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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251 | class ZarrDataset(IterableDataset): # type: ignore
def __init__(
self,
dataset_config: DatasetConfig,
crop_size: Tuple[int, ...],
elastic_deform: bool,
control_point_spacing: int,
control_point_jitter: float,
density: float,
kappa: float,
normalization_factor: float,
):
"""A dataset that serves random samples from a zarr container.
Args:
dataset_config:
A dataset config object pointing to the zarr dataset to use.
The dataset should have shape `(s, c, [t,] [z,] y, x)`, where
`s` = # of samples, `c` = # of channels, `t` = # of frames, and
`z`/`y`/`x` are spatial extents. The dataset should have an
`"axis_names"` attribute that contains the names of the used
axes, e.g., `["s", "c", "y", "x"]` for a 2D dataset.
crop_size:
The size of data crops used during training (distinct from the
"patch size" of the method: from each crop, multiple patches
will be randomly selected and the loss computed on them). This
should be equal to the input size of the model that predicts
the OCEs.
elastic_deform:
Whether to elastically deform data in order to augment training samples?
control_point_spacing:
The distance in pixels between control points used for elastic
deformation of the raw data.
Only used, if `elastic_deform` is set to True.
control_point_jitter:
How much to jitter the control points for elastic deformation
of the raw data, given as the standard deviation of a normal
distribution with zero mean.
Only used if `elastic_deform` is set to True.
density:
Determines the fraction of patches to sample per crop, during training.
kappa:
Neighborhood radius to extract patches from.
normalization_factor:
The factor to use, for dividing the raw image pixel intensities.
If 'None', a factor is chosen based on the dtype of the array .
(e.g., np.uint8 would result in a factor of 1.0/255).
"""
self.dataset_config = dataset_config
self.crop_size = crop_size
self.elastic_deform = elastic_deform
self.control_point_spacing = control_point_spacing
self.control_point_jitter = control_point_jitter
self.normalization_factor = normalization_factor
self.__read_meta_data()
assert len(crop_size) == self.num_spatial_dims, (
f'"crop_size" must have the same dimension as the '
f'spatial(temporal) dimensions of the "{self.dataset_config.dataset_name}" '
f"dataset which is {self.num_spatial_dims}, but it is {crop_size}"
)
self.density = density
self.kappa = kappa
self.output_shape = tuple(int(_ - 16) for _ in self.crop_size)
self.normalization_factor = normalization_factor
self.unbiased_shape = tuple(
int(_ - (2 * self.kappa)) for _ in self.output_shape
)
self.__setup_pipeline()
def __iter__(self):
return iter(self.__yield_sample())
def __setup_pipeline(self):
self.raw = gp.ArrayKey("RAW")
# treat all dimensions as spatial, with a voxel size of 1
raw_spec = gp.ArraySpec(voxel_size=(1,) * self.num_dims, interpolatable=True)
# spatial_dims = tuple(range(self.num_dims - self.num_spatial_dims,
# self.num_dims))
self.pipeline = (
gp.ZarrSource(
self.dataset_config.container_path,
{self.raw: self.dataset_config.dataset_name},
array_specs={self.raw: raw_spec},
)
+ gp.RandomLocation()
+ gp.Normalize(self.raw, factor=self.normalization_factor)
)
if self.elastic_deform:
self.pipeline += gp.ElasticAugment(
control_point_spacing=(self.control_point_spacing,)
* self.num_spatial_dims,
jitter_sigma=(self.control_point_jitter,) * self.num_spatial_dims,
rotation_interval=(0, math.pi / 2),
scale_interval=(0.9, 1.1),
subsample=4,
spatial_dims=self.num_spatial_dims,
)
# + gp.SimpleAugment(mirror_only=spatial_dims, transpose_only=spatial_dims)
def __yield_sample(self):
"""An infinite generator of crops."""
with gp.build(self.pipeline):
while True:
array_is_zero = True
# request one sample, all channels, plus crop dimensions
while array_is_zero:
request = gp.BatchRequest()
request[self.raw] = gp.ArraySpec(
roi=gp.Roi(
(0,) * self.num_dims,
(1, self.num_channels, *self.crop_size),
)
)
sample = self.pipeline.request_batch(request)
sample_data = sample[self.raw].data[0]
if np.max(sample_data) <= 0.0:
pass
else:
array_is_zero = False
anchor_samples, reference_samples = self.sample_coordinates()
yield sample_data, anchor_samples, reference_samples
def __read_meta_data(self):
meta_data = DatasetMetaData.from_dataset_config(self.dataset_config)
self.num_dims = meta_data.num_dims
self.num_spatial_dims = meta_data.num_spatial_dims
self.num_channels = meta_data.num_channels
self.num_samples = meta_data.num_samples
self.sample_dim = meta_data.sample_dim
self.channel_dim = meta_data.channel_dim
self.time_dim = meta_data.time_dim
def get_num_channels(self):
return self.num_channels
def get_num_spatial_dims(self):
return self.num_spatial_dims
def sample_offsets_within_radius(self, radius, number_offsets):
if self.num_spatial_dims == 2:
offsets_x = np.random.randint(-radius, radius + 1, size=2 * number_offsets)
offsets_y = np.random.randint(-radius, radius + 1, size=2 * number_offsets)
offsets_coordinates = np.stack((offsets_x, offsets_y), axis=1)
elif self.num_spatial_dims == 3:
offsets_x = np.random.randint(-radius, radius + 1, size=3 * number_offsets)
offsets_y = np.random.randint(-radius, radius + 1, size=3 * number_offsets)
offsets_z = np.random.randint(-radius, radius + 1, size=3 * number_offsets)
offsets_coordinates = np.stack((offsets_x, offsets_y, offsets_z), axis=1)
in_circle = (offsets_coordinates**2).sum(axis=1) < radius**2
offsets_coordinates = offsets_coordinates[in_circle]
not_zero = np.absolute(offsets_coordinates).sum(axis=1) > 0
offsets_coordinates = offsets_coordinates[not_zero]
if len(offsets_coordinates) < number_offsets:
return self.sample_offsets_within_radius(radius, number_offsets)
return offsets_coordinates[:number_offsets]
def sample_coordinates(self):
num_anchors = self.get_num_anchors()
num_references = self.get_num_references()
if self.num_spatial_dims == 2:
anchor_coordinates_x = np.random.randint(
self.kappa,
self.output_shape[0] - self.kappa + 1,
size=num_anchors,
)
anchor_coordinates_y = np.random.randint(
self.kappa,
self.output_shape[1] - self.kappa + 1,
size=num_anchors,
)
anchor_coordinates = np.stack(
(anchor_coordinates_x, anchor_coordinates_y), axis=1
)
elif self.num_spatial_dims == 3:
anchor_coordinates_x = np.random.randint(
self.kappa,
self.output_shape[0] - self.kappa + 1,
size=num_anchors,
)
anchor_coordinates_y = np.random.randint(
self.kappa,
self.output_shape[1] - self.kappa + 1,
size=num_anchors,
)
anchor_coordinates_z = np.random.randint(
self.kappa,
self.output_shape[2] - self.kappa + 1,
size=num_anchors,
)
anchor_coordinates = np.stack(
(anchor_coordinates_x, anchor_coordinates_y, anchor_coordinates_z),
axis=1,
)
anchor_samples = np.repeat(anchor_coordinates, num_references, axis=0)
offset_in_pos_radius = self.sample_offsets_within_radius(
self.kappa, len(anchor_samples)
)
reference_samples = anchor_samples + offset_in_pos_radius
return anchor_samples, reference_samples
def get_num_anchors(self):
return int(self.density * self.unbiased_shape[0] * self.unbiased_shape[1])
def get_num_references(self):
return int(self.density * self.kappa**2 * np.pi)
def get_num_samples(self):
return self.get_num_anchors() * self.get_num_references()
|