Error in makePlots() : Problem in salience map computation

[adeldaoud]

@cjerzak I am running the following code and the model seems to complete just fine but then at the end the process throughts an error at me and the ImageConfoundingAnalysis_clear never materializes. Any thoughts?

# Causal inference with image confounding ----

## data definitions

obsW <- sim_data$X  # Observed confounder
obsY <- sim_data$Y  # Observed outcome
X <- NULL
nSGD_selected <- 300L  # Number of stochastic gradient descent iterations

# Perform causal inference with image confounding
imageModelClass <- "VisionTransformer"
optimizeImageRep <- TRUE
print(sprintf("Image confounding analysis & optimizeImageRep: %s & imageModelClass: %s", optimizeImageRep, imageModelClass))

# Perform causal inference with image confounding for clear image
ImageConfoundingAnalysis_clear <- causalimages::AnalyzeImageConfounding(
  # input data 
  obsW = obsW,
  obsY = obsY,
  X = NULL, # X[, apply(X, 2, sd) > 0],
  imageKeysOfUnits = KeysOfObservations,
  file = TFRecordName_im,
  
  # modeling parameters 
  batchSize = 16L,
  nBoot = 5L,
  optimizeImageRep = TRUE,
  imageModelClass = imageModelClass,
  nDepth_ImageRep = 4L,
  nWidth_ImageRep = as.integer(2^8),
  learningRateMax = 0.001, nSGD = nSGD_selected,
  dropoutRate = 0.1,
  plotBands = c(1,2,3),
  plotResults = TRUE, figuresTag = "SimConfoundingIm",
  figuresPath = "./"
)
try(dev.off(), TRUE)

Returning output in TransformerBackbone()
Returning ImageRepArm_SpatialArm outputs...
Returning ImageRepArm_batch outputs...
In GetTreatProb_batch() - dense model
Starting GetDense_OneObs()
Returning output and state in GetDense_OneObs()...[1] "Error in h(simpleError(msg, call)) : \n error in evaluating the argument 'x' in selecting a method for function 'brick': incorrect number of dimensions\n"
attr(,"class")
[1] "try-error"
attr(,"condition")
<simpleError in h(simpleError(msg, call)): error in evaluating the argument 'x' in selecting a method for function 'brick': incorrect number of dimensions>
Show Traceback
Rerun with Debug
Error in makePlots() : Problem in salience map computation!

[adeldaoud]

@cjerzak The error still persists. I suspect it has to do with the number of SDG

# Prepare TF records ----
## Write TFRecord (optional, but as in tutorial)
reSaveTFRecord <- TRUE
#TFRecordName_im <- "~/Downloads/SimulatedData_im.tfrecord"
TFRecordName_im <- "C:/Test/SimulatedData_im.tfrecord"

KeysOfImages <- paste0("img_", 1:n_units)  # unique keys

# Keys of observations
KeysOfObservations <- KeysOfImages  # one-to-one for simplicity

# Define acquireImageFxn to identify image matrix  
acquireImageFxn <- function(keys) {
  # This function gets one key at a time (e.g., "img_1") from WriteTfRecord
  index <- as.integer(gsub("img_", "", keys))
  
  # Retrieve the 2D matrix (50x50)
  image_matrix <- sim_data$house_image[[index]]
  
  # Reshape the 2D matrix into a 3D array (50x50x1) to add the channel dimension
  # The TensorFlow backend of causalimages expects this 3D structure
  dim(image_matrix) <- c(1, dim(image_matrix),1)
  
  return(image_matrix)
}

# Check if TFRecord already exists
if(reSaveTFRecord) {
  causalimages::WriteTfRecord(
    file = TFRecordName_im,
    uniqueImageKeys = unique(KeysOfObservations),
    acquireImageFxn = acquireImageFxn
  )
}

# Write TFRecord (optional, but as in tutorial)
reSaveTFRecord <- TRUE
#TFRecordName_im <- "~/Downloads/SimulatedData_im.tfrecord"
TFRecordName_im_systematic_blur <- "C:/Test/SimulatedData_im_systematic_blur.tfrecord"

KeysOfImages <- paste0("img_", 1:n_units)  # unique keys

# Keys of observations
KeysOfObservations <- KeysOfImages  # one-to-one for simplicity

# Define acquireImageFxn to identify image matrix  
acquireImageFxn <- function(keys) {
  # This function gets one key at a time (e.g., "img_1") from WriteTfRecord
  index <- as.integer(gsub("img_", "", keys))
  
  # Retrieve the 2D matrix (50x50)
  image_matrix <- sim_data$systematic_blur_matrix[[index]]
  
  # Reshape the 2D matrix into a 3D array (50x50x1) to add the channel dimension
  # The TensorFlow backend of causalimages expects this 3D structure
  dim(image_matrix) <- c(1, dim(image_matrix),1)
  
  return(image_matrix)
}

# Check if TFRecord already exists
if(reSaveTFRecord) {
  causalimages::WriteTfRecord(
    file = TFRecordName_im_systematic_blur,
    uniqueImageKeys = unique(KeysOfObservations),
    acquireImageFxn = acquireImageFxn
  )
}

# Write TFRecord (optional, but as in tutorial)
reSaveTFRecord <- TRUE
#TFRecordName_im <- "~/Downloads/SimulatedData_im.tfrecord"
TFRecordName_im_noisy_blur <- "C:/Test/SimulatedData_im_noisy_blur.tfrecord"

KeysOfImages <- paste0("img_", 1:n_units)  # unique keys

# Keys of observations
KeysOfObservations <- KeysOfImages  # one-to-one for simplicity

# Define acquireImageFxn to identify image matrix  
acquireImageFxn <- function(keys) {
  # This function gets one key at a time (e.g., "img_1") from WriteTfRecord
  index <- as.integer(gsub("img_", "", keys))
  
  # Retrieve the 2D matrix (50x50)
  image_matrix <- sim_data$noisy_blur_matrix[[index]]
  
  # Reshape the 2D matrix into a 3D array (50x50x1) to add the channel dimension
  # The TensorFlow backend of causalimages expects this 3D structure
  dim(image_matrix) <- c(1, dim(image_matrix),1)
  
  return(image_matrix)
}

# Check if TFRecord already exists
if(reSaveTFRecord) {
  causalimages::WriteTfRecord(
    file = TFRecordName_im_noisy_blur,
    uniqueImageKeys = unique(KeysOfObservations),
    acquireImageFxn = acquireImageFxn
  )
}


# Causal inference with image confounding ----
## data definitions

obsW <- sim_data$X  # Observed treatment
obsY <- sim_data$Y  # Observed outcome
X <- NULL # no adjustment covariates in this example besides image
nSGD_selected <- 50L  # Number of stochastic gradient descent iterations

# Perform causal inference with image confounding
imageModelClass <- "VisionTransformer"
optimizeImageRep <- TRUE
print(sprintf("Image confounding analysis & optimizeImageRep: %s & imageModelClass: %s", optimizeImageRep, imageModelClass))

# Perform causal inference with image confounding for clear image
ImageConfoundingAnalysis_clear <- causalimages::AnalyzeImageConfounding(
  # input data 
  obsW = obsW,
  obsY = obsY,
  X = NULL, # X[, apply(X, 2, sd) > 0],
  imageKeysOfUnits = KeysOfObservations,
  file = TFRecordName_im,
  
  # modeling parameters 
  batchSize = 16L,
  nBoot = 5L,
  optimizeImageRep = TRUE,
  imageModelClass = imageModelClass,
  nDepth_ImageRep = 4L,
  nWidth_ImageRep = as.integer(2^8),
  learningRateMax = 0.001, nSGD = nSGD_selected,
  dropoutRate = 0.1,
  plotBands = 1L,
  plotResults = F, figuresTag = "SimConfoundingIm",
  figuresPath = "./"
)
try(dev.off(), TRUE)

print(ImageConfoundingAnalysis_clear$tauHat_propensityHajek)


# Perform causal inference with image confounding for systematic blur
ImageConfoundingAnalysis_systematic_blur <- causalimages::AnalyzeImageConfounding(
  # input data 
  obsW = obsW,
  obsY = obsY,
  X = NULL, # X[, apply(X, 2, sd) > 0],
  imageKeysOfUnits = KeysOfObservations,
  file = TFRecordName_im_systematic_blur,
  
  # modeling parameters 
  batchSize = 16L,
  nBoot = 5L,
  optimizeImageRep = TRUE,
  imageModelClass = imageModelClass,
  nDepth_ImageRep = 4L,
  nWidth_ImageRep = as.integer(2^8),
  learningRateMax = 0.001, nSGD = nSGD_selected,
  dropoutRate = 0.1,
  plotBands = 1L,
  plotResults = F, figuresTag = "SimConfoundingIm",
  figuresPath = "./"
)
try(dev.off(), TRUE)

print(ImageConfoundingAnalysis_systematic_blur$tauHat_propensityHajek)


# Perform causal inference with noisy blur
ImageConfoundingAnalysis_noisy_blur <- causalimages::AnalyzeImageConfounding(
  # input data 
  obsW = obsW,
  obsY = obsY,
  X = NULL, # X[, apply(X, 2, sd) > 0],
  imageKeysOfUnits = KeysOfObservations,
  file = TFRecordName_im_noisy_blur,
  
  # modeling parameters 
  batchSize = 16L,
  nBoot = 5L,
  optimizeImageRep = TRUE,
  imageModelClass = imageModelClass,
  nDepth_ImageRep = 4L,
  nWidth_ImageRep = as.integer(2^8),
  learningRateMax = 0.001, nSGD = nSGD_selected,
  dropoutRate = 0.1,
  plotBands = 1L,
  plotResults = F, figuresTag = "SimConfoundingIm",
  figuresPath = "./"
)
try(dev.off(), TRUE)

print(ImageConfoundingAnalysis_noisy_blur$tauHat_propensityHajek)