You can, but you don't have to. Especially when you fine-tune all layers it is not necessary as the CNN will adapt to it.
\n\nIf you freeze the EfficientNet I guess normalization is advised.
", "rawMarkdown": "You can, but you don't have to. Especially when you fine-tune all layers it is not necessary as the CNN will adapt to it.\n\nIf you freeze the EfficientNet I guess normalization is advised.", "votes": null }, { "id": "959019", "postDate": "08/05/2020 09:38:56", "content": "Alright. Thanks.
", "rawMarkdown": "Alright. Thanks.", "votes": null }, { "id": "959513", "postDate": "08/05/2020 16:32:19", "content": "@group16 Hi Gilles may I know why it is not necessary to normalize the img data in CNN?\nI though it is a rule of thumb to normalize img data before feeding into CNN as can converge faster without costing much extra resources?
\n\nthanks so much
", "rawMarkdown": "group16 Hi Gilles may I know why it is not necessary to normalize the img data in CNN?\nI though it is a rule of thumb to normalize img data before feeding into CNN as can converge faster without costing much extra resources?\n\nthanks so much", "votes": null }, { "id": "960274", "postDate": "08/06/2020 09:26:00", "content": "Of course you may ask @fiyeroleung !
\n\nI am no expert in deep learning at all, so don't take my word for granted. The normalization is just some linear transformation of your data (subtract mean, divide by std dev), which your neural network can learn as well.
\n\nIt could be much more important to ensure that the pre-processing of your images is the same as the imagenet dataset when you freeze some of the layers. But since we are fine-tuning all of the layers, the network quickly figures out that linear transformation.
", "rawMarkdown": "Of course you may ask @fiyeroleung !\n\nI am no expert in deep learning at all, so don't take my word for granted. The normalization is just some linear transformation of your data (subtract mean, divide by std dev), which your neural network can learn as well.\n\nIt could be much more important to ensure that the pre-processing of your images is the same as the imagenet dataset when you freeze some of the layers. But since we are fine-tuning all of the layers, the network quickly figures out that linear transformation.", "votes": null } ], "comments": [ { "id": 958811, "author_name": "group16", "author_url": "", "post_date": "08/05/2020 06:41:38", "content": "You can, but you don't have to. Especially when you fine-tune all layers it is not necessary as the CNN will adapt to it.
\n\nIf you freeze the EfficientNet I guess normalization is advised.
", "votes": null, "replies": [ { "id": 959019, "author_name": "faraksuli", "author_url": "", "post_date": "08/05/2020 09:38:56", "content": "Alright. Thanks.
", "votes": null, "replies": [] }, { "id": 959513, "author_name": "fiyeroleung", "author_url": "", "post_date": "08/05/2020 16:32:19", "content": "@group16 Hi Gilles may I know why it is not necessary to normalize the img data in CNN?\nI though it is a rule of thumb to normalize img data before feeding into CNN as can converge faster without costing much extra resources?
\n\nthanks so much
", "votes": null, "replies": [] }, { "id": 960274, "author_name": "group16", "author_url": "", "post_date": "08/06/2020 09:26:00", "content": "Of course you may ask @fiyeroleung !
\n\nI am no expert in deep learning at all, so don't take my word for granted. The normalization is just some linear transformation of your data (subtract mean, divide by std dev), which your neural network can learn as well.
\n\nIt could be much more important to ensure that the pre-processing of your images is the same as the imagenet dataset when you freeze some of the layers. But since we are fine-tuning all of the layers, the network quickly figures out that linear transformation.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "958786": "", "958811": "You can, but you don't have to. Especially when you fine-tune all layers it is not necessary as the CNN will adapt to it.\n\nIf you freeze the EfficientNet I guess normalization is advised.", "959019": "Alright. Thanks.", "959513": "group16 Hi Gilles may I know why it is not necessary to normalize the img data in CNN?\nI though it is a rule of thumb to normalize img data before feeding into CNN as can converge faster without costing much extra resources?\n\nthanks so much", "960274": "Of course you may ask @fiyeroleung !\n\nI am no expert in deep learning at all, so don't take my word for granted. The normalization is just some linear transformation of your data (subtract mean, divide by std dev), which your neural network can learn as well.\n\nIt could be much more important to ensure that the pre-processing of your images is the same as the imagenet dataset when you freeze some of the layers. But since we are fine-tuning all of the layers, the network quickly figures out that linear transformation." }, "source": "meta" }