The aucmedi from frankkramer-lab

Implement XAI Method: CNN Fixation

Make AUCMEDI 3D avaible

Update subfunction resize, padding & cropping 3d able (check shape in subfunction)
Check if subfunction standardize can transform volumes
Clean up architectures via 2D / 3D
Add 3D architectures
Add numpy loader for volumes/images
Add npy to accepted formats
Parameter in DataGenerator to pass loader() with default to image_loader()
Add 3D augmentation volumenations (can we just switch ai module depending on 2D/3D?)
Modify ensembler.augmenting function
Add 3D unittesting for ensembler.augmenting
Add volume support in unittesting
Rework DataGenerator to allow better image preparation for large 3D volumes

XAI shouldn't be able to run on 3D :/

Add sampling functions

For binary, multi-class and multi-label

Normal train,test split
Cross-Validation split

sampling seed auf 0 setzen anstatt None bei den parametern

Should all seeds be set to 0 as default parameter or random (None)?

Current implementation is None.

Implement XAI Method: Grad-Cam

https://ieeexplore.ieee.org/document/8237336

Write documentation

Update README
Create wiki
Create some tutorials
Create example

Data Loading with different file structures

/images/...
classes.json/.csv

or

/images/class_a/...
/images/class_b/...
/images/class_c/...

Ensemble Learning OOM

Delete model at the end for ensemble learning stacking or bagging approaches

Implement XAI Method: Grad-Cam++

https://github.com/samson6460/tf_keras_gradcamplusplus/blob/master/gradcam.py
https://github.com/totti0223/gradcamplusplus/blob/master/gradcamutils.py

https://ieeexplore.ieee.org/document/8354201

Don't forget temporary directory removement for prepare_batches mode in pipeline

Add sample_weight computation for multi-label

For each class: c
-> Compute n_samples / (n_classes * n_occurences_for_c)

compute sample weights by multiplying weights together for each sample

Check out this:
https://scikit-learn.org/stable/modules/generated/sklearn.utils.class_weight.compute_sample_weight.html

Implement XAI Method: Integrated Gradient

https://github.com/experiencor/deep-viz-keras
https://keras.io/examples/vision/integrated_gradients/
https://github.com/hiranumn/IntegratedGradients

https://arxiv.org/abs/1703.01365

Implement XAI Method: Guided Grad-Cam

Activations Visualization

Increase default number of epochs for transfer learning first training process

Resizing Speed

Implement cv2 resize and check speed comparison

Have a look in scale augmentations

Does 0.9 to 1.1 scaling means doubling?

4k x 4k images resulted in 8k x 8k?

Inference Augmentation: Techniques

Method:

Init Img Augmentation
Load testing data
Modify x_test by multiplying each index by N (indicies should be in order -> [a, a, a, b, b, b, c, c, c ...] for N=3)
Provide to DataGenerator with ImgAugmentation
Run prediction
Ensemble inference

Implement multiprocessing option for in model train

Test batch generation speed with workers

Configuration

Yaml file with configs
Runner: main.py
Run modi: Training, Inference, Complete/FullAnalysis

Neural Network load model

Delete last model before overwriting when loading a new model

Problem: Tensorflow caches various model variables which can lead to OOM when switching multiple times between different models.

Solution: clear session when loading a new model.
tf.keras.backend.clear_session()

Note for me: clear_session probably not working when using multiple model variables!!!

Note for Ensemble Learning: Delete models at the end to avoid this issue

Subfunctions: Color Constancy

Test aucmedi input_interface with grayscale and rgb images

https://github.com/sicara/tf-explain

cv2 error in google colab?

Das war die Fehlermeldung, die ich bekommen habe, welche ich dann mit dem deinstallieren und neu installieren von opencv_python gelöst habe:

---------------------------------------------------------------------------

ImportError                               Traceback (most recent call last)

<ipython-input-12-8cde4f83d73a> in <module>()
----> 1 from aucmedi.data_processing import io_data
      2 dataset_loader = io_data.input_interface("directory", path_imagedir="/content/Tuberculosis_Classification", path_data=None, training=True, ohe=False)

11 frames

/usr/local/lib/python3.7/dist-packages/cv2/__init__.py in <module>()
      7
      8 from .cv2 import *
----> 9 from .cv2 import _registerMatType
     10 from . import mat_wrapper
     11 from . import gapi

ImportError: cannot import name '_registerMatType' from 'cv2.cv2' (/usr/local/lib/python3.7/dist-packages/cv2/cv2.cpython-37m-x86_64-linux-gnu.so)


---------------------------------------------------------------------------
NOTE: If your import is failing due to a missing package, you can
manually install dependencies using either !pip or !apt.

To view examples of installing some common dependencies, click the
"Open Examples" button below.