It's awesome that we can train models on TPUs directly on Kaggle! And Martin gave us an excellent starter notebook Time to use all 30 hrs per week and see how high we can score.
\n\nWhat is you best score and/or recommendations for Single Model architecture?
\n\nI've added a notebook that I hope helps with making basic tweaks to simple models.
", "messages": [ { "id": "742680", "postDate": "02/11/2020 13:31:28", "content": "It's awesome that we can train models on TPUs directly on Kaggle! And Martin gave us an excellent starter notebook Time to use all 30 hrs per week and see how high we can score.
\n\nWhat is you best score and/or recommendations for Single Model architecture?
\n\nI've added a notebook that I hope helps with making basic tweaks to simple models.
", "rawMarkdown": "It's awesome that we can train models on TPUs directly on Kaggle! And Martin gave us an [excellent starter notebook](https://www.kaggle.com/mgornergoogle/getting-started-with-100-flowers-on-tpu) Time to use all 30 hrs per week and see how high we can score.\n\nWhat is you best score and/or recommendations for Single Model architecture?\n\nI've added a [notebook ](https://www.kaggle.com/rakibilly/flowers-for-beginners-like-me-on-tpu)that I hope helps with making basic tweaks to simple models.", "votes": null }, { "id": "742690", "postDate": "02/11/2020 13:42:01", "content": "Resnet 50: LB 0.925
", "rawMarkdown": "Resnet 50: LB 0.925", "votes": null }, { "id": "743056", "postDate": "02/11/2020 18:45:02", "content": "Try fine-tuning Xception of EfficientNet like in the Five flowers with Keras and Xception on TPU notebook.
", "rawMarkdown": "Try fine-tuning Xception of EfficientNet like in the [Five flowers with Keras and Xception on TPU](https://www.kaggle.com/mgornergoogle/five-flowers-with-keras-and-xception-on-tpu) notebook.", "votes": null }, { "id": "743187", "postDate": "02/11/2020 21:46:53", "content": "Thanks for the tip! I have already tried some other things... I just ran out of submissions today.
\n\nThanks for the great starter notebook!
", "rawMarkdown": "Thanks for the tip! I have already tried some other things... I just ran out of submissions today.\n\nThanks for the great starter notebook!", "votes": null }, { "id": "743419", "postDate": "02/12/2020 03:37:40", "content": "EfficientNetB0: 0.93497
", "rawMarkdown": "EfficientNetB0: 0.93497", "votes": null }, { "id": "766893", "postDate": "03/08/2020 22:56:30", "content": "An interesting \"side note\": I wanted to try NASNetLarge, which of all the Keras application models is the one with the best stats on ImageNet. However, the Keras implementation (possibly due to a bug) only works on images with the default size 331 x 331, so I decided to run some of the top models in that resolution in order to see if it was worth implementing a version of NASNet that works with 512 x 512 images. NASNetLarge delivered an accuracy of 0.91517, which I thought was due to the resolution and assumed the other models might show a marked reduction from their 512 x 512 performance, but then EfficientNet came in at 0.94559, and DenseNET201 at 0.95143, even in 331 x 331, which made me lose interest in NASNet for now.
", "rawMarkdown": "An interesting \"side note\": I wanted to try NASNetLarge, which of all the Keras application models is the one with the best stats on ImageNet. However, the Keras implementation (possibly due to a bug) only works on images with the default size 331 x 331, so I decided to run some of the top models in that resolution in order to see if it was worth implementing a version of NASNet that works with 512 x 512 images. NASNetLarge delivered an accuracy of 0.91517, which I thought was due to the resolution and assumed the other models might show a marked reduction from their 512 x 512 performance, but then EfficientNet came in at 0.94559, and DenseNET201 at 0.95143, even in 331 x 331, which made me lose interest in NASNet for now.", "votes": null }, { "id": "768386", "postDate": "03/10/2020 18:22:06", "content": "@atamazian made NASNETLarge work on TPUs a while back: https://www.kaggle.com/atamazian/100-flowers-on-tpu-with-nasnetlarge
", "rawMarkdown": "atamazian made NASNETLarge work on TPUs a while back: https://www.kaggle.com/atamazian/100-flowers-on-tpu-with-nasnetlarge", "votes": null }, { "id": "768864", "postDate": "03/11/2020 09:22:54", "content": "Thanks for the tip, Martin! Will have a look at it!
", "rawMarkdown": "Thanks for the tip, Martin! Will have a look at it!", "votes": null } ], "comments": [ { "id": 742690, "author_name": "rakibilly", "author_url": "", "post_date": "02/11/2020 13:42:01", "content": "Resnet 50: LB 0.925
", "votes": null, "replies": [] }, { "id": 743056, "author_name": "mgorner", "author_url": "", "post_date": "02/11/2020 18:45:02", "content": "Try fine-tuning Xception of EfficientNet like in the Five flowers with Keras and Xception on TPU notebook.
", "votes": null, "replies": [ { "id": 743187, "author_name": "rakibilly", "author_url": "", "post_date": "02/11/2020 21:46:53", "content": "Thanks for the tip! I have already tried some other things... I just ran out of submissions today.
\n\nThanks for the great starter notebook!
", "votes": null, "replies": [] } ] }, { "id": 743419, "author_name": "rakibilly", "author_url": "", "post_date": "02/12/2020 03:37:40", "content": "EfficientNetB0: 0.93497
", "votes": null, "replies": [] }, { "id": 766893, "author_name": "thomasbrekkunnvik", "author_url": "", "post_date": "03/08/2020 22:56:30", "content": "An interesting \"side note\": I wanted to try NASNetLarge, which of all the Keras application models is the one with the best stats on ImageNet. However, the Keras implementation (possibly due to a bug) only works on images with the default size 331 x 331, so I decided to run some of the top models in that resolution in order to see if it was worth implementing a version of NASNet that works with 512 x 512 images. NASNetLarge delivered an accuracy of 0.91517, which I thought was due to the resolution and assumed the other models might show a marked reduction from their 512 x 512 performance, but then EfficientNet came in at 0.94559, and DenseNET201 at 0.95143, even in 331 x 331, which made me lose interest in NASNet for now.
", "votes": null, "replies": [ { "id": 768386, "author_name": "mgorner", "author_url": "", "post_date": "03/10/2020 18:22:06", "content": "@atamazian made NASNETLarge work on TPUs a while back: https://www.kaggle.com/atamazian/100-flowers-on-tpu-with-nasnetlarge
", "votes": null, "replies": [] }, { "id": 768864, "author_name": "thomasbrekkunnvik", "author_url": "", "post_date": "03/11/2020 09:22:54", "content": "Thanks for the tip, Martin! Will have a look at it!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "742680": "It's awesome that we can train models on TPUs directly on Kaggle! And Martin gave us an [excellent starter notebook](https://www.kaggle.com/mgornergoogle/getting-started-with-100-flowers-on-tpu) Time to use all 30 hrs per week and see how high we can score.\n\nWhat is you best score and/or recommendations for Single Model architecture?\n\nI've added a [notebook ](https://www.kaggle.com/rakibilly/flowers-for-beginners-like-me-on-tpu)that I hope helps with making basic tweaks to simple models.", "742690": "Resnet 50: LB 0.925", "743056": "Try fine-tuning Xception of EfficientNet like in the [Five flowers with Keras and Xception on TPU](https://www.kaggle.com/mgornergoogle/five-flowers-with-keras-and-xception-on-tpu) notebook.", "743187": "Thanks for the tip! I have already tried some other things... I just ran out of submissions today.\n\nThanks for the great starter notebook!", "743419": "EfficientNetB0: 0.93497", "766893": "An interesting \"side note\": I wanted to try NASNetLarge, which of all the Keras application models is the one with the best stats on ImageNet. However, the Keras implementation (possibly due to a bug) only works on images with the default size 331 x 331, so I decided to run some of the top models in that resolution in order to see if it was worth implementing a version of NASNet that works with 512 x 512 images. NASNetLarge delivered an accuracy of 0.91517, which I thought was due to the resolution and assumed the other models might show a marked reduction from their 512 x 512 performance, but then EfficientNet came in at 0.94559, and DenseNET201 at 0.95143, even in 331 x 331, which made me lose interest in NASNet for now.", "768386": "atamazian made NASNETLarge work on TPUs a while back: https://www.kaggle.com/atamazian/100-flowers-on-tpu-with-nasnetlarge", "768864": "Thanks for the tip, Martin! Will have a look at it!" }, "source": "meta" }