File size: 10,260 Bytes
ecf08bc |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 |
# Copyright 2020 Division of Medical Image Computing, German Cancer Research Center (DKFZ), Heidelberg, Germany
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from copy import deepcopy
import numpy as np
from nnunet.experiment_planning.common_utils import get_pool_and_conv_props
from nnunet.experiment_planning.experiment_planner_baseline_3DUNet import ExperimentPlanner
from nnunet.network_architecture.generic_UNet import Generic_UNet
from nnunet.paths import *
class ExperimentPlanner3D_v21(ExperimentPlanner):
"""
Combines ExperimentPlannerPoolBasedOnSpacing and ExperimentPlannerTargetSpacingForAnisoAxis
We also increase the base_num_features to 32. This is solely because mixed precision training with 3D convs and
amp is A LOT faster if the number of filters is divisible by 8
"""
def __init__(self, folder_with_cropped_data, preprocessed_output_folder):
super(ExperimentPlanner3D_v21, self).__init__(folder_with_cropped_data, preprocessed_output_folder)
self.data_identifier = "nnUNetData_plans_v2.1"
self.plans_fname = join(self.preprocessed_output_folder,
"nnUNetPlansv2.1_plans_3D.pkl")
self.unet_base_num_features = 32
def get_target_spacing(self):
"""
per default we use the 50th percentile=median for the target spacing. Higher spacing results in smaller data
and thus faster and easier training. Smaller spacing results in larger data and thus longer and harder training
For some datasets the median is not a good choice. Those are the datasets where the spacing is very anisotropic
(for example ACDC with (10, 1.5, 1.5)). These datasets still have examples with a spacing of 5 or 6 mm in the low
resolution axis. Choosing the median here will result in bad interpolation artifacts that can substantially
impact performance (due to the low number of slices).
"""
spacings = self.dataset_properties['all_spacings']
sizes = self.dataset_properties['all_sizes']
target = np.percentile(np.vstack(spacings), self.target_spacing_percentile, 0)
# This should be used to determine the new median shape. The old implementation is not 100% correct.
# Fixed in 2.4
# sizes = [np.array(i) / target * np.array(j) for i, j in zip(spacings, sizes)]
target_size = np.percentile(np.vstack(sizes), self.target_spacing_percentile, 0)
target_size_mm = np.array(target) * np.array(target_size)
# we need to identify datasets for which a different target spacing could be beneficial. These datasets have
# the following properties:
# - one axis which much lower resolution than the others
# - the lowres axis has much less voxels than the others
# - (the size in mm of the lowres axis is also reduced)
worst_spacing_axis = np.argmax(target)
other_axes = [i for i in range(len(target)) if i != worst_spacing_axis]
other_spacings = [target[i] for i in other_axes]
other_sizes = [target_size[i] for i in other_axes]
has_aniso_spacing = target[worst_spacing_axis] > (self.anisotropy_threshold * max(other_spacings))
has_aniso_voxels = target_size[worst_spacing_axis] * self.anisotropy_threshold < min(other_sizes)
# we don't use the last one for now
#median_size_in_mm = target[target_size_mm] * RESAMPLING_SEPARATE_Z_ANISOTROPY_THRESHOLD < max(target_size_mm)
if has_aniso_spacing and has_aniso_voxels:
spacings_of_that_axis = np.vstack(spacings)[:, worst_spacing_axis]
target_spacing_of_that_axis = np.percentile(spacings_of_that_axis, 10)
# don't let the spacing of that axis get higher than the other axes
if target_spacing_of_that_axis < max(other_spacings):
target_spacing_of_that_axis = max(max(other_spacings), target_spacing_of_that_axis) + 1e-5
target[worst_spacing_axis] = target_spacing_of_that_axis
return target
def get_properties_for_stage(self, current_spacing, original_spacing, original_shape, num_cases,
num_modalities, num_classes):
"""
ExperimentPlanner configures pooling so that we pool late. Meaning that if the number of pooling per axis is
(2, 3, 3), then the first pooling operation will always pool axes 1 and 2 and not 0, irrespective of spacing.
This can cause a larger memory footprint, so it can be beneficial to revise this.
Here we are pooling based on the spacing of the data.
"""
new_median_shape = np.round(original_spacing / current_spacing * original_shape).astype(int)
dataset_num_voxels = np.prod(new_median_shape) * num_cases
# the next line is what we had before as a default. The patch size had the same aspect ratio as the median shape of a patient. We swapped t
# input_patch_size = new_median_shape
# compute how many voxels are one mm
input_patch_size = 1 / np.array(current_spacing)
# normalize voxels per mm
input_patch_size /= input_patch_size.mean()
# create an isotropic patch of size 512x512x512mm
input_patch_size *= 1 / min(input_patch_size) * 512 # to get a starting value
input_patch_size = np.round(input_patch_size).astype(int)
# clip it to the median shape of the dataset because patches larger then that make not much sense
input_patch_size = [min(i, j) for i, j in zip(input_patch_size, new_median_shape)]
network_num_pool_per_axis, pool_op_kernel_sizes, conv_kernel_sizes, new_shp, \
shape_must_be_divisible_by = get_pool_and_conv_props(current_spacing, input_patch_size,
self.unet_featuremap_min_edge_length,
self.unet_max_numpool)
# we compute as if we were using only 30 feature maps. We can do that because fp16 training is the standard
# now. That frees up some space. The decision to go with 32 is solely due to the speedup we get (non-multiples
# of 8 are not supported in nvidia amp)
ref = Generic_UNet.use_this_for_batch_size_computation_3D * self.unet_base_num_features / \
Generic_UNet.BASE_NUM_FEATURES_3D
here = Generic_UNet.compute_approx_vram_consumption(new_shp, network_num_pool_per_axis,
self.unet_base_num_features,
self.unet_max_num_filters, num_modalities,
num_classes,
pool_op_kernel_sizes, conv_per_stage=self.conv_per_stage)
while here > ref:
axis_to_be_reduced = np.argsort(new_shp / new_median_shape)[-1]
tmp = deepcopy(new_shp)
tmp[axis_to_be_reduced] -= shape_must_be_divisible_by[axis_to_be_reduced]
_, _, _, _, shape_must_be_divisible_by_new = \
get_pool_and_conv_props(current_spacing, tmp,
self.unet_featuremap_min_edge_length,
self.unet_max_numpool,
)
new_shp[axis_to_be_reduced] -= shape_must_be_divisible_by_new[axis_to_be_reduced]
# we have to recompute numpool now:
network_num_pool_per_axis, pool_op_kernel_sizes, conv_kernel_sizes, new_shp, \
shape_must_be_divisible_by = get_pool_and_conv_props(current_spacing, new_shp,
self.unet_featuremap_min_edge_length,
self.unet_max_numpool,
)
here = Generic_UNet.compute_approx_vram_consumption(new_shp, network_num_pool_per_axis,
self.unet_base_num_features,
self.unet_max_num_filters, num_modalities,
num_classes, pool_op_kernel_sizes,
conv_per_stage=self.conv_per_stage)
#print(new_shp)
#print(here, ref)
input_patch_size = new_shp
batch_size = Generic_UNet.DEFAULT_BATCH_SIZE_3D # This is what wirks with 128**3
batch_size = int(np.floor(max(ref / here, 1) * batch_size))
# check if batch size is too large
max_batch_size = np.round(self.batch_size_covers_max_percent_of_dataset * dataset_num_voxels /
np.prod(input_patch_size, dtype=np.int64)).astype(int)
max_batch_size = max(max_batch_size, self.unet_min_batch_size)
batch_size = max(1, min(batch_size, max_batch_size))
do_dummy_2D_data_aug = (max(input_patch_size) / input_patch_size[
0]) > self.anisotropy_threshold
plan = {
'batch_size': batch_size,
'num_pool_per_axis': network_num_pool_per_axis,
'patch_size': input_patch_size,
'median_patient_size_in_voxels': new_median_shape,
'current_spacing': current_spacing,
'original_spacing': original_spacing,
'do_dummy_2D_data_aug': do_dummy_2D_data_aug,
'pool_op_kernel_sizes': pool_op_kernel_sizes,
'conv_kernel_sizes': conv_kernel_sizes,
}
return plan
|