\n
![\"Data](\"https://lionbridge.ai/wp-content/uploads/2020/03/2020-03-26_data-augmentation.jpg\")
Hi, All.
\nIs there any fine-tuning classification networks for bigger than 224x224 size as input image ?
\nI wonder …
\nThanks, in advance.
\nBest,
\n@bemoregt.
Hi, All.
\nIs there any fine-tuning classification networks for bigger than 224x224 size as input image ?
\nI wonder …
\nThanks, in advance.
\nBest,
\n@bemoregt.
You mean pretrained networks that you can finetune and that easily work with input images sizes bigger than 224 by 224? E.g. the various EfficientNet networks have as their default inputs (i.e. what they are trained on and when you don't want to just plug them in) 224 by 224 for EfficientNet-B0 going all the way up to EfficientNet-B8 with 672 by 672. Each EfficientNet has slightly larger input resolution, more depth etc. (all scaling in a ratio that the authors of the paper suggest may be a good approximation to jointly scaling them in an optimal trade-off between computation and performance).
", "rawMarkdown": "You mean pretrained networks that you can finetune and that easily work with input images sizes bigger than 224 by 224? E.g. the various EfficientNet networks have as their default inputs (i.e. what they are trained on and when you don't want to just plug them in) 224 by 224 for EfficientNet-B0 going all the way up to EfficientNet-B8 with 672 by 672. Each EfficientNet has slightly larger input resolution, more depth etc. (all scaling in a ratio that the authors of the paper suggest may be a good approximation to jointly scaling them in an optimal trade-off between computation and performance).", "votes": null }, { "id": "1122250", "postDate": "12/22/2020 09:53:28", "content": "HI, Bjo:rn.
\nThat's great.
\nThe EfficientNet-B8 with 672 by 672 for input image size!
\nThanks a lot.
\n", "rawMarkdown": "HI, Bjo:rn.\n\nThat's great.\n\nThe EfficientNet-B8 with 672 by 672 for input image size! \n\nThanks a lot.\n\n@bemoregt.", "votes": null }, { "id": "1122436", "postDate": "12/22/2020 12:46:57", "content": "Hey dude0!
\nI believe you don't know yet but the answer is in: 🤓😎😅
\nData Augmentation Methodology
why take a look and see if it helps?
\nData Augmentation Techniques for Larger Image Datasets:
\n
Image Manipulations
\nImage flipping, cropping, rotations, and translations are some obvious first steps. We can also change the color space of the image using contrast, sharpening, white balancing, color jittering, random color manipulation and many other techniques (called photometric transformations). If you’ve ever used Instagram filters or Snapseed, then you’ll get what I’m saying. You can get as creative as you want.
Moreover, you can mix images together, randomly erase segments of an image, and of course, combine all the above in all sorts of various ways.
\n😜🙄
![\"\"](\"https://lionbridge.ai/wp-content/uploads/2020/03/data-aug-01.png\")
Just be aware that trading the larger EfficientNet versions gets challenging. B4 and B5 is still straightforward enough, but especially B7 and B8 becomes challenging on the standard GPU Kaggle provides (lots of notebooks using the architecture in all its variants out there though).
", "rawMarkdown": "Just be aware that trading the larger EfficientNet versions gets challenging. B4 and B5 is still straightforward enough, but especially B7 and B8 becomes challenging on the standard GPU Kaggle provides (lots of notebooks using the architecture in all its variants out there though).", "votes": null }, { "id": "1122999", "postDate": "12/22/2020 20:58:39", "content": "Hi,
\nIt's not matter of augmentation,
\nbut matter of size of input image for pre-trained network.
Thanks, at any rate.
\n", "rawMarkdown": "Hi,\n\nIt's not matter of augmentation,\nbut matter of size of input image for pre-trained network. \n\nThanks, at any rate.\n\n@bemoregt.", "votes": null } ], "comments": [ { "id": 1122171, "author_name": "bjoernholzhauer", "author_url": "", "post_date": "12/22/2020 08:46:21", "content": "You mean pretrained networks that you can finetune and that easily work with input images sizes bigger than 224 by 224? E.g. the various EfficientNet networks have as their default inputs (i.e. what they are trained on and when you don't want to just plug them in) 224 by 224 for EfficientNet-B0 going all the way up to EfficientNet-B8 with 672 by 672. Each EfficientNet has slightly larger input resolution, more depth etc. (all scaling in a ratio that the authors of the paper suggest may be a good approximation to jointly scaling them in an optimal trade-off between computation and performance).
", "votes": null, "replies": [ { "id": 1122250, "author_name": "bemorekgg", "author_url": "", "post_date": "12/22/2020 09:53:28", "content": "HI, Bjo:rn.
\nThat's great.
\nThe EfficientNet-B8 with 672 by 672 for input image size!
\nThanks a lot.
\n", "votes": null, "replies": [] }, { "id": 1122556, "author_name": "bjoernholzhauer", "author_url": "", "post_date": "12/22/2020 14:22:45", "content": "Just be aware that trading the larger EfficientNet versions gets challenging. B4 and B5 is still straightforward enough, but especially B7 and B8 becomes challenging on the standard GPU Kaggle provides (lots of notebooks using the architecture in all its variants out there though).
", "votes": null, "replies": [] } ] }, { "id": 1122436, "author_name": "franckepeixoto", "author_url": "", "post_date": "12/22/2020 12:46:57", "content": "Hey dude0!
\nI believe you don't know yet but the answer is in: 🤓😎😅
\nData Augmentation Methodology
why take a look and see if it helps?
\nData Augmentation Techniques for Larger Image Datasets:
\n
Image Manipulations
\nImage flipping, cropping, rotations, and translations are some obvious first steps. We can also change the color space of the image using contrast, sharpening, white balancing, color jittering, random color manipulation and many other techniques (called photometric transformations). If you’ve ever used Instagram filters or Snapseed, then you’ll get what I’m saying. You can get as creative as you want.
Moreover, you can mix images together, randomly erase segments of an image, and of course, combine all the above in all sorts of various ways.
\n😜🙄
Hi,
\nIt's not matter of augmentation,
\nbut matter of size of input image for pre-trained network.
Thanks, at any rate.
\n", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1122108": "Hi, All.\n\nIs there any fine-tuning classification networks for bigger than 224x224 size as input image ? \n\nI wonder ...\n\nThanks, in advance.\nBest,\n@bemoregt.", "1122171": "You mean pretrained networks that you can finetune and that easily work with input images sizes bigger than 224 by 224? E.g. the various EfficientNet networks have as their default inputs (i.e. what they are trained on and when you don't want to just plug them in) 224 by 224 for EfficientNet-B0 going all the way up to EfficientNet-B8 with 672 by 672. Each EfficientNet has slightly larger input resolution, more depth etc. (all scaling in a ratio that the authors of the paper suggest may be a good approximation to jointly scaling them in an optimal trade-off between computation and performance).", "1122250": "HI, Bjo:rn.\n\nThat's great.\n\nThe EfficientNet-B8 with 672 by 672 for input image size! \n\nThanks a lot.\n\n@bemoregt.", "1122436": "Hey dude0!\nI believe you don't know yet but the answer is in: 🤓😎😅\nData Augmentation Methodology\n\nwhy take a look and see if it helps?\n\n**Data Augmentation Techniques for Larger Image Datasets:**\n\n\n[Image Manipulations](https://lionbridge.ai/articles/data-augmentation-with-machine-learning-an-overview/)\nImage flipping, cropping, rotations, and translations are some obvious first steps. We can also change the color space of the image using contrast, sharpening, white balancing, color jittering, random color manipulation and many other techniques (called photometric transformations). If you’ve ever used Instagram filters or Snapseed, then you’ll get what I’m saying. You can get as creative as you want.\n\nMoreover, you can mix images together, randomly erase segments of an image, and of course, combine all the above in all sorts of various ways.\n😜🙄\n* [A survey on Image Data Augmentation for Deep Learning](https://journalofbigdata.springeropen.com/articles/10.1186/s40537-019-0197-0)\n* [How to Configure Image Data Augmentation in Keras](https://machinelearningmastery.com/how-to-configure-image-data-augmentation-when-training-deep-learning-neural-networks/)\n\n", "1122556": "Just be aware that trading the larger EfficientNet versions gets challenging. B4 and B5 is still straightforward enough, but especially B7 and B8 becomes challenging on the standard GPU Kaggle provides (lots of notebooks using the architecture in all its variants out there though).", "1122999": "Hi,\n\nIt's not matter of augmentation,\nbut matter of size of input image for pre-trained network. \n\nThanks, at any rate.\n\n@bemoregt." }, "source": "meta" }