-
Notifications
You must be signed in to change notification settings - Fork 0
/
distributed_segment.py
226 lines (199 loc) · 9.38 KB
/
distributed_segment.py
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
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
import os
import sys
import time
import numpy as np
from tqdm import tqdm
from tifffile import imwrite
import dask
from dask.diagnostics import ProgressBar
from lsm.dataio import get_data
from lsm.utils.logger import Logger
from lsm.distributed import get_model
from lsm.utils.console_log import log
from lsm.utils.train_utils import count_trainable_parameters
from lsm.utils.load_config import create_args_parser, load_config, backup
from lsm.utils.distributed_util import (
init_env,
get_rank,
is_master,
get_local_rank,
get_world_size,
)
def main_function(args):
init_env(args)
rank = get_rank()
local_rank = get_local_rank()
world_size = get_world_size()
exp_dir = args.training.exp_dir
logger = Logger(
log_dir=exp_dir,
save_dir=os.path.join(exp_dir, "chunks"),
monitoring=args.training.get("monitoring", "tensorboard"),
monitoring_dir=os.path.join(exp_dir, "events"),
rank=rank,
is_master=is_master(),
multi_process_logging=(world_size > 1),
)
log.info(f"Segmentation directory: {exp_dir}")
# lazy load data as a dask array
dataset = get_data(args)
gt_vol = next(iter(dataset))[-1]["orig_vol"]
# run distributed segmentation
for model in tqdm(args.segmentation.models):
save_dir = os.path.join(exp_dir, f"{model}_seg")
if not os.path.exists(save_dir):
os.makedirs(save_dir, exist_ok=True)
print(f"Running stitching with {model}...")
# load model, as a segmentation function
segment_func = get_model(model=model)
if model == "cellpose":
img_vol = next(iter(dataset))[-1]
cfg_dict = {
"image": img_vol["orig_vol"],
"debug": args.model.debug,
"channels": args.model.channels,
"boundary": args.model.boundary,
"diameter": args.model.diameter,
"use_anisotropy": args.model.use_anisotropy,
"iou_depth": args.model.stitching.iou_depth,
"iou_threshold": args.model.stitching.iou_threshold,
}
elif model in ["anystar", "anystar-gaussian", "anystar-spherical"]:
img_vol = next(iter(dataset))[-1]
cfg_dict = {
"image": img_vol["orig_vol"],
"scale": args.model.scale,
"debug": args.model.debug,
"boundary": args.model.boundary,
"diameter": args.model.diameter,
"use_anisotropy": args.model.use_anisotropy,
"iou_depth": args.model.stitching.iou_depth,
"iou_threshold": args.model.stitching.iou_threshold,
}
# separate check for weights and hyperparameters
if model == "anystar":
cfg_dict["model_folder"] = args.model.anystar.model_folder
cfg_dict["model_name"] = args.model.anystar.model_name
cfg_dict["weight_name"] = args.model.anystar.weight_name
cfg_dict["prob_thresh"] = args.model.anystar.prob_thresh
cfg_dict["nms_thresh"] = args.model.anystar.nms_thresh
elif model == "anystar-gaussian":
cfg_dict["model_folder"] = args.model.anystar_gaussian.model_folder
cfg_dict["model_name"] = args.model.anystar_gaussian.model_name
cfg_dict["weight_name"] = args.model.anystar_gaussian.weight_name
cfg_dict["prob_thresh"] = args.model.anystar_gaussian.prob_thresh
cfg_dict["nms_thresh"] = args.model.anystar_gaussian.nms_thresh
elif model == "anystar-spherical":
cfg_dict["model_folder"] = args.model.anystar_spherical.model_folder
cfg_dict["model_name"] = args.model.anystar_spherical.model_name
cfg_dict["weight_name"] = args.model.anystar_spherical.weight_name
cfg_dict["prob_thresh"] = args.model.anystar_spherical.prob_thresh
cfg_dict["nms_thresh"] = args.model.anystar_spherical.nms_thresh
else:
raise NotImplementedError(
f"Anystar model type: {model} not implemented. Try one of [anystar, anystar-gaussian, anystar-spherical]"
)
else:
raise NotImplementedError
if args.model.save_gt_proxy:
print(f"Saving ground truth proxy for stitching analysis (model: {model})")
impath = os.path.join(save_dir, f"gt_proxy.tiff")
try:
if model == "cellpose":
from cellpose import models
model = models.Cellpose(gpu=True, model_type="nuclei")
gt_proxy, _, _, _ = model.eval(
gt_vol,
channels=args.model.channels,
z_axis=0,
channel_axis=3,
diameter=args.model.diameter[1],
do_3D=True,
anisotropy=args.model.use_anisotropy,
augment=True,
tile=True,
)
imwrite(impath, gt_proxy)
elif model in ["anystar", "anystar-gaussian", "anystar-spherical"]:
from stardist.models import StarDist3D
# normalize voxel
x = gt_vol.compute() # convert to numpy array
upper = np.percentile(x, 99.9)
x = np.clip(x, 0, upper)
x = (x - x.min()) / (x.max() - x.min())
x = x[..., 0]
# choose appropriate model weights
if model == "anystar":
seg_network = StarDist3D(
None,
name=args.model.anystar.model_name,
basedir=args.model.anystar.model_folder,
)
seg_network.load_weights(name=args.model.anystar.weight_name)
seg_network.trainable = False
seg_network.keras_model.trainable = False
gt_proxy, _ = seg_network.predict_instances(
x,
prob_thresh=args.model.anystar.prob_thresh,
nms_thresh=args.model.anystar.nms_thresh,
n_tiles=None,
scale=args.model.scale,
)
imwrite(impath, gt_proxy)
elif model == "anystar-gaussian":
seg_network = StarDist3D(
None,
name=args.model.anystar_gaussian.model_name,
basedir=args.model.anystar_gaussian.model_folder,
)
seg_network.load_weights(
name=args.model.anystar_gaussian.weight_name
)
seg_network.trainable = False
seg_network.keras_model.trainable = False
gt_proxy, _ = seg_network.predict_instances(
x,
prob_thresh=args.model.anystar_gaussian.prob_thresh,
nms_thresh=args.model.anystar_gaussian.nms_thresh,
n_tiles=None,
scale=args.model.scale,
)
imwrite(impath, gt_proxy)
elif model == "anystar-spherical":
seg_network = StarDist3D(
None,
name=args.model.anystar_spherical.model_name,
basedir=args.model.anystar_spherical.model_folder,
)
seg_network.load_weights(
name=args.model.anystar_spherical.weight_name
)
seg_network.trainable = False
seg_network.keras_model.trainable = False
gt_proxy, _ = seg_network.predict_instances(
x,
prob_thresh=args.model.anystar_spherical.prob_thresh,
nms_thresh=args.model.anystar_spherical.nms_thresh,
n_tiles=None,
scale=args.model.scale,
)
imwrite(impath, gt_proxy)
except:
raise MemoryError(
f"Voxel chunk size is large. Consider reducing shape from {args.segmentation.voxel_shape}"
)
for chunk in tqdm(args.segmentation.chunk_sizes):
print(f"Running segmentation for chunk size: {chunk}")
cfg_dict["chunk"] = chunk
seg_vol = segment_func(**cfg_dict)
with ProgressBar():
with dask.config.set(scheduler="synchronous"):
seg_vol = seg_vol.compute()
fpath = os.path.join(save_dir, f"chunk_{chunk}.tiff")
imwrite(fpath, seg_vol)
if __name__ == "__main__":
parser = create_args_parser()
parser.add_argument("--ddp", action="store_true", help="Distributed processing")
args, unknown = parser.parse_known_args()
config = load_config(args, unknown)
main_function(config)