such good score as yours(my best score is ~0.14).
\n\nAs I see here you use CrossEntropyLoss and I'm using SoftmaxWithLoss, seems they are the same, so maybe some other tricks were used?