Preparation
The data originators have provides TXT files listing the image chips into separate training, validation, and testing datasets. So, we first read in these lists as data frames. The data lists are not in the correct format for geodl’s defineSegDataset() function, which requires that the following columns are present: “chpN”, “chpPth”, and “mskPth”. The “chpN” column must provides the name of the chip, and the “chpPth” and “MskPth” columns must provide the path to the image and associated mask chips relative to the folder that houses the chips. Using dplyr, we create data frames in this correct format from the provided tables. The mask files have the suffix “_m” added, so we also must add this suffix to the image name when defining the associated mask.
Lastly, we randomly select out 3,000 training, 500 validation, and 500 testing chips to speed up the training and inference processes for this demonstration.
testDF <- read.csv("C:/myFiles/data/landcoverai/test.txt", header=FALSE)
trainDF <- read.csv("C:/myFiles/data/landcoverai/train.txt", header=FALSE)
valDF <- read.csv("C:/myFiles/data/landcoverai/val.txt", header=FALSE)trainDF <- data.frame(chpN=trainDF$V1,
chpPth=paste0("images/", trainDF$V1, ".tif"),
mskPth=paste0("masks/", trainDF$V1, "_m.tif")) |>
sample_frac(1, replace=FALSE) |> sample_n(3000)
testDF <- data.frame(chpN=testDF$V1,
chpPth=paste0("images/", testDF$V1, ".tif"),
mskPth=paste0("masks/", testDF$V1, "_m.tif")) |>
sample_frac(1, replace=FALSE) |> sample_n(500)
valDF <- data.frame(chpN=valDF$V1,
chpPth=paste0("images/", valDF$V1, ".tif"),
mskPth=paste0("masks/", valDF$V1, "_m.tif")) |>
sample_frac(1, replace=FALSE) |> sample_n(500)As a check, I next view a randomly selected set of 25 chips using the viewChips() function.
viewChips(chpDF=trainDF,
folder="C:/myFiles/data/landcoverai/train/",
nSamps = 25,
mode = "both",
justPositive = FALSE,
cCnt = 5,
rCnt = 5,
r = 1,
g = 2,
b = 3,
rescale = FALSE,
rescaleVal = 1,
cNames= c("Background", "Buildings", "Woodland", "Water", "Road"),
cColors= c("gray", "red", "green", "blue", "black"),
useSeed = TRUE,
seed = 42)
I am now ready to define the training and validation datasets. Here are a few key points:
- I am not normalizing the data, so do not need to provide band means and standard deviations.
- The images are scaled from 0 to 255 (8-bit), so I rescale the data from 0 to 1 using a rescaleFactor() of 255. The masks do not need to be rescaled. I also set mskAdd to 1 so that the class codes will begin at 1 instead of 0.
- Lastly, I apply a maximum of 1 augmentation per chip. The augmentations used are random vertical flips, horizontal flips, brightness changes, and saturation changes. Horizontal and vertical flips have a 50% chance of being randomly applied, brightness changes have a 10% chance of being applied, and saturation changes have a 20% chance of being applied.
I use the same settings for the training and validation datasets other than not applying random augmentations for the validation data.
trainDS <- defineSegDataSet(
chpDF=trainDF,
folder="C:/myFiles/data/landcoverai/train/",
normalize = FALSE,
rescaleFactor = 255,
mskRescale= 1,
bands = c(1,2,3),
mskAdd=1,
doAugs = TRUE,
maxAugs = 1,
probVFlip = .5,
probHFlip = .5,
probBrightness = .1,
probContrast = 0,
probGamma = 0,
probHue = 0,
probSaturation = .2,
brightFactor = c(.9,1.1),
contrastFactor = c(.9,1.1),
gammaFactor = c(.9, 1.1, 1),
hueFactor = c(-.1, .1),
saturationFactor = c(.9, 1.1))
valDS <- defineSegDataSet(
chpDF=valDF,
folder="C:/myFiles/data/landcoverai/val/",
normalize = FALSE,
rescaleFactor = 255,
mskRescale = 1,
mskAdd=1,
bands = c(1,2,3),
doAugs = FALSE,
maxAugs = 0,
probVFlip = 0,
probHFlip = 0)Next, a print the length of the datasets to make sure the number of samples are as expected.
length(trainDS)
[1] 3000
length(valDS)
[1] 500Now that I have datasets defined, I generate DataLoaders using the dataloader() function from torch. I use a mini-batch size of 15. You may need to change the mini-batch size depending on your computer’s hardware. The training data are shuffled to reduce autocorrelation; however, the validation data are not. I drop the last mini-batch for both the training and validation data.
trainDL <- torch::dataloader(trainDS,
batch_size=15,
shuffle=TRUE,
drop_last = TRUE)
valDL <- torch::dataloader(valDS,
batch_size=15,
shuffle=FALSE,
drop_last = TRUE)As checks, I next view a batch of the training and validation data using the viewBatch() function. I also use describeBatch() to obtain summary info for a mini-batch. Here are a few points to consider.
- Each mini-batch of images should have a shape of (Mini-Batch, Channels, Width, Height).
- Each mini-batch of masks should have a shape of (Mini-Batch, Class, Width, Height).
- Each image should have a shape of (Channels, Width, Height) and a 32-bit float data type,
- Each mask should have a shape of (Class, Width, Height) and have a long integer data type.
- The number of channels and rows and columns of pixels should match the data being used.
- If you specified a subset of bands, the number of channels should match your subset.
- The range of values in the image bands should be as expected, such as 0 to 1.
- The range of class indices should be as expected. Note whether class codes start at 0 or 1.
- Viewing a mini-batch can help you visualize the augmentations being applied and whether or not they are too extreme or too subtle.
- Viewing the mini-batch can help you determine if there are any oddities with the data, such as the predictor variables and masks not aligning or the bands being in the incorrect order.
viewBatch(dataLoader=trainDL,
nCols = 5,
r = 1,
g = 2,
b = 3,
cNames=c("Background", "Buildings", "Woodland", "Water", "Road"),
cColors=c("gray", "red", "green", "blue", "black"))
viewBatch(dataLoader=valDL,
nCols = 5,
r = 1,
g = 2,
b = 3,
cNames=c("Background", "Buildings", "Woodland", "Water", "Road"),
cColors=c("gray", "red", "green", "blue", "black"))
trainStats <- describeBatch(trainDL,
zeroStart=FALSE)
valStats <- describeBatch(valDL,
zeroStart=FALSE)print(trainStats)
$batchSize
[1] 15
$imageDataType
[1] "Float"
$maskDataType
[1] "Long"
$imageShape
[1] "15" "3" "512" "512"
$maskShape
[1] "15" "1" "512" "512"
$bndMns
[1] 0.3979441 0.4192438 0.3617396
$bandSDs
[1] 0.1630734 0.1325573 0.1078829
$maskCount
[1] 2618263 14723 1063879 15848 219447
$minIndex
[1] 1
$maxIndex
[1] 5
print(valStats)
$batchSize
[1] 15
$imageDataType
[1] "Float"
$maskDataType
[1] "Long"
$imageShape
[1] "15" "3" "512" "512"
$maskShape
[1] "15" "1" "512" "512"
$bndMns
[1] 0.3086319 0.3634965 0.3161972
$bandSDs
[1] 0.1500528 0.1296813 0.1046080
$maskCount
[1] 2183194 53566 1374415 259548 61437
$minIndex
[1] 1
$maxIndex
[1] 5
