In order to train and validate deep learning models using torch, input data must be converted into tensors (i.e., multidimensional arrays) with the correct dimensionality and data types. These tensors need to then be provided to the algorithm as mini-batches, since algorithms are trained incrementally on subsets of the available training set. During the training loop, data within each training mini-batch will be predicted. These predictions will then be compared to the reference labels to calculate a loss (and, optionally, additional assessment metrics). Once the loss is calculated, backpropagation is performed and the model parameters are updated using an optimization algorithm, such as stochastic gradient descent, RMSProp, or Adam.
In the torch environment, how a single image chip is processed is defined by a DataSet subclass. Specifically, the .getitem() method defines how a single input sample is processed to a tensor. The DataLoader is responsible for using the DataSet subclass to generate a mini-batch of samples to pass to the algorithm during the training or validation process.
DataSets and DataLoaders are used for preparing training, validation, and testing datasets. As a result, they are an important component of the torch workflow. geodl provides the defineSegDataSet() function for processing input data. An instantiated DataSet generated with this function can then be passed to torch’s default DataLoader class.
The defineSegDataSet() function from geodl instantiates a subclass of torch::dataset() for geospatial semantic segmentation. This function is specifically designed to load data generated using the makeChips() or makeChipsMultiClass() function. It accepts a data frame created with the makeChipsDF() function.
It can also define random augmentations to combat overfitting. Note that horizontal and vertical flips will impact the alignment of the image chip and associated mask. As a result, the same augmentation will be applied to both the image and the mask. Changes in brightness, contrast, gamma, hue, and saturation are not applied to the masks since alignment is not impacted by these transformations. Predictor variables are generated with three dimensions (channel/variable, width, height) regardless of the number of channels/variables. Masks are generated as three dimensional tensors (class index, width, height).
In this example, we step through how to prepare data for input into a modeling workflow. The general workflow is as follows:
Since the processes of creating masks and image chips are described in other articles, we will begin here with creating the chips data frame. The makeChipsDF() function requires (1) an input folder path where the chips are saved and (2) a mode (either “All”, “Divided”, or “Positive”). In our example, we are using the “Positive” mode so that only chips with at least one pixel mapped to the positive class are listed. You can also choose to save the data frame to a CSV file and write it to disk and shuffle the rows to reduce autocorrelation. In our example, we are not writing the data frame out to a CSV file and not shuffling the rows. Shuffling is not necessary in this case because the data are already randomly sequenced.
You can use the head() function to view the first few rows of the table.
chpDF <- makeChipsDF(folder = "C:/myFiles/data/toChipBinary/chips/",
outCSV = "C:/myFiles/data/toChipBinary/chips/chipsDF.csv",
extension = ".tif",
mode="Positive",
shuffle=FALSE,
saveCSV=FALSE)
head(chpDF)
chpN chpPth mskPth
1 topoOut_1_1.tif images/topoOut_1_1.tif masks/topoOut_1_1.tif
2 topoOut_1_1793.tif images/topoOut_1_1793.tif masks/topoOut_1_1793.tif
3 topoOut_1_2049.tif images/topoOut_1_2049.tif masks/topoOut_1_2049.tif
4 topoOut_1_2305.tif images/topoOut_1_2305.tif masks/topoOut_1_2305.tif
5 topoOut_1_257.tif images/topoOut_1_257.tif masks/topoOut_1_257.tif
6 topoOut_1_2817.tif images/topoOut_1_2817.tif masks/topoOut_1_2817.tif
Once the chips have been listed into a data frame, they can be described with the describeChips() function. This can be useful for checking for issues, obtaining statistical for data normalization, or estimating class proportions to implement weighted loss functions. To speed up the calculations, it is recommend to use a random subset of chips and a random subset of pixels per chip.
chpDescript <- describeChips(folder= "C:/myFiles/data/toChipBinary/chips/",
extension = ".tif",
mode = "Positive",
subSample = TRUE,
numChips = 100,
numChipsBack = 100,
subSamplePix = TRUE,
sampsPerChip = 400)
print(chpDescript)
$ImageStats
vars n mean sd median trimmed mad min max range skew kurtosis se
B1 1 40000 213.95 34.14 219 217.71 32.62 0 255 255 -1.20 2.46 0.17
B2 2 40000 205.13 51.06 226 212.01 41.51 0 255 255 -0.95 -0.06 0.26
B3 3 40000 162.87 51.13 169 163.69 41.51 0 255 255 -0.21 -0.29 0.26
$mskStats
# A tibble: 2 × 2
value cnt
<dbl> <dbl>
1 0 5366555
2 1 1187045
Once the chips have been listed into a data frame and described, you are ready to define a DataSet using defineSegDataSet(). This function has many parameters, many of which relate to data augmentations. Here is a description of the parameters:
In our example, the data are being normalized, and we have provided band means and standard deviations, which were estimated using the desribeChips() function. We are also rescaling the data by a factor of 255. This is because the data are currently represented as 8-bit integer values as opposed to 32-bit float values scaled from 0 to 1.
Augmentations are also being performed. A maximum of 2 augmentations will be applied per chip. Vertical flips, horizontal flips, brightness changes, and saturation changes will be considered. The flips have a 30% probability of being applied while the brightness and saturation changes have a 10% chance of being applied. We often struggle with determining the appropriate amount of data augmentations to apply. We tend to conservatively apply augmentations; however, others may disagree.
Not all augmentations are appropriate for all data types. For example, saturation and hue adjustments are only applicable to RGB data.
myDS <- defineSegDataSet(chpDF,
folder="C:/myFiles/data/toChipBinary/chips/",
normalize = TRUE,
rescaleFactor = 255,
mskRescale=1,
mskAdd=1,
bands = c(1,2,3),
bMns=c(214,206,163),
bSDs=c(33,50,51),
doAugs = TRUE,
maxAugs = 2,
probVFlip = .3,
probHFlip = .3,
probBrightness = .1,
probContrast = 0,
probGamma = 0,
probHue = 0,
probSaturation = .1,
brightFactor = c(.8,1.2),
contrastFactor = c(.8,1.2),
gammaFactor = c(.8, 1.2, 1),
hueFactor = c(-.2, .2),
saturationFactor = c(.8, 1.2))
Once the DataSet is defined, it can be fed to the dataloader() function from torch. Again, the purpose of the DataLoader is to generate mini-batches of data during the training or inference processes. In our example, a mini-batch size of 4 is being used. The maximum possible mini-batch size will depend on several factors including (1) the complexity of the model, (2) the number of rows and columns of pixels in each chip, (3) the number of channels in each chip, and (4) the available hardware. Training on multiple GPUs allows for using a larger mini-batch size. However, studies have shown that using a large mini-batch size can result in overfitting or more sensitivity to changes in the learning rate. As a result, we generally use smaller mini-batch sizes.
In our example, we are also randomly shuffling the data in an attempt to further reduce autocorrelation. The last mini-batch is being dropped. This is because, unless the number of available chips is perfectly divisible by the mini-batch size, the last mini-batch will be smaller than the other mini-batches. This can result in issues with loss and assessment metric calculations. So, we generally recommend dropping the last mini-batch unless the total number of chips is evenly divisible by the mini-batch size.
Once a DataSet and DataLoader have been instantiated, you are ready to define and initialize a training loop. This process will be explored in other articles.
Optionally, you can check your data by (1) plotting a mini-batch of chips and associated masks with the viewBatch() function and/or (2) obtaining summary info for a single mini-batch using the describeBatch() function. We generally recommend running these checks. We recommend considering the following as you review the data:
viewBatch(dataLoader=myDL,
nCols = 2,
r = 1,
g = 2,
b = 3,
cNames=c("Background", "Mine"),
cColor=c("#D4C2AD","#BA8E7A")
)
print(myBatchStats)
$batchSize
[1] 4
$imageDataType
[1] "Float"
$maskDataType
[1] "Long"
$imageShape
[1] "4" "3" "256" "256"
$maskShape
[1] "4" "1" "256" "256"
$bndMns
[1] -0.0004852064 -0.0005571313 -0.0006161333
$bandSDs
[1] 0.004172502 0.004146057 0.003736094
$maskCount
[1] 200514 61630
$minIndex
[1] 1
$maxIndex
[1] 2