Metadata-Version: 2.1
Name: tfops_aug
Version: 0.0.2
Summary: Package to augment images using Tensorflow Ops
Project-URL: Homepage, https://github.com/TillBeemelmanns/auto-augment-tf2-operations
Project-URL: Bug Tracker, https://github.com/TillBeemelmanns/auto-augment-tf2-operations/issues
Author-email: Till Beemelmanns <till.beemelmanns@rwth-aachen.de>
License-File: LICENSE
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: OS Independent
Classifier: Programming Language :: Python :: 3
Requires-Python: >=3.9
Requires-Dist: matplotlib
Requires-Dist: numpy
Requires-Dist: opencv-python
Requires-Dist: tensorflow-addons==0.19.0
Description-Content-Type: text/markdown

# 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:
 
![](assets/augmentation_policy.gif)

### 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 | ![](assets/add_noise.gif) |         0.02 s          |
| Adjust Brightness | ![](assets/adjust_brightness.gif) |         0.01 s          |
| Adjust Contrast | ![](assets/adjust_contrast.gif) |         0.02 s          |
| Adjust Gamma | ![](assets/adjust_gamma.gif) |         0.02 s          |
| Adjust Hue | ![](assets/adjust_hue.gif) |         0.01 s          |
| Adjust JPEG Quality | ![](assets/adjust_jpeg_quality.gif) |         0.035 s         |
| Adjust Saturation | ![](assets/adjust_saturation.gif) |         0.02 s          |
| Histogramm Equalization | ![](assets/equalize.gif) |         0.04 s          |
| Invert | ![](assets/invert.gif) |         0.01 s          |
| Posterize | ![](assets/posterize.gif) |         0.02 s          |
| Solarize | ![](assets/solarize.gif) |         0.01 s          |
| Unbiased Gamma Sampling | ![](assets/unbiased_gamma_sampling.gif) |         0.02 s          |
| Gaussian Blur | ![](assets/gaussian_blur.gif) |         0.14 s          |
| Sharpen | ![](assets/sharpen.gif) |         0.05 s          |
| Shear X | ![](assets/shear_x.gif) |         0.04 s          |
| Shear Y | ![](assets/shear_y.gif) |         0.04 s          |
| Translate X | ![](assets/translate_x.gif) |         0.05 s          |
| Translate Y | ![](assets/translate_y.gif) |         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`
