Hello guys,
\n\nwhile looking at the data, I was just wondering why and how are people here using only a 3 channel RGB image per site as their input of the CNN. The provided scans have 6 channels per site, but I barely saw anyone using a 6 channel input in their kernel. Did I get something wrong here ?
\n\nThanks in advance
", "messages": [ { "id": "597330", "postDate": "08/12/2019 07:27:26", "content": "Hello guys,
\n\nwhile looking at the data, I was just wondering why and how are people here using only a 3 channel RGB image per site as their input of the CNN. The provided scans have 6 channels per site, but I barely saw anyone using a 6 channel input in their kernel. Did I get something wrong here ?
\n\nThanks in advance
", "rawMarkdown": "Hello guys,\n\nwhile looking at the data, I was just wondering why and how are people here using only a 3 channel RGB image per site as their input of the CNN. The provided scans have 6 channels per site, but I barely saw anyone using a 6 channel input in their kernel. Did I get something wrong here ?\n\nThanks in advance", "votes": null }, { "id": "597335", "postDate": "08/12/2019 07:34:37", "content": "The reason why people are using 3 channel RGB is because it is faster, which is important for kernels. It is also slightly easier to work with, and allows graphical representation, again good fit for kernels. But I believe everyone uses original 6 channel in their \"production\" models.
", "rawMarkdown": "The reason why people are using 3 channel RGB is because it is faster, which is important for kernels. It is also slightly easier to work with, and allows graphical representation, again good fit for kernels. But I believe everyone uses original 6 channel in their \"production\" models.", "votes": null }, { "id": "597345", "postDate": "08/12/2019 07:54:28", "content": "Hey, the rxrx1 repo contains functions that load or transform a 6-channel image into an RGB image. Another advantage of using RGB would be to be able to use pretrained weights for example on ImageNet. However, it's pretty easy (at least in pytorch) to modify the architecture of a network to receive 6-channel inputs, and likewise it's easy to modify the pretrained weights to accommodate the 6-channel-input first layer. Your choice!
", "rawMarkdown": "Hey, the rxrx1 repo contains functions that load or transform a 6-channel image into an RGB image. Another advantage of using RGB would be to be able to use pretrained weights for example on ImageNet. However, it's pretty easy (at least in pytorch) to modify the architecture of a network to receive 6-channel inputs, and likewise it's easy to modify the pretrained weights to accommodate the 6-channel-input first layer. Your choice!", "votes": null }, { "id": "597351", "postDate": "08/12/2019 07:58:21", "content": "BTW, this has already been discussed https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/100624#latest-581498
", "rawMarkdown": "BTW, this has already been discussed https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/100624#latest-581498", "votes": null }, { "id": "598170", "postDate": "08/13/2019 08:25:54", "content": "Hey I just checked out the rxrx1 toolkit and saw that conversion function. Now it makes sense what the logic behind the compression is. Thanks everyone for the answers !
", "rawMarkdown": "Hey I just checked out the rxrx1 toolkit and saw that conversion function. Now it makes sense what the logic behind the compression is. Thanks everyone for the answers !", "votes": null }, { "id": "614861", "postDate": "09/01/2019 06:50:40", "content": "> it's easy to modify the pretrained weights to accommodate the 6-channel-input first layer\nWhat do you think is the most logical way to do that?
", "rawMarkdown": "> it's easy to modify the pretrained weights to accommodate the 6-channel-input first layer\nWhat do you think is the most logical way to do that?", "votes": null } ], "comments": [ { "id": 597335, "author_name": "zaharch", "author_url": "", "post_date": "08/12/2019 07:34:37", "content": "The reason why people are using 3 channel RGB is because it is faster, which is important for kernels. It is also slightly easier to work with, and allows graphical representation, again good fit for kernels. But I believe everyone uses original 6 channel in their \"production\" models.
", "votes": null, "replies": [] }, { "id": 597345, "author_name": "giuliasavorgnan", "author_url": "", "post_date": "08/12/2019 07:54:28", "content": "Hey, the rxrx1 repo contains functions that load or transform a 6-channel image into an RGB image. Another advantage of using RGB would be to be able to use pretrained weights for example on ImageNet. However, it's pretty easy (at least in pytorch) to modify the architecture of a network to receive 6-channel inputs, and likewise it's easy to modify the pretrained weights to accommodate the 6-channel-input first layer. Your choice!
", "votes": null, "replies": [ { "id": 614861, "author_name": "abyaadrafid", "author_url": "", "post_date": "09/01/2019 06:50:40", "content": "> it's easy to modify the pretrained weights to accommodate the 6-channel-input first layer\nWhat do you think is the most logical way to do that?
", "votes": null, "replies": [] } ] }, { "id": 597351, "author_name": "giuliasavorgnan", "author_url": "", "post_date": "08/12/2019 07:58:21", "content": "BTW, this has already been discussed https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/100624#latest-581498
", "votes": null, "replies": [] }, { "id": 598170, "author_name": "stanislavmalorodov", "author_url": "", "post_date": "08/13/2019 08:25:54", "content": "Hey I just checked out the rxrx1 toolkit and saw that conversion function. Now it makes sense what the logic behind the compression is. Thanks everyone for the answers !
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "597330": "Hello guys,\n\nwhile looking at the data, I was just wondering why and how are people here using only a 3 channel RGB image per site as their input of the CNN. The provided scans have 6 channels per site, but I barely saw anyone using a 6 channel input in their kernel. Did I get something wrong here ?\n\nThanks in advance", "597335": "The reason why people are using 3 channel RGB is because it is faster, which is important for kernels. It is also slightly easier to work with, and allows graphical representation, again good fit for kernels. But I believe everyone uses original 6 channel in their \"production\" models.", "597345": "Hey, the rxrx1 repo contains functions that load or transform a 6-channel image into an RGB image. Another advantage of using RGB would be to be able to use pretrained weights for example on ImageNet. However, it's pretty easy (at least in pytorch) to modify the architecture of a network to receive 6-channel inputs, and likewise it's easy to modify the pretrained weights to accommodate the 6-channel-input first layer. Your choice!", "597351": "BTW, this has already been discussed https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/100624#latest-581498", "598170": "Hey I just checked out the rxrx1 toolkit and saw that conversion function. Now it makes sense what the logic behind the compression is. Thanks everyone for the answers !", "614861": "> it's easy to modify the pretrained weights to accommodate the 6-channel-input first layer\nWhat do you think is the most logical way to do that?" }, "source": "meta" }