Is anyone else having a hard time understanding how augmentations work in Ultralytics? The documentation is not very clear, and most of the questions on GitHub regarding augmentation are answered by AI. I thought I'd write this post in case anyone else has faced the same problem, and has found some clever solution.
\nFirst off, there doesn't seem to be any reliable way to disable augmentations. Setting augment=False in the .train() method doesn't seem to have any effect, nor does setting all augmentation parameters to 0. No matter which augmentation parameters I change, this line always appears in the logs:
albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01, num_output_channels=3, method=), CLAHE(p=0.01, clip_limit=(1.0, 4.0), tile_grid_size=(8, 8))\nThe only way I've found to remove this line is by uninstalling albumentations.
\nFrom what I understand, augmentations can be configured either via the YAML config file or by passing parameters directly to the .train() method. I've tried both approaches, but neither seems to work. Nothing I've done affects what is printed in the albumentations section of the logs. Either the augmentations I defined are being applied and not shown in the logs, or they're not being applied at all.
I've also struggled with implementing custom augmentations, as the options Ultralytics provides are quite limited. Based on this discussion, it seems there's no clean way to use custom augmentations other than modifying the source code. If anyone has found a better solution, I'd be grateful to hear about it.
", "messages": [ { "id": "3185309", "postDate": "04/23/2025 06:17:57", "content": "Is anyone else having a hard time understanding how augmentations work in Ultralytics? The documentation is not very clear, and most of the questions on GitHub regarding augmentation are answered by AI. I thought I'd write this post in case anyone else has faced the same problem, and has found some clever solution.
\nFirst off, there doesn't seem to be any reliable way to disable augmentations. Setting augment=False in the .train() method doesn't seem to have any effect, nor does setting all augmentation parameters to 0. No matter which augmentation parameters I change, this line always appears in the logs:
albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01, num_output_channels=3, method=), CLAHE(p=0.01, clip_limit=(1.0, 4.0), tile_grid_size=(8, 8))\nThe only way I've found to remove this line is by uninstalling albumentations.
\nFrom what I understand, augmentations can be configured either via the YAML config file or by passing parameters directly to the .train() method. I've tried both approaches, but neither seems to work. Nothing I've done affects what is printed in the albumentations section of the logs. Either the augmentations I defined are being applied and not shown in the logs, or they're not being applied at all.
I've also struggled with implementing custom augmentations, as the options Ultralytics provides are quite limited. Based on this discussion, it seems there's no clean way to use custom augmentations other than modifying the source code. If anyone has found a better solution, I'd be grateful to hear about it.
", "rawMarkdown": "Is anyone else having a hard time understanding how augmentations work in Ultralytics? The documentation is not very clear, and most of the questions on GitHub regarding augmentation are answered by AI. I thought I'd write this post in case anyone else has faced the same problem, and has found some clever solution.\n\n### Disabling Augmentations\nFirst off, there doesn't seem to be any reliable way to disable augmentations. Setting `augment=False` in the `.train()` method doesn't seem to have any effect, nor does setting all augmentation parameters to 0. No matter which augmentation parameters I change, this line always appears in the logs:\n\n```bash\nalbumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01, num_output_channels=3, method='weighted_average'), CLAHE(p=0.01, clip_limit=(1.0, 4.0), tile_grid_size=(8, 8))\n```\nThe only way I've found to remove this line is by uninstalling albumentations.\n\n\n### Customizing Augmentations Provided by Ultralytics\nFrom what I understand, augmentations can be configured either via the YAML config file or by passing parameters directly to the `.train()` method. I've tried both approaches, but neither seems to work. Nothing I've done affects what is printed in the albumentations section of the logs. Either the augmentations I defined are being applied and not shown in the logs, or they're not being applied at all.\n\n\n### Custom Augmentations\nI've also struggled with implementing custom augmentations, as the options Ultralytics provides are quite limited. Based on [this discussion](https://github.com/ultralytics/ultralytics/issues/9192#issuecomment-2708507624), it seems there's no clean way to use custom augmentations other than modifying the source code. If anyone has found a better solution, I'd be grateful to hear about it.", "votes": null }, { "id": "3185419", "postDate": "04/23/2025 09:30:44", "content": "You are correct with the conclusion that those albumentations are applied automatically unless you uninstall the library. Use this code snippet to create your own pipeline.
\n ultralytics.data.augment Albumentations\n ultralytics.utils LOGGER, colorstr\n albumentations A\n\n ():\n .p = p\n .transform = \n prefix = colorstr()\n :\n albumentations A\n\n \n\n \n T = [\n A.RandomRotate90(p=),\n A.HorizontalFlip(p=),\n A.VerticalFlip(p=),\n A.Rotate(limit=, p=),\n ]\n .transform = A.Compose(T, bbox_params=A.BboxParams(=, label_fields=[]))\n\n LOGGER.info(prefix + .join(.replace(, ) x T x.p))\n ImportError: \n \n Exception e:\n LOGGER.info()\n\nAlbumentations.__init__ = __init__\nHacky, but it works! Thank you @andreizamfir!
", "rawMarkdown": "Hacky, but it works! Thank you @andreizamfir!", "votes": null }, { "id": "3186160", "postDate": "04/24/2025 09:33:56", "content": "Full Discussion : https://github.com/ultralytics/ultralytics/issues/257
", "rawMarkdown": "Full Discussion : [https://github.com/ultralytics/ultralytics/issues/257](https://github.com/ultralytics/ultralytics/issues/257)", "votes": null }, { "id": "3186172", "postDate": "04/24/2025 09:56:32", "content": "For newer versions:
\n ():\n .p = p\n .transform = \n prefix = colorstr()\n\n :\n spatial_transforms = {\n ,\n ,\n \n }\n\n T = [\n A.HorizontalFlip(p=),\n A.RandomRotate90(p=),\n A.VerticalFlip(p=)\n ]\n\n .contains_spatial = (transform.__class__.__name__ spatial_transforms transform T)\n\n .transform = (\n A.Compose(T, bbox_params=A.BboxParams(=, label_fields=[]))\n .contains_spatial\n A.Compose(T)\n )\n\n (.transform, ):\n .transform.set_random_seed(torch.initial_seed())\n\n LOGGER.info(prefix + .join(.replace(, ) x T x.p))\n\n Exception e:\n LOGGER.info()\n\nAlbumentations.__init__ = __init__\nHave you tried to fine tune with this custom augmentations?
\nI recently used tune from yolo was too comput cost for me.
Thanks for your insight, that is right we can have full control over augmentations during our training this way.
", "rawMarkdown": "Thanks for your insight, that is right we can have full control over augmentations during our training this way.", "votes": null }, { "id": "3188814", "postDate": "04/28/2025 08:46:53", "content": "Thanks for your contribution
", "rawMarkdown": "Thanks for your contribution", "votes": null }, { "id": "3189212", "postDate": "04/29/2025 00:30:05", "content": "Very informative 😃
", "rawMarkdown": "Very informative 😃", "votes": null } ], "comments": [ { "id": 3185419, "author_name": "andreizamfir", "author_url": "", "post_date": "04/23/2025 09:30:44", "content": "You are correct with the conclusion that those albumentations are applied automatically unless you uninstall the library. Use this code snippet to create your own pipeline.
\n ultralytics.data.augment Albumentations\n ultralytics.utils LOGGER, colorstr\n albumentations A\n\n ():\n .p = p\n .transform = \n prefix = colorstr()\n :\n albumentations A\n\n \n\n \n T = [\n A.RandomRotate90(p=),\n A.HorizontalFlip(p=),\n A.VerticalFlip(p=),\n A.Rotate(limit=, p=),\n ]\n .transform = A.Compose(T, bbox_params=A.BboxParams(=, label_fields=[]))\n\n LOGGER.info(prefix + .join(.replace(, ) x T x.p))\n ImportError: \n \n Exception e:\n LOGGER.info()\n\nAlbumentations.__init__ = __init__\nHacky, but it works! Thank you @andreizamfir!
", "votes": null, "replies": [ { "id": 3186172, "author_name": "ravaghi", "author_url": "", "post_date": "04/24/2025 09:56:32", "content": "For newer versions:
\n ():\n .p = p\n .transform = \n prefix = colorstr()\n\n :\n spatial_transforms = {\n ,\n ,\n \n }\n\n T = [\n A.HorizontalFlip(p=),\n A.RandomRotate90(p=),\n A.VerticalFlip(p=)\n ]\n\n .contains_spatial = (transform.__class__.__name__ spatial_transforms transform T)\n\n .transform = (\n A.Compose(T, bbox_params=A.BboxParams(=, label_fields=[]))\n .contains_spatial\n A.Compose(T)\n )\n\n (.transform, ):\n .transform.set_random_seed(torch.initial_seed())\n\n LOGGER.info(prefix + .join(.replace(, ) x T x.p))\n\n Exception e:\n LOGGER.info()\n\nAlbumentations.__init__ = __init__\nFull Discussion : https://github.com/ultralytics/ultralytics/issues/257
", "votes": null, "replies": [] }, { "id": 3186398, "author_name": "mathieuduverne", "author_url": "", "post_date": "04/24/2025 15:43:55", "content": "Have you tried to fine tune with this custom augmentations?
\nI recently used tune from yolo was too comput cost for me.
Thanks for your insight, that is right we can have full control over augmentations during our training this way.
", "votes": null, "replies": [] }, { "id": 3188814, "author_name": "mohanapavanbezawada", "author_url": "", "post_date": "04/28/2025 08:46:53", "content": "Thanks for your contribution
", "votes": null, "replies": [] }, { "id": 3189212, "author_name": "dataneed99", "author_url": "", "post_date": "04/29/2025 00:30:05", "content": "Very informative 😃
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3185309": "Is anyone else having a hard time understanding how augmentations work in Ultralytics? The documentation is not very clear, and most of the questions on GitHub regarding augmentation are answered by AI. I thought I'd write this post in case anyone else has faced the same problem, and has found some clever solution.\n\n### Disabling Augmentations\nFirst off, there doesn't seem to be any reliable way to disable augmentations. Setting `augment=False` in the `.train()` method doesn't seem to have any effect, nor does setting all augmentation parameters to 0. No matter which augmentation parameters I change, this line always appears in the logs:\n\n```bash\nalbumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01, num_output_channels=3, method='weighted_average'), CLAHE(p=0.01, clip_limit=(1.0, 4.0), tile_grid_size=(8, 8))\n```\nThe only way I've found to remove this line is by uninstalling albumentations.\n\n\n### Customizing Augmentations Provided by Ultralytics\nFrom what I understand, augmentations can be configured either via the YAML config file or by passing parameters directly to the `.train()` method. I've tried both approaches, but neither seems to work. Nothing I've done affects what is printed in the albumentations section of the logs. Either the augmentations I defined are being applied and not shown in the logs, or they're not being applied at all.\n\n\n### Custom Augmentations\nI've also struggled with implementing custom augmentations, as the options Ultralytics provides are quite limited. Based on [this discussion](https://github.com/ultralytics/ultralytics/issues/9192#issuecomment-2708507624), it seems there's no clean way to use custom augmentations other than modifying the source code. If anyone has found a better solution, I'd be grateful to hear about it.", "3185419": "You are correct with the conclusion that those albumentations are applied automatically unless you uninstall the library. Use this code snippet to create your own pipeline.\n\n```\nfrom ultralytics.data.augment import Albumentations\nfrom ultralytics.utils import LOGGER, colorstr\nimport albumentations as A\n\ndef __init__(self, p=1.0):\n self.p = p\n self.transform = None\n prefix = colorstr(\"albumentations: \")\n try:\n import albumentations as A\n\n # check_version(A.__version__, \"1.0.3\", hard=True) # version requirement\n\n # Transforms\n T = [\n A.RandomRotate90(p=0.8),\n A.HorizontalFlip(p=0.5),\n A.VerticalFlip(p=0.5),\n A.Rotate(limit=15, p=0.3),\n ]\n self.transform = A.Compose(T, bbox_params=A.BboxParams(format=\"yolo\", label_fields=[\"class_labels\"]))\n\n LOGGER.info(prefix + \", \".join(f\"{x}\".replace(\"always_apply=False, \", \"\") for x in T if x.p))\n except ImportError: # package not installed, skip\n pass\n except Exception as e:\n LOGGER.info(f\"{prefix}{e}\")\n\nAlbumentations.__init__ = __init__\n```", "3185441": "Hacky, but it works! Thank you @andreizamfir!", "3186160": "Full Discussion : [https://github.com/ultralytics/ultralytics/issues/257](https://github.com/ultralytics/ultralytics/issues/257)", "3186172": "For newer versions:\n```python\ndef __init__(self, p=1.0):\n self.p = p\n self.transform = None\n prefix = colorstr(\"albumentations: \")\n\n try:\n spatial_transforms = {\n \"HorizontalFlip\",\n \"RandomRotate90\",\n \"VerticalFlip\"\n }\n \n T = [\n A.HorizontalFlip(p=0.3),\n A.RandomRotate90(p=0.3),\n A.VerticalFlip(p=0.3)\n ]\n\n self.contains_spatial = any(transform.__class__.__name__ in spatial_transforms for transform in T)\n \n self.transform = (\n A.Compose(T, bbox_params=A.BboxParams(format=\"yolo\", label_fields=[\"class_labels\"]))\n if self.contains_spatial\n else A.Compose(T)\n )\n \n if hasattr(self.transform, \"set_random_seed\"):\n self.transform.set_random_seed(torch.initial_seed())\n \n LOGGER.info(prefix + \", \".join(f\"{x}\".replace(\"always_apply=False, \", \"\") for x in T if x.p))\n \n except Exception as e:\n LOGGER.info(f\"{prefix}{e}\")\n\nAlbumentations.__init__ = __init__\n```", "3186398": "Have you tried to fine tune with this custom augmentations?\nI recently used tune from yolo was too comput cost for me.", "3188089": "Thanks for your insight, that is right we can have full control over augmentations during our training this way.", "3188814": "Thanks for your contribution", "3189212": "Very informative 😃" }, "source": "meta" }