Skip to content

Data and code to implement Buscombe et al (2019) optical wave gauging using deep neural networks

Notifications You must be signed in to change notification settings

OpticalWaveGauging/OpticalWaveGauging_DNN

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

37 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

About

Data and code to implement optical wave gauging (OWG) using deep neural networks, detailed in the paper Buscombe et al (2019):

Buscombe, Carini, Harrison, Chickadel, and Warrick (2019) Optical wave gauging using deep neural networks. Coastal Engineering https://doi.org/10.1016/j.coastaleng.2019.103593

Software and data for training deep convolutional neural network models to estimate wave height and wave period from surf zone imagery

This software was tested on Windows 10 and Ubuntu Linux with python 3.7, tensorflow 2. This software was written by Dr Daniel Buscombe at Northern Arizona University, 2018-2019.

How to use this software

You can run this software in two ways:

Either:

  1. Run the provided jupyter notebooks (with .ipynb extension) on Google Colab (a free cloud computing service from Google that provides a runtime for executing notebooks, including GPU support). If you are keen to see what this tool looks like, the two provided juypyter notebooks are available through these links: model training and model testing

or

  1. Set up an anaconda python on your GPU-enabled computer and run the provided python scripts

THIS SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND. IF YOU ENCOUNTER A PROBLEM/BUG OR HAVE A QUESTION OR SUGGESTION, PLEASE USE THE "ISSUES" TAB ON GITHUB. OTHERWISE, THIS SOFTWARE IS UNSUPPORTED.

Folder structure

  • \conda_env contains yml files for setting up a conda environment
  • \config contains the configuration file with user-definable settings
  • \train contains files using for training models
  • \im128 is a file structure that will contain results from model training

Setting up computing environments

Install Anaconda python distribution

Install the latest version of Anaconda (https://www.anaconda.com/distribution/)

When installed, open an anaconda command prompt

Get this github repository

Use git to clone the github directory

git clone --depth 1 [email protected]:dbuscombe-usgs/OpticalWaveGauging_DNN.git

navigate to the OpticalWaveGauging_DNN directory

cd OpticalWaveGauging_DNN

It is strongly recommended that you use a GPU-enabled tensorflow installation. CPU training of a model can take several hours to several days (more likely the latter). However, the following instructions are for a CPU install. To use gpu, replace tensorflow with tensorflow-gpu in conda_env/owg.yml

First, if you are a regular conda user, I would recommend some conda housekeeping (this might take a while):

conda clean --packages
conda update -n base conda

Otherwise (i.e. this is a fresh conda install), no housekeeping required.

Create a conda virtual environment

Create a new conda environment called owg

conda env create -f conda_env/owg.yml

(If you get an error related to installing tensorflow, replace tensorflow-gpu==2.0 with tensorflow-gpu or simply the CPU version tensorflow )

This takes a few minutes. When it is done, activate environment:

conda activate owg

A note on versions and releases

This is an evolving project and things move fast in the world of deep learning. During the 8 months the Coastal Engineering paper associated with this repository was in review, Tensorflow and Keras libraries underwent some major changes. Tensorflow did a major upgrade from version 1 to version 2, and keras got subsumed as into Tensorflow as tf.keras.

This is the version of the software that worked in October 2019 prior to the official release of Tensorflow 2, and may be used to exactly reproduce the results of the paper, if you can get it to work using tensorflow 1.X. The release version "11.11.19" will not work in tensorflow 2.X.

Later releases, including this release, work in Tensorflow 2 using tf.keras

Setting up the model

Configuration files are in JSON format, like this:

{
  "samplewise_std_normalization" : true,
  "samplewise_center"  : true,
  "input_image_format" : "jpg",
  "input_csv_file"     : "snap-training-dataset.csv", 
  "category"           : "H",
  "prc_lower_withheld": 5,
  "prc_upper_withheld": 5,
  
  "horizontal_flip"    : false,
  "vertical_flip"      : false,
  "rotation_range"     : 10,
  "width_shift_range"  : 0.1,
  "height_shift_range" : 0.1,
  "shear_range"        : 0.05,
  "zoom_range"         : 0.2,
  "fill_mode"          : "reflect",
  
  "img_size"           : 128,
  "num_epochs"         : 5,
  "test_size"          : 0.4,
  "dropout_rate"       : 0.5,
  "epsilon"            : 0.0001,
  "min_lr"             : 0.0001,
  "factor"             : 0.8
}

Training inputs

  • imsize : size of image to use (pixels)
  • category: 'H' for wave height, 'T' for wave period
  • input_image_format: image file extension
  • input_csv_file: name of file that has wave height and period per image

Model hyperparameters

  • num_epochs = number of training epochs
  • test_size = proportion of data set to use for training
  • dropout_rate: proportion of neurons to randomly drop in dropout layer
  • factor: factor by which the learning rate will be reduced. new_lr = lr * factor
  • epsilon: threshold for measuring the new optimum, to only focus on significant changes.
  • min_lr: lower bound on the learning rate.

Image pre-processing

  • samplewise_std_normalization: Bool. if True, Divide each input by its std.
  • samplewise_center: Bool. if True, set each sample mean to 0.

Image augmentation parameters:

  • rotation_range: Int. Degree range for random rotations.
  • width_shift_range: Float, 1-D array-like or int float: fraction of total width, if < 1, or pixels if >= 1. 1-D array-like: random elements from the array. int: integer number of pixels from interval (-width_shift_range, +width_shift_range) With width_shift_range=2 possible values are integers [-1, 0, +1], same as with width_shift_range=[-1, 0, +1], while with width_shift_range=1.0 possible values are floats in the half-open interval [-1.0, +1.0[.
  • height_shift_range: Float, 1-D array-like or int float: fraction of total height, if < 1, or pixels if >= 1. 1-D array-like: random elements from the array. int: integer number of pixels from interval (-height_shift_range, +height_shift_range) With height_shift_range=2 possible values are integers [-1, 0, +1], same as with height_shift_range=[-1, 0, +1], while with height_shift_range=1.0 possible values are floats in the half-open interval [-1.0, +1.0[.
  • brightness_range: Tuple or list of two floats. Range for picking a brightness shift value from.
  • shear_range: Float. Shear Intensity (Shear angle in counter-clockwise direction in degrees)
  • zoom_range: Float or [lower, upper]. Range for random zoom. If a float, [lower, upper] = [1-zoom_range, 1+zoom_range].
  • fill_mode: One of {"constant", "nearest", "reflect" or "wrap"}.
  • horizontal_flip: Boolean. Randomly flip inputs horizontally.
  • vertical_flip: Boolean. Randomly flip inputs vertically.

Training models

(On Google Colab, run OWG_train.ipynb)

If you do not wish to retrain all the OWGs yourself from scratch, MobileNetV1-based OWGs (only) are provided. See im128/res/200epoch/H/model1 for wave height and im128/res/200epoch/T/model1 for wave period. The hdf5 format files contain the model's weights. The .json format files contain the model architecture. The image files show training outputs and are automatically generated by python train_OWG.py, described below.

To train models yourself (recommended), the following scripts will do so for all combinations of 4 models (MobileNetV1, DenseNet201, InceptionV3, and InceptionResnet2), and 4 batch sizes (16, 32, 64, and 128 images).

Note that in this Tensorflow 2 version of the repository, DenseNet201 is used instead of MobileNetV2, which was used in the paper and the original repository (see release version 11.11.18) but is not implemented correctly at the time of writing using keras in Tensorflow 2 without modifying the MobileNetV2 source script. Therefore DenseNet201 is used instead, which gives similar results to MobileNetV2

python train_OWG.py -c configfile.json

In the above, configfile.json is one of the config files in the . Just provide the name of the json file, including the 'json' file extension, not the full path to the file, like this:

python train_OWG.py -c config_nearshore_H.json

The best models are obtained using larger numbers of epochs (say, 100+), but you'll probably want to train them on a GPU (install tensorflow-gpu instead of tensorflow).

To train OWGs for wave period, change the category in the config file to 'T' and run the above again

Tidying up

Model result files (*.hdf5 format, which contain the model's weights) are organized in the following file structure

im128

---res

------100epoch

---------H

------------model1

---------------batch16

---------------batch32

---------------batch64

---------------batch128

------------model2

---------------batch16

---------------batch32

---------------batch64

---------------batch128

------------model3

---------------batch16

---------------batch32

---------------batch64

---------------batch128

------------model4

---------------batch16

---------------batch32

---------------batch64

---------------batch128

---------T

------------model1

---------------batch16

---------------batch32

---------------batch64

---------------batch128

------------model2

---------------batch16

---------------batch32

---------------batch64

---------------batch128

------------model3

---------------batch16

---------------batch32

---------------batch64

---------------batch128

------------model4

---------------batch16

---------------batch32

---------------batch64

---------------batch128

Finally, compile and plot results from all models using

python compile_results.py

Operational Mode

Testing model on a folder of images

(On Google Colab, run Estimate_per_image_in_folder.ipynb)

python test_OWG_folder.py -i snap_images/data

This program will read the configuration file, conf/config_test.json

This file should be set up with the following information:

{
  "samplewise_std_normalization" : true,
  "samplewise_center"  : true,
  "weights_path" : "im128/res/100epoch/H/model1/batch16/waveheight_weights_model1_16batch.best.nearshore.hdf5",
  "input_csv_file"     : "train/snap-training-dataset.csv", 
  "category"           : "H",
  "im_size"            : 128,
  "image_direc"        : "train/snap_images"
}

where the image_direc is the folder where the test set of images are; weights_path is the hdf5 file associated with the model you wish to use; and input_csv_file should be a comma delimited file like the one used to train with.

The other variables, im_size, category, samplewise_std_normalization and samplewise_center are the same as used in model training

Making a time-series plot from a folder of images

(On Google Colab, not available yet)

python gen_ts_from_folder.py -i snap_images/data

This program will read the configuration file, conf/config_test.json, load the model, predict the quantity of interest, and make two plots. The first plot, short_time_series shows just the first 100 records in the time series and the other time_series plot shows the whole time-series, observed versus estimated overlain.

Testing model on a single image

(On Google Colab, not available yet)

python predict_image.py -i path/to/image/file.ext

for example:

python predict_image.py -i snap_images/data/1516401000.cx.snap.jpg

The following variables are also read from the conf/config_test.json file: weights_path, im_size, category, samplewise_std_normalization and samplewise_center. These should be the same as used in model training

Forthcoming: ensemble model prediction

Next update ...

Wrapping up

Deactivate environment when finished:

conda deactivate