def prepareInputsForGPU(image, full_mask, gaa_mean, gaa_std):
"""
This packs a series of operations that transfer input to GPU and then format and normalizes it.
Inputs:
image, full_mask = MR images and their masks
gaa_mean, gaa_std = input statistics
Outputs:
image and target for neural network on GPU
"""
# transfer input to GPU
gImage = fast_transfer(image)
gImage = gpuDecompress(gImage)
gMask = fast_transfer(full_mask, isImage=False)
gMask = gpuDecompressMask(gMask, 88, 576, 576)
image1, target = gpuNormalise(gImage, gMask, gaa_mean, gaa_std)
return image1, target
output_dir = r'C:\Users\wfok007\femuring\imaging\AtriaSeg_2018_training\transformation\ResNet'
debug_dir = r'C:\Users\wfok007\femuring\imaging\AtriaSeg_2018_training\debug'
hpc = r'Z:\mpi_heart'
weight_dir = 'Z:\mpi_heart\weight_checkpoints'
mean, std = load_stat()
num_ranks = 125
num_ranks_train = 375
od, oh, ow = 88, 112, 112
multiple = 0.25
model = createModel(od, oh, ow, multiple)
# put the statistics on the GPU
aa_mean, aa_std = addAxis(mean, std)
gaa_mean, gaa_std = fast_transfer(aa_mean), fast_transfer(aa_std)
# initialize the data for validation and training
generator_valid = metaGen('valid', num_ranks, hpc)
generator_train = metaGen('train', num_ranks_train, hpc)
criterion = Dice_Loss()
accuracy = non_differentiable_Dice_coef()
learning_rate = math.pow(10, -1.5)
T = 50
lookup, storedKeys, lrs = createCyclical(T,
learning_rate,
'highLow',
min_fraction=1 / 10)