def tile_model_reconstruction_comparison(data, reconstruction_data, time_step, z_slice, threshold):
rows = 4
cols = 3
tiling = np.zeros((rows * data.shape[1], cols * data.shape[2]), dtype = data.dtype)
rmses = []
for j in range(cols):
img = data[z_slice,:,:,time_step,j+1]
reconstruction = reconstruction_data[z_slice,:,:,time_step,j+1]
difference = np.abs(img-reconstruction)
# Input
i = 0
tiling[
i*data.shape[1]:(i+1)*data.shape[1],
j*data.shape[2]:(j+1)*data.shape[2]] = img
# Reconstruction
i = 1
tiling[
i*data.shape[1]:(i+1)*data.shape[1],
j*data.shape[2]:(j+1)*data.shape[2]] = reconstruction
# Difference
i = 2
tiling[
i*data.shape[1]:(i+1)*data.shape[1],
j*data.shape[2]:(j+1)*data.shape[2]] = difference
# Thresholded difference
difference[difference <= threshold] = 0.
i = 3
tiling[
i*data.shape[1]:(i+1)*data.shape[1],
j*data.shape[2]:(j+1)*data.shape[2]] = difference
error = rmse(img, reconstruction)
rmses.append(error)
difference = np.abs(data[z_slice,:,:,time_step,:] - reconstruction_data[z_slice,:,:,time_step,:])
difference[difference <= threshold] = 0.
return tiling, difference, rmses
# For the selected subject index, grab the full data for that one subject: We only evaluate one data which has information from 0:64 only
subject_sliced = images_vl_sl[0:64,:,:,:,:]
# Initialize the while loop variables
start_idx = 0
end_idx = config["batch_size"]
subject_rmse = []
while end_idx <= config["spatial_size_z"]:
# Selected slices for batch and run through the model
feed_dict = {model.image_matrix: subject_sliced[start_idx:end_idx]}
out_mu = model.sess.run(model.decoder_output, feed_dict)
# Compute rmse of these slices and append it to the subject error
error = rmse(subject_sliced[start_idx:end_idx], out_mu)
subject_rmse.append(error)
# Visualization
if config["visualization_mode"] == 'all':
out_path = os.path.join(project_code_root,'Results/Evaluation/' + model_name + '/validation_segmentedFlowMRI/' + 'Subject_' + str(subject_index) + '_' + str(start_idx) + '_' + str(end_idx) + '.png')
plot_batch_3d_complete(subject_sliced[start_idx:end_idx], out_mu, every_x_time_step=1, out_path= out_path)
# Save it to the hdf5 file
#If all the subjects are saved into same hdf5 file, then change hdf5 path description and bring the multiplication back.
#Not sure if I need to multiple with subject_index here as we are only dealing with one data. So I remove the multiplication for the moment.
print("out_mu shape: ")
print(out_mu.shape)
dataset['reconstruction'][start_idx:end_idx, :, :, :, :] = out_mu
# dataset['reconstruction'][start_idx:end_idx, :, :, :, 0:3] = out_mu
# dataset['reconstruction'][start_idx:end_idx, :, :, :, 4] = segmentation_prob?
# Selected slices for batch and run through the model
# digitize and subtract 1 to get classes starting from 0 (digitize starts counting at 1 for the first bin)
bins = np.linspace(
0, 64,
9) # 64 slices, 8 classes + 1 since outer limit excluded
batch_slice_info = np.digitize(range(start_idx, end_idx),
bins) - 1
feed_dict = {
model.image_matrix: subject_sliced[start_idx:end_idx],
model.batch_slice_info: batch_slice_info
}
out_mu = model.sess.run(model.decoder_output, feed_dict)
# Compute rmse of these slices and append it to the subject error
error = rmse(subject_sliced[start_idx:end_idx], out_mu)
subject_rmse.append(error)
# Visualization
if config["visualization_mode"] == 'all':
out_path = os.path.join(
project_code_root, 'Results/Evaluation/' + model_name +
'/' + which_dataset + '/' + 'Subject_' + str(i) + '_' +
str(start_idx) + '_' + str(end_idx) + '.png')
plot_batch_3d_complete(subject_sliced[start_idx:end_idx],
out_mu,
every_x_time_step=1,
out_path=out_path)
# Save it to the hdf5 file
dataset['reconstruction'][