How are the dimensions generated in the uniform sampler? I tried debugging the image size it seems it works for some of the iterations it does not for others. Any idea how to fix this. My configuration is given below :
[CUSTOM]
num_classes: 14
output_prob: True
label_normalisation: True
softmax: True
min_sampling_ratio: 0
compulsory_labels: (0, 1)
rand_samples: 0
min_numb_labels: 1
proba_connect: True
evaluation_units: foreground
image: ('images',)
label: ('label',)
weight: ()
sampler: ()
name: net_segment
[CONFIG_FILE]
[IMAGES]
csv_file:
path_to_search: /home/ubuntu/med_deacthalon/Task_all_same_names/imagesTr_1
filename_contains: ()
filename_not_contains: ('lung',)
interp_order: 3
loader: None
pixdim: (1.0, 1.0, 1.0)
axcodes: ('A', 'R', 'S')
spatial_window_size: (51, 51, 51)
[LABEL]
-csv_file:
path_to_search: /home/ubuntu/med_deacthalon/Task_all_same_names/labelsTr_1
filename_contains: ()
filename_not_contains: ('lung',)
interp_order: 3
loader: None
pixdim: (1.0, 1.0, 1.0)
axcodes: ('A', 'R', 'S')
spatial_window_size: (9, 9, 9)
[SYSTEM]
cuda_devices: ""
num_threads: 2
num_gpus: 1
model_dir: /home/ubuntu/models_nifty/deepmedic/all_task_same_name_rename_labels
dataset_split_file: ./dataset_split.csv
action: train
[NETWORK]
name: deepmedic
activation_function: relu
batch_size: 32
decay: 0.0
reg_type: L2
volume_padding_size: (21, 21, 21)
volume_padding_mode: minimum
window_sampling: uniform
queue_length: 128
multimod_foreground_type: and
histogram_ref_file: histogram_standardisation_alltask.txt
norm_type: percentile
cutoff: (0.01, 0.99)
foreground_type: otsu_plus
normalisation: False
whitening: True
normalise_foreground_only: True
weight_initializer: he_normal
bias_initializer: zeros
keep_prob: 1.0
weight_initializer_args: {}
bias_initializer_args: {}
[TRAINING]
optimiser: adam
sample_per_volume: 32
rotation_angle: (-10.0, 10.0)
rotation_angle_x: ()
rotation_angle_y: ()
rotation_angle_z: ()
scaling_percentage: (-10.0, 10.0)
random_flipping_axes: -1
do_elastic_deformation: False
num_ctrl_points: 4
deformation_sigma: 15
proportion_to_deform: 0.5
lr: 0.001
loss_type: Dice
starting_iter: 0
save_every_n: 45
tensorboard_every_n: 20
max_iter: 10
max_checkpoints: 20
validation_every_n: -1
validation_max_iter: 1
exclude_fraction_for_validation: 0.0
exclude_fraction_for_inference: 0.0
[INFERENCE]
spatial_window_size: (57, 57, 57)
inference_iter: -1
dataset_to_infer:
save_seg_dir: ./deepmedic/alltask_newname
output_postfix: _niftynet_out
output_interp_order: 0
border: (36, 36, 36)
CRITICAL:niftynet: Don't know how to generate sampling locations: Spatial dimensions of the grouped input sources are not consistent. {(477, 451, 187), (391, 369, 147)}
Exception in thread Thread-2:
Traceback (most recent call last):
File "/home/ubuntu/anaconda3/envs/python3/lib/python3.6/threading.py", line 916, in _bootstrap_inner
self.run()
File "/home/ubuntu/anaconda3/envs/python3/lib/python3.6/threading.py", line 864, in run
self._target(*self._args, **self._kwargs)
File "/home/ubuntu/niftynet/NiftyNet/niftynet/engine/image_window_buffer.py", line 148, in _push
for output_dict in self():
File "/home/ubuntu/niftynet/NiftyNet/niftynet/engine/sampler_uniform.py", line 81, in layer_op
self.window.n_samples)
File "/home/ubuntu/niftynet/NiftyNet/niftynet/engine/sampler_uniform.py", line 151, in _spatial_coordinates_generator
_infer_spatial_size(img_sizes, win_sizes)
File "/home/ubuntu/niftynet/NiftyNet/niftynet/engine/sampler_uniform.py", line 238, in _infer_spatial_size
raise NotImplementedError
NotImplementedError
The issue is resolved here: https://github.com/NifTK/NiftyNet/issues/170
In summary images and labels should have the same voxel spacing values stored in their header when pixdim
is set in the configuration file.