# KerasAug
<!-- markdownlint-disable MD033 -->
[](https://github.com/james77777778/keras-aug/actions/workflows/actions.yml?query=branch%3Amain++)
[](https://codecov.io/gh/james77777778/keras-aug)
[](https://pypi.org/project/keras-aug/)
[](https://huggingface.co/spaces/james77777778/KerasAug)
## Description
KerasAug is a library that includes Keras 3 preprocessing and augmentation layers, providing support for various data types such as images, labels, bounding boxes, segmentation masks, and more.
<div align="center">
<img width="45%" src="https://github.com/user-attachments/assets/bf9488c4-5c6b-4c87-8fa8-30170a67c92c" alt="object_detection.gif"> <img width="45%" src="https://github.com/user-attachments/assets/556db949-9461-438a-b1cf-3621ec63416e" alt="semantic_segmentation.gif">
</div>
> [!NOTE]
> See `docs/*.py` for the GIF generation. YOLOV8-like pipeline for bounding boxes and segmentation masks.
KerasAug aims to provide fast, robust and user-friendly preprocessing and augmentation layers, facilitating seamless integration with Keras 3 and `tf.data`.
The APIs largely follow `torchvision`, and the correctness of the layers has been verified through unit tests.
Also, you can check out the demo app on HF:
App here: [](https://huggingface.co/spaces/james77777778/KerasAug)
## Why KerasAug
- 🚀 Supports many preprocessing & augmentation layers across all backends (JAX, TensorFlow and Torch).
- 🧰 Seamlessly integrates with `tf.data`, offering a performant and scalable data pipeline.
- 🔥 Follows the same API design as `torchvision`.
- 🙌 Depends only on Keras 3.
## Installation
```bash
pip install keras keras-aug -U
```
> [!IMPORTANT]
> Make sure you have installed a supported backend for Keras.
## Quickstart
### Rock, Paper and Scissors Image Classification
[](https://colab.research.google.com/drive/11xc0nW06iWQ_R-oH4wLB_MYV4GY4mNwy?usp=sharing)
```python
import keras
import tensorflow as tf
import tensorflow_datasets as tfds
from keras_aug import layers as ka_layers
BATCH_SIZE = 64
NUM_CLASSES = 3
INPUT_SIZE = (128, 128)
# Create a `tf.data.Dataset`-compatible preprocessing pipeline.
# Note that this example works with all backends.
train_dataset, validation_dataset = tfds.load(
"rock_paper_scissors", as_supervised=True, split=["train", "test"]
)
train_dataset = (
train_dataset.batch(BATCH_SIZE)
.map(
lambda images, labels: {
"images": tf.cast(images, "float32") / 255.0,
"labels": tf.one_hot(labels, NUM_CLASSES),
}
)
.map(ka_layers.vision.Resize(INPUT_SIZE))
.shuffle(128)
.map(ka_layers.vision.RandAugment())
.map(ka_layers.vision.CutMix(num_classes=NUM_CLASSES))
.map(ka_layers.vision.Rescale(scale=2.0, offset=-1)) # [0, 1] to [-1, 1]
.map(lambda data: (data["images"], data["labels"]))
.prefetch(tf.data.AUTOTUNE)
)
validation_dataset = (
validation_dataset.batch(BATCH_SIZE)
.map(
lambda images, labels: {
"images": tf.cast(images, "float32") / 255.0,
"labels": tf.one_hot(labels, NUM_CLASSES),
}
)
.map(ka_layers.vision.Resize(INPUT_SIZE))
.map(ka_layers.vision.Rescale(scale=2.0, offset=-1)) # [0, 1] to [-1, 1]
.map(lambda data: (data["images"], data["labels"]))
.prefetch(tf.data.AUTOTUNE)
)
# Create a model using MobileNetV2 as the backbone.
backbone = keras.applications.MobileNetV2(
input_shape=(*INPUT_SIZE, 3), include_top=False
)
backbone.trainable = False
inputs = keras.Input((*INPUT_SIZE, 3))
x = backbone(inputs)
x = keras.layers.GlobalAveragePooling2D()(x)
outputs = keras.layers.Dense(NUM_CLASSES, activation="softmax")(x)
model = keras.Model(inputs, outputs)
model.summary()
model.compile(
loss="categorical_crossentropy",
optimizer=keras.optimizers.SGD(learning_rate=1e-3, momentum=0.9),
metrics=["accuracy"],
)
# Train and evaluate your model
model.fit(train_dataset, validation_data=validation_dataset, epochs=8)
model.evaluate(validation_dataset)
```
The above example runs with all backends (JAX, TensorFlow, Torch).
### More Examples
- [YOLOV8 object detection pipeline](guides/voc_yolov8_aug.py) [](https://colab.research.google.com/drive/1AgnnvfTRMHKq--7gvmHP7RyxTeQResV4?usp=sharing)
- [YOLOV8 semantic segmentation pipeline](guides/oxford_yolov8_aug.py) [](https://colab.research.google.com/drive/1IJwUPiHreO7iIJ3VewgfLRoBdFdcQcJE?usp=sharing)
## Gradio App
```bash
gradio deploy
```
## Citing KerasAug
```bibtex
@misc{chiu2023kerasaug,
title={KerasAug},
author={Hongyu, Chiu},
year={2023},
howpublished={\url{https://github.com/james77777778/keras-aug}},
}
```
Raw data
{
"_id": null,
"home_page": null,
"name": "keras-aug",
"maintainer": null,
"docs_url": null,
"requires_python": ">=3.9",
"maintainer_email": "Hong-Yu Chiu <james77777778@gmail.com>",
"keywords": "deep-learning, preprocessing, augmentation, keras, jax, tensorflow, torch",
"author": null,
"author_email": "Hong-Yu Chiu <james77777778@gmail.com>",
"download_url": "https://files.pythonhosted.org/packages/82/68/f61fe767218fcfa4f4a180c311416489162c81acce626b52b0220d2b63cf/keras_aug-1.1.1.tar.gz",
"platform": null,
"description": "# KerasAug\n\n<!-- markdownlint-disable MD033 -->\n\n\n[](https://github.com/james77777778/keras-aug/actions/workflows/actions.yml?query=branch%3Amain++)\n[](https://codecov.io/gh/james77777778/keras-aug)\n[](https://pypi.org/project/keras-aug/)\n\n[](https://huggingface.co/spaces/james77777778/KerasAug)\n\n## Description\n\nKerasAug is a library that includes Keras 3 preprocessing and augmentation layers, providing support for various data types such as images, labels, bounding boxes, segmentation masks, and more.\n\n<div align=\"center\">\n<img width=\"45%\" src=\"https://github.com/user-attachments/assets/bf9488c4-5c6b-4c87-8fa8-30170a67c92c\" alt=\"object_detection.gif\"> <img width=\"45%\" src=\"https://github.com/user-attachments/assets/556db949-9461-438a-b1cf-3621ec63416e\" alt=\"semantic_segmentation.gif\">\n</div>\n\n> [!NOTE]\n> See `docs/*.py` for the GIF generation. YOLOV8-like pipeline for bounding boxes and segmentation masks.\n\nKerasAug aims to provide fast, robust and user-friendly preprocessing and augmentation layers, facilitating seamless integration with Keras 3 and `tf.data`.\n\nThe APIs largely follow `torchvision`, and the correctness of the layers has been verified through unit tests.\n\nAlso, you can check out the demo app on HF:\nApp here: [](https://huggingface.co/spaces/james77777778/KerasAug)\n\n## Why KerasAug\n\n- \ud83d\ude80 Supports many preprocessing & augmentation layers across all backends (JAX, TensorFlow and Torch).\n- \ud83e\uddf0 Seamlessly integrates with `tf.data`, offering a performant and scalable data pipeline.\n- \ud83d\udd25 Follows the same API design as `torchvision`.\n- \ud83d\ude4c Depends only on Keras 3.\n\n## Installation\n\n```bash\npip install keras keras-aug -U\n```\n\n> [!IMPORTANT] \n> Make sure you have installed a supported backend for Keras.\n\n## Quickstart\n\n### Rock, Paper and Scissors Image Classification\n\n[](https://colab.research.google.com/drive/11xc0nW06iWQ_R-oH4wLB_MYV4GY4mNwy?usp=sharing)\n\n```python\nimport keras\nimport tensorflow as tf\nimport tensorflow_datasets as tfds\n\nfrom keras_aug import layers as ka_layers\n\nBATCH_SIZE = 64\nNUM_CLASSES = 3\nINPUT_SIZE = (128, 128)\n\n# Create a `tf.data.Dataset`-compatible preprocessing pipeline.\n# Note that this example works with all backends.\ntrain_dataset, validation_dataset = tfds.load(\n \"rock_paper_scissors\", as_supervised=True, split=[\"train\", \"test\"]\n)\ntrain_dataset = (\n train_dataset.batch(BATCH_SIZE)\n .map(\n lambda images, labels: {\n \"images\": tf.cast(images, \"float32\") / 255.0,\n \"labels\": tf.one_hot(labels, NUM_CLASSES),\n }\n )\n .map(ka_layers.vision.Resize(INPUT_SIZE))\n .shuffle(128)\n .map(ka_layers.vision.RandAugment())\n .map(ka_layers.vision.CutMix(num_classes=NUM_CLASSES))\n .map(ka_layers.vision.Rescale(scale=2.0, offset=-1)) # [0, 1] to [-1, 1]\n .map(lambda data: (data[\"images\"], data[\"labels\"]))\n .prefetch(tf.data.AUTOTUNE)\n)\nvalidation_dataset = (\n validation_dataset.batch(BATCH_SIZE)\n .map(\n lambda images, labels: {\n \"images\": tf.cast(images, \"float32\") / 255.0,\n \"labels\": tf.one_hot(labels, NUM_CLASSES),\n }\n )\n .map(ka_layers.vision.Resize(INPUT_SIZE))\n .map(ka_layers.vision.Rescale(scale=2.0, offset=-1)) # [0, 1] to [-1, 1]\n .map(lambda data: (data[\"images\"], data[\"labels\"]))\n .prefetch(tf.data.AUTOTUNE)\n)\n\n# Create a model using MobileNetV2 as the backbone.\nbackbone = keras.applications.MobileNetV2(\n input_shape=(*INPUT_SIZE, 3), include_top=False\n)\nbackbone.trainable = False\ninputs = keras.Input((*INPUT_SIZE, 3))\nx = backbone(inputs)\nx = keras.layers.GlobalAveragePooling2D()(x)\noutputs = keras.layers.Dense(NUM_CLASSES, activation=\"softmax\")(x)\nmodel = keras.Model(inputs, outputs)\nmodel.summary()\nmodel.compile(\n loss=\"categorical_crossentropy\",\n optimizer=keras.optimizers.SGD(learning_rate=1e-3, momentum=0.9),\n metrics=[\"accuracy\"],\n)\n\n# Train and evaluate your model\nmodel.fit(train_dataset, validation_data=validation_dataset, epochs=8)\nmodel.evaluate(validation_dataset)\n```\n\nThe above example runs with all backends (JAX, TensorFlow, Torch).\n\n### More Examples\n\n- [YOLOV8 object detection pipeline](guides/voc_yolov8_aug.py) [](https://colab.research.google.com/drive/1AgnnvfTRMHKq--7gvmHP7RyxTeQResV4?usp=sharing)\n\n- [YOLOV8 semantic segmentation pipeline](guides/oxford_yolov8_aug.py) [](https://colab.research.google.com/drive/1IJwUPiHreO7iIJ3VewgfLRoBdFdcQcJE?usp=sharing)\n\n## Gradio App\n\n```bash\ngradio deploy\n```\n\n## Citing KerasAug\n\n```bibtex\n@misc{chiu2023kerasaug,\n title={KerasAug},\n author={Hongyu, Chiu},\n year={2023},\n howpublished={\\url{https://github.com/james77777778/keras-aug}},\n}\n```\n",
"bugtrack_url": null,
"license": "Apache License 2.0",
"summary": "A library that includes Keras 3 preprocessing and augmentation layers",
"version": "1.1.1",
"project_urls": {
"Documentation": "https://github.com/james77777778/keras-aug",
"Homepage": "https://github.com/james77777778/keras-aug",
"Issues": "https://github.com/james77777778/keras-aug/issues",
"Repository": "https://github.com/james77777778/keras-aug.git"
},
"split_keywords": [
"deep-learning",
" preprocessing",
" augmentation",
" keras",
" jax",
" tensorflow",
" torch"
],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "1f6dfc4a1efea49d4d6ae30de6bc0690ac04a1aa30ff8bf3c0a0c995ce48b163",
"md5": "e4a567c6b50a86e7d2af8abb21f6052f",
"sha256": "74d0925c9846e8b7b9b8b94aec246fbb0a4ce7ddadfa1bc5831bc0f2c83dce63"
},
"downloads": -1,
"filename": "keras_aug-1.1.1-py3-none-any.whl",
"has_sig": false,
"md5_digest": "e4a567c6b50a86e7d2af8abb21f6052f",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.9",
"size": 146114,
"upload_time": "2024-08-12T14:28:58",
"upload_time_iso_8601": "2024-08-12T14:28:58.371411Z",
"url": "https://files.pythonhosted.org/packages/1f/6d/fc4a1efea49d4d6ae30de6bc0690ac04a1aa30ff8bf3c0a0c995ce48b163/keras_aug-1.1.1-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "8268f61fe767218fcfa4f4a180c311416489162c81acce626b52b0220d2b63cf",
"md5": "6bfcdbff4f1975d2255d65dc08ce5a10",
"sha256": "8b9c22c288498c50c17f740675e51fe7cd07f9dc5c24c5f0c79c588007d18bf1"
},
"downloads": -1,
"filename": "keras_aug-1.1.1.tar.gz",
"has_sig": false,
"md5_digest": "6bfcdbff4f1975d2255d65dc08ce5a10",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.9",
"size": 82842,
"upload_time": "2024-08-12T14:28:59",
"upload_time_iso_8601": "2024-08-12T14:28:59.890188Z",
"url": "https://files.pythonhosted.org/packages/82/68/f61fe767218fcfa4f4a180c311416489162c81acce626b52b0220d2b63cf/keras_aug-1.1.1.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2024-08-12 14:28:59",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "james77777778",
"github_project": "keras-aug",
"travis_ci": false,
"coveralls": false,
"github_actions": true,
"requirements": [
{
"name": "torch",
"specs": []
},
{
"name": "torchvision",
"specs": []
},
{
"name": "keras",
"specs": []
}
],
"lcname": "keras-aug"
}
JSON
