Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/tillbeemelmanns/tfops-aug
TFOps-Aug: Implementation of policy-based image augmentation techniques based on TF2 Operations. All augmentations as efficient Tensorflow 2.11.0 operations. Easy integration into a tf.data API pipeline.
https://github.com/tillbeemelmanns/tfops-aug
augmentation-libraries augmentation-methods augmentation-operation augmentation-policy google-autoaugment keras tf2
Last synced: 3 months ago
JSON representation
TFOps-Aug: Implementation of policy-based image augmentation techniques based on TF2 Operations. All augmentations as efficient Tensorflow 2.11.0 operations. Easy integration into a tf.data API pipeline.
- Host: GitHub
- URL: https://github.com/tillbeemelmanns/tfops-aug
- Owner: TillBeemelmanns
- License: mit
- Created: 2020-04-20T20:26:33.000Z (almost 5 years ago)
- Default Branch: master
- Last Pushed: 2022-12-12T18:58:04.000Z (about 2 years ago)
- Last Synced: 2024-11-15T03:44:44.244Z (3 months ago)
- Topics: augmentation-libraries, augmentation-methods, augmentation-operation, augmentation-policy, google-autoaugment, keras, tf2
- Language: Python
- Homepage:
- Size: 55.2 MB
- Stars: 14
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# TFOps-Aug: Implementation of policy-based image augmentation techniques based on TF2 Operations
The principle for the augmentation mechanism of __TFOps-Aug__ relies on Google's AutoAugment paper "Learning
Augmentation Policies from Data" [1]. This repository implements the augmentation policy logic and many of the
augmentation functions. Like in the original implementation, the augmentation procedure is defined as a policy,
which consists of several subpolicies and operations.
The augmentation operations rely on Tensorflow 2 operations which allow scalability and high computational throughput
even with large images. Furthermore, the augmentation pipeline can be easily integrated into the `tf.data` API, because
all operations rely on Tensorflow operations and can be execute on image representations of type `tf.Tensor`. Currently,
only image representations of type `tf.uint8` are supported.
### Installation
The package is available on [pypi.org](https://pypi.org/project/tfops-aug/) and can be installed with `pip`:
```bash
pip install tfops-aug
```
### Example for an augmentation policy
```python
policy = {'sub_policy0': {'op0': ['adjust_saturation', 0.2, 2],
'op1': ['equalize', 0.1, 6],
'op2': ['add_noise', 0.9, 6]},
'sub_policy1': {'op0': ['adjust_contrast', 0.1, 7],
'op1': ['add_noise', 0.0, 10]},
'sub_policy2': {'op0': ['posterize', 0.9, 6],
'op1': ['unbiased_gamma_sampling', 0.5, 1]},
'sub_policy3': {'op0': ['adjust_brightness', 0.3, 1],
'op1': ['adjust_hue', 0.4, 5]},
'sub_policy4': {'op0': ['adjust_saturation', 0.2, 9],
'op1': ['add_noise', 0.1, 0]},
'sub_policy5': {'op0': ['adjust_contrast', 1.0, 1],
'op1': ['unbiased_gamma_sampling', 0.4, 9]},
'sub_policy6': {'op0': ['unbiased_gamma_sampling', 0.3, 0],
'op1': ['adjust_hue', 0.1, 6]},
'sub_policy7': {'op0': ['solarize', 0.6, 0],
'op1': ['adjust_gamma', 0.3, 6]},
'sub_policy8': {'op0': ['adjust_jpeg_quality', 0.7, 10],
'op1': ['adjust_hue', 0.1, 2]},
'sub_policy9': {'op0': ['equalize', 0.6, 0],
'op1': ['solarize', 0.0, 6]}}
```
Similar to Google's AutoAugment, a single augmentation policy consists of several subpolicies, which in turn consists of
one or more augmentation operation. Each operation is defined as a tuple of **augmentation method**,
**probability** and **intensity**. Several operations within one subpolicy are applied in sequence.
The augmentation policy from above would augment the [original image](assets/test_image.jpg) to the following output:
### Usage
A full example script for image classification can be found in [classification_example.py](utils/classification_example.py).
This excerpt demonstrates the simplicity for the usage of the augmentation methods:
```python
import tensorflow as tf
from tfops_aug.augmentation_utils import apply_augmentation_policy
def augmentor_func(img, label):
img = apply_augmentation_policy(img, policy)
return img, label
train_dataset = tf.keras.preprocessing.image_dataset_from_directory(
"PetImages",
subset="training",
image_size=(180, 180),
batch_size=1
).unbatch()
train_dataset = train_dataset.map(augmentor_func).batch(32).prefetch(32)
```
### Augmentation Methods
A list of all implemented augmentation techniques is given here. Additional, methods will be implemented in the near
future. Performance is measured with the `test_image.jpg` which has size `2048 x 1024`. All augmentation methods are
executed with `level=5`. Averaged over 500 samples on the Intel Core i7 Prozessor 8665U.
| Augmentation | Image | Performance (per Image) |
|----------|:-------------|:-----------------------:|
| Additive Gaussian Noise |  | 0.02 s |
| Adjust Brightness |  | 0.01 s |
| Adjust Contrast |  | 0.02 s |
| Adjust Gamma |  | 0.02 s |
| Adjust Hue |  | 0.01 s |
| Adjust JPEG Quality |  | 0.035 s |
| Adjust Saturation |  | 0.02 s |
| Histogramm Equalization |  | 0.04 s |
| Invert |  | 0.01 s |
| Posterize |  | 0.02 s |
| Solarize |  | 0.01 s |
| Unbiased Gamma Sampling |  | 0.02 s |
| Gaussian Blur |  | 0.14 s |
| Sharpen |  | 0.05 s |
| Shear X |  | 0.04 s |
| Shear Y |  | 0.04 s |
| Translate X |  | 0.05 s |
| Translate Y |  | 0.05 s |
### Reference
```
[1] AutoAugment: Learning Augmentation Policies from Data - 2019
Ekin Dogus Cubuk and Barret Zoph and Dandelion Mane and Vijay Vasudevan and Quoc V. Le
https://arxiv.org/pdf/1805.09501.pdf
```
### TODO
- [ ] More Augmentation Methods
- [X] Shear X
- [X] Shear Y
- [X] Translate X
- [X] Translate Y
- [ ] Random Translation
- [ ] Random Rotation
- [ ] Make augmentation min and max values configurable
- [X] Implement Learning Pipeline
- [ ] Implement augmentation policies identical to these in [1]
- [ ] Implement augmentation policy search with Ray Tune
- [ ] Clean up Code (Unified Docstrings)
- [X] Create Python package
- [ ] Support image representation types, other than `uint8`