The data comes from the Chest X-Ray Images (Pneumonia) data-set(s) in publicly available from kaggle
The data was prepared as seen in the following R code:
# FULL LIST OF FILES - TRAIN
trn.nrm.all <- list.files( "./chest_xray/train/NORMAL/" )
trn.pnm.all <- list.files( "./chest_xray/train/PNEUMONIA/" )
# FULL LIST OF FILES - TEST
tst.nrm.all <- list.files( "./chest_xray/test/NORMAL/" )
tst.pnm.all <- list.files( "./chest_xray/test/PNEUMONIA/" )
# SPLIT IN THREE RANDOM SETS
split <- function( files, ngroups = 3 ) {
n <- length( files ) / ngroups
files <- sample( files )
lapply( seq( ngroups ), function( ii ) {
start <- 1 + ( ii - 1 ) * n
end <- n * ii
if( ii == ngroups & n - floor( n ) != 0 ) {
end <- n * ii + 1
}
return( files[ start: end ] )
} )
}
trn.nrm.3 <- split( trn.nrm.all, 3 )
trn.pnm.3 <- split( trn.pnm.all, 3 )
tst.nrm.3 <- split( tst.nrm.all, 3 )
tst.pnm.3 <- split( tst.pnm.all, 3 )
save_ids <- function( split_files, filename ) {
for( ii in seq( length( split_files ) ) ) {
localname <- paste0( filename, ii, ".csv" )
data <- data.frame( image_name = split_files[[ ii ]] )
write.csv( data, file = localname, quote = FALSE, row.names = FALSE )
}
}
save_ids( trn.nrm.3, "./data_ids/three_dataseties_scenario/train.nrm.3_" )
save_ids( trn.pnm.3, "./data_ids/three_dataseties_scenario/train.pnm.3_" )
save_ids( tst.nrm.3, "./data_ids/three_dataseties_scenario/test.nrm.3_" )
save_ids( tst.pnm.3, "./data_ids/three_dataseties_scenario/test.pnm.3_" )
# SPLINT IN TWO RANDOM SETS
trn.nrm.2 <- split( trn.nrm.all, 2 )
trn.pnm.2 <- split( trn.pnm.all, 2 )
tst.nrm.2 <- split( tst.nrm.all, 2 )
tst.pnm.2 <- split( tst.pnm.all, 2 )
save_ids( trn.nrm.2, "./data_ids/two_datasites_scenario/train.nrm.2_" )
save_ids( trn.pnm.2, "./data_ids/two_datasites_scenario/train.pnm.2_" )
save_ids( tst.nrm.2, "./data_ids/two_datasites_scenario/test.nrm.2_" )
save_ids( tst.pnm.2, "./data_ids/two_datasites_scenario/test.pnm.2_" )Therefore:
- There are two possible scenarios: two or three data-sites available
- Each file contains a half or a third of the full data-sets
- Each files contains the same variable (aka.
image_name) with the name of the image that is included in the set
The model to test is the following:
- 2d Convolution layer (input layer; in: 3, out: 12)
- 2d normalize (1st hidden layer; in: 12, out: 12)
- Rectified linear unit (2nd hidden layer; in: 12, out: 12)
- 2d pooling (max) (3th hidden layer; in: 12; out: 12)
- 2d Convolution layer (4th hidden layer; in: 12, out: 20)
- Rectified linear unit (5th hidden layer; in: 20, out: 20)
- 2d Convolution layer (6th hidden layer; in: 20, out: 32)
- 2d normalize (7th hidden layer; in: 32, out: 32)
- Rectified linear unit (8th hidden layer; in: 32, out: 32)
- Linear discriminant (output layer; in: 32, out: 2)
Depicted as:
In order to validate the results of the FL DL exercise, we will take the following as the ground truth model:
class CnnModel( nn.Module ) :
def __init__( self, num_classes = 2 ):
super( CnnModel, self ).__init__()
self.conv1 = nn.Conv2d( in_channels = 3, out_channels = 12, kernel_size = 3, stride = 1, padding = 1 )
self.bn1 = nn.BatchNorm2d( num_features = 12 )
self.relu1 = nn.ReLU()
self.pool = nn.MaxPool2d( kernel_size = 2 )
self.conv2 = nn.Conv2d( in_channels = 12, out_channels = 20, kernel_size = 3, stride = 1, padding = 1 )
self.relu2 = nn.ReLU()
self.conv3 = nn.Conv2d( in_channels = 20, out_channels = 32, kernel_size = 3, stride = 1, padding = 1 )
self.bn3 = nn.BatchNorm2d( num_features = 32 )
self.relu3 = nn.ReLU()
self.fc = nn.Linear( in_features = 32 * 112 * 112, out_features = num_classes )
def forward( self, input ):
output = self.conv1( input )
output = self.bn1( output )
output = self.relu1( output )
output = self.pool( output )
output = self.conv2( output )
output = self.relu2( output )
output = self.conv3( output )
output = self.bn3( output )
output = self.relu3( output )
output = output.view( -1, 32*112*112 )
output = self.fc( output )
return outputThis implementation is python based using:
torchtorchvision
The model was trained in 10 epochs and using the Adam optimizer and loss of cross entropy criterion using only local CPU, and using the full set of train data-set as seen as:
Epoch 0 - Loss: 107.02728892862797
Epoch 1 - Loss: 16.680565030500293
Epoch 2 - Loss: 7.684169912338257
Epoch 3 - Loss: 8.067254842817784
Epoch 4 - Loss: 3.4069480776786802
Epoch 5 - Loss: 4.979365282598883
Epoch 6 - Loss: 3.159924500063062
Epoch 7 - Loss: 1.560024076141417
Epoch 8 - Loss: 1.0648583032190797
Epoch 9 - Loss: 0.7512097196653486
Model test accuracy: 0.7756410256410257; Model test loss: 19.418610334396362Also, for this model, I resized all pictures as (3, 224, 224).