If I apply the following augmentations to my images as shown:
\ntrain_augmentations = A.Compose([\n A.RandomCrop(image_size, image_size, p=1),\n A.CoarseDropout(p=0.5),\n A.Cutout(p=0.5),\n A.Flip(p=0.5),\n A.ShiftScaleRotate(p=0.5),\n A.HueSaturationValue(p=0.5, hue_shift_limit=0.2, sat_shift_limit=0.2, val_shift_limit=0.2),\n A.RandomBrightnessContrast(p=0.5, brightness_limit=(-0.2,0.2), contrast_limit=(-0.2, 0.2)),\n A.ToFloat()\n ], p=1)
My efficientnet doesn't train properly, train accuracy goes up but val accuracy just gets stuck at 0.1+. But if I remove the last augmentation, ToFloat(), then it trains fine. Could someone please explain to me why? I didn't have this problem with Xception/Inception..
\nEdit: I believe the problem might not lie with converting to float itself, but rather the rescaling by 1/255 when the transformation ToFloat() is applied.. Still don't understand why though.
", "messages": [ { "id": "1165852", "postDate": "01/23/2021 09:21:01", "content": "If I apply the following augmentations to my images as shown:
\ntrain_augmentations = A.Compose([\n A.RandomCrop(image_size, image_size, p=1),\n A.CoarseDropout(p=0.5),\n A.Cutout(p=0.5),\n A.Flip(p=0.5),\n A.ShiftScaleRotate(p=0.5),\n A.HueSaturationValue(p=0.5, hue_shift_limit=0.2, sat_shift_limit=0.2, val_shift_limit=0.2),\n A.RandomBrightnessContrast(p=0.5, brightness_limit=(-0.2,0.2), contrast_limit=(-0.2, 0.2)),\n A.ToFloat()\n ], p=1)
My efficientnet doesn't train properly, train accuracy goes up but val accuracy just gets stuck at 0.1+. But if I remove the last augmentation, ToFloat(), then it trains fine. Could someone please explain to me why? I didn't have this problem with Xception/Inception..
\nEdit: I believe the problem might not lie with converting to float itself, but rather the rescaling by 1/255 when the transformation ToFloat() is applied.. Still don't understand why though.
", "rawMarkdown": "If I apply the following augmentations to my images as shown:\n\n`train_augmentations = A.Compose([\n A.RandomCrop(image_size, image_size, p=1),\n A.CoarseDropout(p=0.5),\n A.Cutout(p=0.5),\n A.Flip(p=0.5),\n A.ShiftScaleRotate(p=0.5),\n A.HueSaturationValue(p=0.5, hue_shift_limit=0.2, sat_shift_limit=0.2, val_shift_limit=0.2),\n A.RandomBrightnessContrast(p=0.5, brightness_limit=(-0.2,0.2), contrast_limit=(-0.2, 0.2)),\n A.ToFloat()\n ], p=1)`\n\nMy efficientnet doesn't train properly, train accuracy goes up but val accuracy just gets stuck at 0.1+. But if I remove the last augmentation, ToFloat(), then it trains fine. Could someone please explain to me why? I didn't have this problem with Xception/Inception.. \n\nEdit: I believe the problem might not lie with converting to float itself, but rather the rescaling by 1/255 when the transformation ToFloat() is applied.. Still don't understand why though.", "votes": null }, { "id": "1165920", "postDate": "01/23/2021 10:09:51", "content": "Look at the first few layers of Efficientnet model - they contain Rescaling layer and Normalization layer - so you don't need to rescale the input going into the model.
\n```\nLayer (type) Output Shape Param # Connected to \n==================================================================================================\ninput_1 (InputLayer) [(None, 512, 512, 3) 0 \n__________________________________________________________________________________________________\nrescaling (Rescaling) (None, 512, 512, 3) 0 input_1[0][0] \n__________________________________________________________________________________________________\nnormalization (Normalization) (None, 512, 512, 3) 7 rescaling[0][0] \n```
", "rawMarkdown": "Look at the first few layers of Efficientnet model - they contain Rescaling layer and Normalization layer - so you don't need to rescale the input going into the model.\n\n```\n```\nLayer (type) Output Shape Param # Connected to \n==================================================================================================\ninput_1 (InputLayer) [(None, 512, 512, 3) 0 \n__________________________________________________________________________________________________\nrescaling (Rescaling) (None, 512, 512, 3) 0 input_1[0][0] \n__________________________________________________________________________________________________\nnormalization (Normalization) (None, 512, 512, 3) 7 rescaling[0][0] \n```\n```", "votes": null }, { "id": "1165931", "postDate": "01/23/2021 10:21:47", "content": "Thanks Jimmy!
", "rawMarkdown": "Thanks Jimmy!", "votes": null }, { "id": "1166567", "postDate": "01/23/2021 17:32:01", "content": "Your welcome - sadly I learned this the hard way - spent lots of hours trying to understand why code similiar to yours did not work :)
\nThere is a post buried somewhere in this Discussion were someone pointed this out for Efficientnets. I found the post after lots of wasted hours.
\nI know as a good bit of advice that I should read all the recent Discussion posts every day - but I also know that I very often fail to follow my own advice. :)
", "rawMarkdown": "Your welcome - sadly I learned this the hard way - spent lots of hours trying to understand why code similiar to yours did not work :)\n\nThere is a post buried somewhere in this Discussion were someone pointed this out for Efficientnets. I found the post after lots of wasted hours.\n\nI know as a good bit of advice that I should read all the recent Discussion posts every day - but I also know that I very often fail to follow my own advice. :)", "votes": null }, { "id": "1166926", "postDate": "01/23/2021 23:07:42", "content": "Maybe a further helping comment according to my understanding… The above screenshot seems to come from the tensorflow.keras.application package. However if you use PyPI efficientnet==1.1.1 (because the tensorflow version of the TPU environment does not include efficientnet yet) then this rescaling and normalization layers are not present. As far as I know, each tensorflow.keras.application (and also PyPI efficientnet package) usually have a preprocess_input funtion in the corresponding module that you should apply to the image with floating values in the range [0., 255.].
Thanks, that's good to know!
", "rawMarkdown": "Thanks, that's good to know!", "votes": null }, { "id": "1167116", "postDate": "01/24/2021 04:10:46", "content": "Thanks. I just noticed this detail.
", "rawMarkdown": "Thanks. I just noticed this detail.", "votes": null }, { "id": "1167150", "postDate": "01/24/2021 05:05:46", "content": "\nYour correct - my screenshot was from the tf.keras.application version of efficientnet. The scaling and normalization issues had me confused for almost two weeks.
\nThe lesson I learned - ALWAYS generate a model.summary and take a close look at the first several layers and the last couple. I generated the summaries but never looked at them.
", "rawMarkdown": "[tmoroder](https://www.kaggle.com/morodertobias)\n\n\nYour correct - my screenshot was from the tf.keras.application version of efficientnet. The scaling and normalization issues had me confused for almost two weeks.\n\nThe lesson I learned - ALWAYS generate a model.summary and take a close look at the first several layers and the last couple. I generated the summaries but never looked at them.", "votes": null } ], "comments": [ { "id": 1165920, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/23/2021 10:09:51", "content": "Look at the first few layers of Efficientnet model - they contain Rescaling layer and Normalization layer - so you don't need to rescale the input going into the model.
\n```\nLayer (type) Output Shape Param # Connected to \n==================================================================================================\ninput_1 (InputLayer) [(None, 512, 512, 3) 0 \n__________________________________________________________________________________________________\nrescaling (Rescaling) (None, 512, 512, 3) 0 input_1[0][0] \n__________________________________________________________________________________________________\nnormalization (Normalization) (None, 512, 512, 3) 7 rescaling[0][0] \n```
", "votes": null, "replies": [ { "id": 1165931, "author_name": "junyingsg", "author_url": "", "post_date": "01/23/2021 10:21:47", "content": "Thanks Jimmy!
", "votes": null, "replies": [] }, { "id": 1166567, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/23/2021 17:32:01", "content": "Your welcome - sadly I learned this the hard way - spent lots of hours trying to understand why code similiar to yours did not work :)
\nThere is a post buried somewhere in this Discussion were someone pointed this out for Efficientnets. I found the post after lots of wasted hours.
\nI know as a good bit of advice that I should read all the recent Discussion posts every day - but I also know that I very often fail to follow my own advice. :)
", "votes": null, "replies": [] }, { "id": 1166926, "author_name": "morodertobias", "author_url": "", "post_date": "01/23/2021 23:07:42", "content": "Maybe a further helping comment according to my understanding… The above screenshot seems to come from the tensorflow.keras.application package. However if you use PyPI efficientnet==1.1.1 (because the tensorflow version of the TPU environment does not include efficientnet yet) then this rescaling and normalization layers are not present. As far as I know, each tensorflow.keras.application (and also PyPI efficientnet package) usually have a preprocess_input funtion in the corresponding module that you should apply to the image with floating values in the range [0., 255.].
Thanks, that's good to know!
", "votes": null, "replies": [] }, { "id": 1167116, "author_name": "zhangeng", "author_url": "", "post_date": "01/24/2021 04:10:46", "content": "Thanks. I just noticed this detail.
", "votes": null, "replies": [] }, { "id": 1167150, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/24/2021 05:05:46", "content": "\nYour correct - my screenshot was from the tf.keras.application version of efficientnet. The scaling and normalization issues had me confused for almost two weeks.
\nThe lesson I learned - ALWAYS generate a model.summary and take a close look at the first several layers and the last couple. I generated the summaries but never looked at them.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1165852": "If I apply the following augmentations to my images as shown:\n\n`train_augmentations = A.Compose([\n A.RandomCrop(image_size, image_size, p=1),\n A.CoarseDropout(p=0.5),\n A.Cutout(p=0.5),\n A.Flip(p=0.5),\n A.ShiftScaleRotate(p=0.5),\n A.HueSaturationValue(p=0.5, hue_shift_limit=0.2, sat_shift_limit=0.2, val_shift_limit=0.2),\n A.RandomBrightnessContrast(p=0.5, brightness_limit=(-0.2,0.2), contrast_limit=(-0.2, 0.2)),\n A.ToFloat()\n ], p=1)`\n\nMy efficientnet doesn't train properly, train accuracy goes up but val accuracy just gets stuck at 0.1+. But if I remove the last augmentation, ToFloat(), then it trains fine. Could someone please explain to me why? I didn't have this problem with Xception/Inception.. \n\nEdit: I believe the problem might not lie with converting to float itself, but rather the rescaling by 1/255 when the transformation ToFloat() is applied.. Still don't understand why though.", "1165920": "Look at the first few layers of Efficientnet model - they contain Rescaling layer and Normalization layer - so you don't need to rescale the input going into the model.\n\n```\n```\nLayer (type) Output Shape Param # Connected to \n==================================================================================================\ninput_1 (InputLayer) [(None, 512, 512, 3) 0 \n__________________________________________________________________________________________________\nrescaling (Rescaling) (None, 512, 512, 3) 0 input_1[0][0] \n__________________________________________________________________________________________________\nnormalization (Normalization) (None, 512, 512, 3) 7 rescaling[0][0] \n```\n```", "1165931": "Thanks Jimmy!", "1166567": "Your welcome - sadly I learned this the hard way - spent lots of hours trying to understand why code similiar to yours did not work :)\n\nThere is a post buried somewhere in this Discussion were someone pointed this out for Efficientnets. I found the post after lots of wasted hours.\n\nI know as a good bit of advice that I should read all the recent Discussion posts every day - but I also know that I very often fail to follow my own advice. :)", "1166926": "Maybe a further helping comment according to my understanding... The above screenshot seems to come from the ``tensorflow.keras.application`` package. However if you use PyPI ``efficientnet==1.1.1`` (because the tensorflow version of the TPU environment does not include efficientnet yet) then this rescaling and normalization layers are not present. As far as I know, each tensorflow.keras.application (and also PyPI efficientnet package) usually have a ``preprocess_input`` funtion in the corresponding module that you should apply to the image with floating values in the range ``[0., 255.]``.", "1167012": "Thanks, that's good to know!", "1167116": "Thanks. I just noticed this detail.", "1167150": "[tmoroder](https://www.kaggle.com/morodertobias)\n\n\nYour correct - my screenshot was from the tf.keras.application version of efficientnet. The scaling and normalization issues had me confused for almost two weeks.\n\nThe lesson I learned - ALWAYS generate a model.summary and take a close look at the first several layers and the last couple. I generated the summaries but never looked at them." }, "source": "meta" }