To verify my code pipeline, I have started by reproducing baseline. I trained a ResNet50 initialized with ImageNet and trained it for 30 epochs with lr0.1. The validation accuracy becomes close to 100% however the submission accuracy is close to 2%. Therefore, it would be great if you can provide more details about the baseline. It will help us see if our implementation pipeline is okay.
", "messages": [ { "id": "832473", "postDate": "05/04/2020 06:52:48", "content": "To verify my code pipeline, I have started by reproducing baseline. I trained a ResNet50 initialized with ImageNet and trained it for 30 epochs with lr0.1. The validation accuracy becomes close to 100% however the submission accuracy is close to 2%. Therefore, it would be great if you can provide more details about the baseline. It will help us see if our implementation pipeline is okay.
", "rawMarkdown": "To verify my code pipeline, I have started by reproducing baseline. I trained a ResNet50 initialized with ImageNet and trained it for 30 epochs with lr0.1. The validation accuracy becomes close to 100% however the submission accuracy is close to 2%. Therefore, it would be great if you can provide more details about the baseline. It will help us see if our implementation pipeline is okay.", "votes": null }, { "id": "832521", "postDate": "05/04/2020 07:54:27", "content": "By the way, the sample solution does not exist: https://github.com/cvl-umass/semi-inat-2020/blob/master/solution_sample.csv
", "rawMarkdown": "By the way, the sample solution does not exist: https://github.com/cvl-umass/semi-inat-2020/blob/master/solution_sample.csv", "votes": null }, { "id": "833057", "postDate": "05/04/2020 15:19:56", "content": "Thanks for noticing, I have fixed the link.
", "rawMarkdown": "Thanks for noticing, I have fixed the link.", "votes": null }, { "id": "833070", "postDate": "05/04/2020 15:24:56", "content": "Hi, I believe that using a pre-trained resnet50 you can at least get the baseline performance. I used a batch size of 32 with Adam optimizer trained with one single GPU. You can tune the learning rate, weight decay, etc.
", "rawMarkdown": "Hi, I believe that using a pre-trained resnet50 you can at least get the baseline performance. I used a batch size of 32 with Adam optimizer trained with one single GPU. You can tune the learning rate, weight decay, etc.", "votes": null }, { "id": "833838", "postDate": "05/05/2020 05:04:46", "content": "Thank you.
", "rawMarkdown": "Thank you.", "votes": null }, { "id": "834057", "postDate": "05/05/2020 09:05:46", "content": "Can you comment on difference between validation and test accuracy?
", "rawMarkdown": "Can you comment on difference between validation and test accuracy?", "votes": null }, { "id": "834483", "postDate": "05/05/2020 15:08:55", "content": "The val and test accuracy should be similar if you train on training set only.
", "rawMarkdown": "The val and test accuracy should be similar if you train on training set only.", "votes": null } ], "comments": [ { "id": 832521, "author_name": "deepawais", "author_url": "", "post_date": "05/04/2020 07:54:27", "content": "By the way, the sample solution does not exist: https://github.com/cvl-umass/semi-inat-2020/blob/master/solution_sample.csv
", "votes": null, "replies": [ { "id": 833057, "author_name": "jcfredsu", "author_url": "", "post_date": "05/04/2020 15:19:56", "content": "Thanks for noticing, I have fixed the link.
", "votes": null, "replies": [] } ] }, { "id": 833070, "author_name": "jcfredsu", "author_url": "", "post_date": "05/04/2020 15:24:56", "content": "Hi, I believe that using a pre-trained resnet50 you can at least get the baseline performance. I used a batch size of 32 with Adam optimizer trained with one single GPU. You can tune the learning rate, weight decay, etc.
", "votes": null, "replies": [ { "id": 833838, "author_name": "deepawais", "author_url": "", "post_date": "05/05/2020 05:04:46", "content": "Thank you.
", "votes": null, "replies": [] } ] }, { "id": 834057, "author_name": "deepawais", "author_url": "", "post_date": "05/05/2020 09:05:46", "content": "Can you comment on difference between validation and test accuracy?
", "votes": null, "replies": [ { "id": 834483, "author_name": "jcfredsu", "author_url": "", "post_date": "05/05/2020 15:08:55", "content": "The val and test accuracy should be similar if you train on training set only.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "832473": "To verify my code pipeline, I have started by reproducing baseline. I trained a ResNet50 initialized with ImageNet and trained it for 30 epochs with lr0.1. The validation accuracy becomes close to 100% however the submission accuracy is close to 2%. Therefore, it would be great if you can provide more details about the baseline. It will help us see if our implementation pipeline is okay.", "832521": "By the way, the sample solution does not exist: https://github.com/cvl-umass/semi-inat-2020/blob/master/solution_sample.csv", "833057": "Thanks for noticing, I have fixed the link.", "833070": "Hi, I believe that using a pre-trained resnet50 you can at least get the baseline performance. I used a batch size of 32 with Adam optimizer trained with one single GPU. You can tune the learning rate, weight decay, etc.", "833838": "Thank you.", "834057": "Can you comment on difference between validation and test accuracy?", "834483": "The val and test accuracy should be similar if you train on training set only." }, "source": "meta" }