Skip to content

Latest commit

 

History

History
513 lines (338 loc) · 24.2 KB

README.md

File metadata and controls

513 lines (338 loc) · 24.2 KB
Raw

DOI

Deep Neural Networks for Active Wave Breaking Classification

This repository contains code and data to reproduce the results of the paper Deep Neural Networks for Active Wave Breaking Classification published by Scientific Reports.

Contents

Deep Neural Networks for Active Wave Breaking Classification

1. Dependencies

Using conda:

  • Windows
conda env create -f environment_win.yml
  • Linux
conda env create -f environment_linux.yml

Manually

Package list 
# install GIT
conda install git

# create a new environment
conda create --name tf python=3

# activate your new environment
conda activate tf

# If you have a nvidia GPU installed and properly configured
pip install --upgrade pip
pip install tensorflow

# extras
pip install -q git+https://github.com/tensorflow/examples.git
pip install tensorflow_addons

# conda packages
conda install -y numba natsort pandas scikit-learn scikit-image natsort  matplotlib seaborn netCDF4 xarray ipython tqdm
pip install h5netcdf

# Extra thresholding methods
pip install pythreshold

# fitting circles to data
pip install miniball

# parallel computations
pip install pebble

2. Data

2.1. Manual Creation

2.2. Production Ready

Model Link Alternative link
Train (10k) -
Train (20k) -
Test (1k) -
Black Sea (200k) Upcoming -
La Jument 2019 (10k) Upcoming -

3. Training

Open In Colab | Jupyter Notebook

Note: The training dataset used here is a smaller version (10k) of the published dataset so it can run on Google Colab. The 20K dataset takes over 6 hours to train and Google will disconnect your session.

The data needs to be in a folder which has sub-folders "0" and "1"

For example:

train
    ├───0
    ├───1

There are 5 backbones implemented: VGG16, ResNet50V2, InceptionResNetV2, MobileNetV2 and EfficientNet

Note that the weights from these pre-trained models will be reset and updated from the scratch here. These models have no knowledge of the present data and, consequently, transferred learning does not work well.

Example

python train.py --data "train/" --backbone "VGG16" --model "vgg_test" --logdir "logs/" --random-state 11 --validation-size 0.2 --learning-rate 0.00001 --epochs 200 --batch-size 64 --dropout 0.5 --input-size 256 256
Arguments:

  • --data Input train data path.

  • --model Model name.

  • --backbone Which backbone to use. See above.

  • --random-state Random seed for reproducibility. Default is 11.

  • --validation-size Size of the validation dataset. Default is 0.2.

  • --epochs Number of epochs (iterations) to train the model. Default is 200.

  • --batch-size Number of images to process in each step. Decrease if running into memory issues. Default is 64.

  • --dropout Droput percentage. Default is 0.5.

  • --input-size Image input size. Decrease if running into memory issues. Default is 256x256px.


The neural network looks something like this:

Pre-trained Models

Please use the links below to download pre-trained models:

Scientific Reports (20K dataset)

Model Link Alternative link
VGG16 -
ResNet50V2 -
InceptionResNetV2 -
MobileNet -
EfficientNet -

La Jument (10K dataset)

Model Link Alternative link
VGG16 Upcoming -
ResNet50V2 Upcoming -
InceptionResNetV2 Upcoming -
MobileNet Upcoming -
EfficientNet Upcoming -

Note: Work in progress.

4. Model Performance

4.1. Evaluating

Open In Colab | Jupyter Notebook

To evaluate a pre-trained model on test data, use the test script.

Example:

python test.py --data "path/to/test/data/" --model "VGG16.h5" --threshold 0.5 -- output "path/to/results.csv"
Arguments:

  • --data Input test data. Use same structure as when training.

  • --model Pre-trained model.

  • --threshold Threshold for binary classification. Default is 0.5

  • --output path to save the results.


The classification report with be printed on the screen. For example:

              precision    recall  f1-score   support

         0.0       0.88      0.99      0.94      1025
         1.0       0.87      0.23      0.36       175

    accuracy                           0.88      1200
   macro avg       0.88      0.61      0.65      1200
weighted avg       0.88      0.88      0.85      1200

To summarize the model metrics do:

python metrics.py --data "path/to/data/" --model "VGG16.h5" --threshold 0.5 -- output "path/to/metrics.csv"

The arguments are the same as above.

The results look something like this:

VGG16 Binary_Accuracy True_Positives False_Positives True_Negatives False_Negatives Precision Recall AUC
Train 0.89 771.00 248.00 5680.00 521.00 0.76 0.60 0.92
Validation 0.87 100.00 19.00 1463.00 222.00 0.84 0.31 0.90
Test 0.88 40.00 6.00 1019.00 135.00 0.87 0.23 0.82

To plot the training curves and a confusion matrix, do:

python plot_history_and_confusion_matrix.py --history "path/to/history.csv" --results "path/to/results.csv" --output "figure.png"
Arguments:

  • --history Training history. Comes from train_wave_breaking_classifier_v2.py.

  • --results Classification results from the test data. Comes from test_wave_breaking_classifier.py.

  • --output Figure name.


The results look like this:

4.2. Results

The table below summarizes the results presented in the paper. Results are sorted by AUC.

Train

Model Accuracy TP FP TN FN Precision Recall AUC
ResNetV250 0.97 1414 198 13978 280 0.877 0.835 0.989
VGG16 0.93 855 273 13911 831 0.758 0.507 0.943
InceptionResnetV2 0.927 886 359 13823 802 0.712 0.525 0.932
EfficientNet 0.772 1403 3346 10920 297 0.295 0.825 0.874
MobileNet 0.904 436 268 13916 1250 0.619 0.259 0.848

Validation

Model Accuracy TP FP TN FN Precision Recall AUC
VGG16 0.932 221 65 3478 204 0.773 0.52 0.946
ResNetV250 0.919 197 97 3450 224 0.67 0.468 0.873
InceptionResnetV2 0.921 190 81 3466 231 0.701 0.451 0.93
EfficientNet 0.809 353 687 2856 72 0.339 0.831 0.897
MobileNet 0.908 123 64 3479 302 0.658 0.289 0.878

Test

Model Accuracy TP FP TN FN Precision Recall AUC
VGG16 0.876 106 80 945 69 0.57 0.606 0.855
ResNetV250 0.881 95 63 962 80 0.601 0.543 0.843
InceptionResnetV2 0.882 91 57 968 84 0.615 0.52 0.839
EfficientNet 0.873 88 65 960 87 0.575 0.503 0.827
MobileNet 0.875 30 5 1020 145 0.857 0.171 0.768

5. Using a Pre-trained Neural Network

5.1 Predicting on New Data

Create a dataset either manually or with the provided tools then use the predict script. The data structure is as follows:

pred
    ├───images
        ├───img_00001.png
        ├───img_00002.png
        ├───...
        ├───img_0000X.png

Example:

python predict.py --data "pred/" --model "VGG16.h5" --threshold 0.5 --output "results.csv"
Arguments:

  • --data Input test data.

  • --model Pre-trained model.

  • --threshold Threshold for binary classification. Default is 0.5

  • --output A csv file with the classification results.


5.2. Predicting from the Results of the Naïve Detector

Use predict from naïve candidates and the results from the naive wave breaking detector and a pre-trained neural network to obtain only active wave breaking instances. This script runs on CPU but can be much faster on GPU.

Example:

python predict_from_naive_candidates.py --debug --input "naive_results.csv" --model "path/to/model.h5" --frames "path/to/frames/folder/"  --region-of-interest "region_of_interest.csv" --output "robust_results.csv" --temporary-path "tmp" --frames-to-plot 1000 --threshold 0.5
Arguments:

  • --debug Runs in debug mode and will save output plots.

  • -i [--input] Input data obtained from naive_wave_breaking_detector.

  • -m [--model] Pre-trained Tensorflow model.

  • -o [--output] Output file name (see below for explanation).

  • -frames [--frames] Input path with images.

  • --region-of-interest File with region of interest. Use minimun bounding geometry to generate a valid input file.

  • -temporary-path Output path for debug plots.

  • --frames-to-process Number of frames to process.

  • --from-frame Start frame.

  • --regex Regular expression to find input frames. Default is "[0-9]{6,}".

  • --threshold Threshold for activation in the last (sigmoid) layer of the model. Default is 0.5.

    Note: The input data must have at least the following entries: ic, jc, ir, and frame.


The output of this script is a comma-separated value (csv) file. It looks like exactly like the output of naive wave breaking detector but adding a extra column with the results of the classification.

5.3. Plotting Wave Breaking Detection Results

Plot the results of the wave breaking detection algorithms. Can handle outputs of any algorithm, as long as the input data is correct. Ideally the results from cluster.py are used as input.

Example:

python plot_wave_breaking_detection_results.py --input "clustered_events.csv" --output "path/to/output/" --frames "path/to/frames/" --region-of-interest "path/to/roi.csv" --frames-to-plot 1000
Arguments:

  • -i [--input] Input csv file.

  • -o [--output] Output path.

  • -frames-path Path with frames.

  • --region-of-interest File with region of interest. Use minimun bounding geometry to generate a valid input file.

  • --frames-to-plot Number of frames to plot.

  • --from-frame Number of frames to plot.

  • --regex Regular expression to find input frames. Default is "[0-9]{6,}".

Note: The input data must have at least the following entries: ic, jc, ir, frame, wave_breaking_event.

6. Model Interpretation

Open In Colab | Jupyter Notebook

Use interpret.py to apply Grad-CAM to data samples. Organize your data as follows:

gradcam
    ├───images
        ├───img_00001.png
        ├───img_00002.png
        ├───...
        ├───img_0000X.png

Example:

python interpret.py --data "path/to/gradcam" --model "VGG16.h5" -o "path/to/output"
Arguments:

  • -data Input image data path.

  • -o [--output] Output path.

  • -model pre-trained VGG16 model.

Note: This script will only work with VVG16 models.


7. Active Wave Breaking Segmentation

The neural networks developed here can also be used for image segmentation. Please refer to Image Segmentation.

8. Wave Tracking

Please refer to Wave Tracking.

9. Wave Breaking Statistics

Please refer to Wave Breaking Statistics.

Gallery

La Jument:

Black Sea:

Aqua Alta:

Grad-CAM (Black Sea):

Image Segmentation (La Jument):

Standard Variable Names

The following variables are standard across this repository and scripts that output these quantities should use these names. If a given script has extra output variables, these are documented in each script.

Variables
Variable Description
x x-coordinate in metric coordinates.
y y-coordinate in metric coordinates.
z z-coordinate in metric coordinates.
time date and time. Use a format that pandas.to_datetime() can understand.
frame sequential number.
i pixel coordinate in pixel units. Use Matplotlib coordinate system.
j pixel coordinate in pixel units. Use Matplotlib coordinate system.
ic center of a circle or ellipse in pixel coordinates.
jc center of a circle or ellipse in pixel coordinates.
xc center of a circle or ellipse in metric coordinates.
yc center of a circle or ellipse in metric coordinates.
ir radius in the i-direction.
jr radius in the j-direction.
xr radius in the x-direction.
yr radius in the y-direction.
theta_ij angle of rotation of an ellipse with respect to the x-axis counter-clockwise.
theta_xy angle of rotation of an ellipse with respect to the x-axis counter-clockwise.
wave_breaking_event unique wave breaking event id.
vx velocity in the x-direction in m/s.
vy velocity in the y-direction in m/s.
vi velocity in the x-direction in pixels/frame.
vj velocity in the y-direction in pixels/frame.

Disclaimer

There is no warranty for the program, to the extent permitted by applicable law except when otherwise stated in writing the copyright holders and/or other parties provide the program “as is” without warranty of any kind, either expressed or implied, including, but not limited to, the implied warranties of merchantability and fitness for a particular purpose. the entire risk as to the quality and performance of the program is with you. should the program prove defective, you assume the cost of all necessary servicing, repair or correction.