Hi,
\n\nI'm implementing SSD for this challenge which uses VGG 16 as its base network. Can I use pretrained VGG weights or should I go for training the model from scratch? Given that the images in this challege are chest xrays and the pretrained VGG was trained on imagenet (with RBG images).
", "messages": [ { "id": "399146", "postDate": "10/05/2018 09:36:12", "content": "Hi,
\n\nI'm implementing SSD for this challenge which uses VGG 16 as its base network. Can I use pretrained VGG weights or should I go for training the model from scratch? Given that the images in this challege are chest xrays and the pretrained VGG was trained on imagenet (with RBG images).
", "rawMarkdown": "Hi,\n\nI'm implementing SSD for this challenge which uses VGG 16 as its base network. Can I use pretrained VGG weights or should I go for training the model from scratch? Given that the images in this challege are chest xrays and the pretrained VGG was trained on imagenet (with RBG images).", "votes": null }, { "id": "399149", "postDate": "10/05/2018 09:43:27", "content": "Well it would certainly help you if you use pretrained weights, as the first few layers just detect lines, curves, colour and other general features that would remain useful for this dataset as well.
", "rawMarkdown": "Well it would certainly help you if you use pretrained weights, as the first few layers just detect lines, curves, colour and other general features that would remain useful for this dataset as well.", "votes": null }, { "id": "401862", "postDate": "10/10/2018 19:43:10", "content": "Probably yes. But it's not so obvious for me. Cloud-like objects are not the same as real imagenet objects.
", "rawMarkdown": "Probably yes. But it's not so obvious for me. Cloud-like objects are not the same as real imagenet objects.", "votes": null }, { "id": "403704", "postDate": "10/14/2018 11:48:39", "content": "Pretrained weights should give you at least a good weights initialisation, if nothing else. It's worth to check both but in most cases using transfer learning even between different domains does work better.
", "rawMarkdown": "Pretrained weights should give you at least a good weights initialisation, if nothing else. It's worth to check both but in most cases using transfer learning even between different domains does work better.", "votes": null }, { "id": "404448", "postDate": "10/15/2018 18:00:26", "content": "@ Dmytro, Are you using basic or heavy augmentation techniques? I'm using Mask RCCN and I noticed that there is a lot of variation in the val loss once augmentations are applied. So I was just curious to know if heavy augmentation methods are being used or if I am just over thinking.
", "rawMarkdown": "Dmytro, Are you using basic or heavy augmentation techniques? I'm using Mask RCCN and I noticed that there is a lot of variation in the val loss once augmentations are applied. So I was just curious to know if heavy augmentation methods are being used or if I am just over thinking.", "votes": null }, { "id": "404536", "postDate": "10/15/2018 23:15:42", "content": "I don't use very heavy augmentation, usually I try to apply the maximum level of augmentations while ensuring the labels are still correct and recognisable (unless the dataset is very large and augmentations are not as important). In this case I can't tell if changing the brightness for example still allows to recognise pneumonia, so I decided to limit augmentations changing pixel brightness to very mild range. I still apply horizontal flips, small rotations and scale.
", "rawMarkdown": "I don't use very heavy augmentation, usually I try to apply the maximum level of augmentations while ensuring the labels are still correct and recognisable (unless the dataset is very large and augmentations are not as important). In this case I can't tell if changing the brightness for example still allows to recognise pneumonia, so I decided to limit augmentations changing pixel brightness to very mild range. I still apply horizontal flips, small rotations and scale.", "votes": null }, { "id": "404541", "postDate": "10/15/2018 23:36:03", "content": "@Dmytro, thank you very much.
", "rawMarkdown": "Dmytro, thank you very much.", "votes": null }, { "id": "404986", "postDate": "10/16/2018 17:39:29", "content": "@rishabh where are you training your model?
", "rawMarkdown": "rishabh where are you training your model?", "votes": null }, { "id": "407993", "postDate": "10/22/2018 05:47:10", "content": "On my college server's GPU and https://vast.ai
", "rawMarkdown": "On my college server's GPU and https://vast.ai", "votes": null } ], "comments": [ { "id": 399149, "author_name": "nazimgirach", "author_url": "", "post_date": "10/05/2018 09:43:27", "content": "Well it would certainly help you if you use pretrained weights, as the first few layers just detect lines, curves, colour and other general features that would remain useful for this dataset as well.
", "votes": null, "replies": [ { "id": 401862, "author_name": "sergeyzlobin", "author_url": "", "post_date": "10/10/2018 19:43:10", "content": "Probably yes. But it's not so obvious for me. Cloud-like objects are not the same as real imagenet objects.
", "votes": null, "replies": [] } ] }, { "id": 403704, "author_name": "dmytropoplavskiy", "author_url": "", "post_date": "10/14/2018 11:48:39", "content": "Pretrained weights should give you at least a good weights initialisation, if nothing else. It's worth to check both but in most cases using transfer learning even between different domains does work better.
", "votes": null, "replies": [ { "id": 404448, "author_name": "dskswu", "author_url": "", "post_date": "10/15/2018 18:00:26", "content": "@ Dmytro, Are you using basic or heavy augmentation techniques? I'm using Mask RCCN and I noticed that there is a lot of variation in the val loss once augmentations are applied. So I was just curious to know if heavy augmentation methods are being used or if I am just over thinking.
", "votes": null, "replies": [] }, { "id": 404536, "author_name": "dmytropoplavskiy", "author_url": "", "post_date": "10/15/2018 23:15:42", "content": "I don't use very heavy augmentation, usually I try to apply the maximum level of augmentations while ensuring the labels are still correct and recognisable (unless the dataset is very large and augmentations are not as important). In this case I can't tell if changing the brightness for example still allows to recognise pneumonia, so I decided to limit augmentations changing pixel brightness to very mild range. I still apply horizontal flips, small rotations and scale.
", "votes": null, "replies": [] }, { "id": 404541, "author_name": "dskswu", "author_url": "", "post_date": "10/15/2018 23:36:03", "content": "@Dmytro, thank you very much.
", "votes": null, "replies": [] } ] }, { "id": 404986, "author_name": "nitishsingh41", "author_url": "", "post_date": "10/16/2018 17:39:29", "content": "@rishabh where are you training your model?
", "votes": null, "replies": [ { "id": 407993, "author_name": "rishabhiitbhu", "author_url": "", "post_date": "10/22/2018 05:47:10", "content": "On my college server's GPU and https://vast.ai
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "399146": "Hi,\n\nI'm implementing SSD for this challenge which uses VGG 16 as its base network. Can I use pretrained VGG weights or should I go for training the model from scratch? Given that the images in this challege are chest xrays and the pretrained VGG was trained on imagenet (with RBG images).", "399149": "Well it would certainly help you if you use pretrained weights, as the first few layers just detect lines, curves, colour and other general features that would remain useful for this dataset as well.", "401862": "Probably yes. But it's not so obvious for me. Cloud-like objects are not the same as real imagenet objects.", "403704": "Pretrained weights should give you at least a good weights initialisation, if nothing else. It's worth to check both but in most cases using transfer learning even between different domains does work better.", "404448": "Dmytro, Are you using basic or heavy augmentation techniques? I'm using Mask RCCN and I noticed that there is a lot of variation in the val loss once augmentations are applied. So I was just curious to know if heavy augmentation methods are being used or if I am just over thinking.", "404536": "I don't use very heavy augmentation, usually I try to apply the maximum level of augmentations while ensuring the labels are still correct and recognisable (unless the dataset is very large and augmentations are not as important). In this case I can't tell if changing the brightness for example still allows to recognise pneumonia, so I decided to limit augmentations changing pixel brightness to very mild range. I still apply horizontal flips, small rotations and scale.", "404541": "Dmytro, thank you very much.", "404986": "rishabh where are you training your model?", "407993": "On my college server's GPU and https://vast.ai" }, "source": "meta" }