-
Notifications
You must be signed in to change notification settings - Fork 668
/
unet_evaluation_workflows.py
197 lines (173 loc) · 7.97 KB
/
unet_evaluation_workflows.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
# Copyright (c) MONAI Consortium
# 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.
"""
This example shows how to execute distributed evaluation based on PyTorch native `DistributedDataParallel` module
and MONAI workflows. It can run on several nodes with multiple GPU devices on every node.
Main steps to set up the distributed evaluation:
- Execute `torch.distributed.launch` to create processes on every node for every GPU.
It receives parameters as below:
`--nproc_per_node=NUM_GPUS_PER_NODE`
`--nnodes=NUM_NODES`
`--node_rank=INDEX_CURRENT_NODE`
`--master_addr="192.168.1.1"`
`--master_port=1234`
For more details, refer to https://github.com/pytorch/pytorch/blob/master/torch/distributed/launch.py.
Alternatively, we can also use `torch.multiprocessing.spawn` to start program, but in that case, need to handle
all the above parameters and compute `rank` manually, then set to `init_process_group`, etc.
`torch.distributed.launch` is even more efficient than `torch.multiprocessing.spawn`.
- Use `init_process_group` to initialize every process, every GPU runs in a separate process with unique rank.
Here we use `NVIDIA NCCL` as the backend and must set `init_method="env://"` if use `torch.distributed.launch`.
- We need the `--use_env` to avoid args.local_rank, ignite.distributed will handle that for us.
- Wrap the model with `DistributedDataParallel` after moving to expected device.
- Put model file on every node, then load and map to expected GPU device in every process.
- Wrap Dataset with `DistributedSampler`, disable the `shuffle` in sampler and DataLoader.
- Add `StatsHandler` to the master process which is `dist.get_rank() == 0`.
- ignite can automatically reduce metrics for distributed evaluation, refer to:
https://github.com/pytorch/ignite/blob/v0.4.2/ignite/metrics/metric.py#L507
Note:
`torch.distributed.launch` will launch `nnodes * nproc_per_node = world_size` processes in total.
Suggest setting exactly the same software environment for every node, especially `PyTorch`, `nccl`, etc.
A good practice is to use the same MONAI docker image for all nodes directly.
Example script to execute this program on every node:
python -m torch.distributed.launch --nproc_per_node=NUM_GPUS_PER_NODE
--nnodes=NUM_NODES --node_rank=INDEX_CURRENT_NODE
--master_addr="192.168.1.1" --master_port=1234 --use_env
unet_evaluation_workflows.py -d DIR_OF_TESTDATA
This example was tested with [Ubuntu 16.04/20.04], [NCCL 2.6.3].
Referring to: https://pytorch.org/tutorials/intermediate/ddp_tutorial.html
"""
import argparse
import logging
import os
import sys
from glob import glob
import nibabel as nib
import numpy as np
import torch
import torch.distributed as dist
import ignite.distributed as idist
from ignite.metrics import Accuracy
from torch.nn.parallel import DistributedDataParallel
from torch.utils.data.distributed import DistributedSampler
import monai
from monai.data import DataLoader, Dataset, create_test_image_3d
from monai.engines import SupervisedEvaluator
from monai.handlers import CheckpointLoader, MeanDice, StatsHandler, from_engine
from monai.inferers import SlidingWindowInferer
from monai.transforms import (
Activationsd,
EnsureChannelFirstd,
AsDiscreted,
Compose,
KeepLargestConnectedComponentd,
LoadImaged,
ScaleIntensityd,
SaveImaged,
)
def evaluate(args):
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
# initialize the distributed evaluation process, every GPU runs in a process
dist.init_process_group(backend="nccl", init_method="env://")
if idist.get_local_rank() == 0 and not os.path.exists(args.dir):
# create 16 random image, mask paris for evaluation
print(f"generating synthetic data to {args.dir} (this may take a while)")
os.makedirs(args.dir)
# set random seed to generate same random data for every node
np.random.seed(seed=0)
for i in range(16):
im, seg = create_test_image_3d(128, 128, 128, num_seg_classes=1, channel_dim=-1)
n = nib.Nifti1Image(im, np.eye(4))
nib.save(n, os.path.join(args.dir, f"img{i:d}.nii.gz"))
n = nib.Nifti1Image(seg, np.eye(4))
nib.save(n, os.path.join(args.dir, f"seg{i:d}.nii.gz"))
idist.barrier()
images = sorted(glob(os.path.join(args.dir, "img*.nii.gz")))
segs = sorted(glob(os.path.join(args.dir, "seg*.nii.gz")))
val_files = [{"image": img, "label": seg} for img, seg in zip(images, segs)]
# define transforms for image and segmentation
val_transforms = Compose(
[
LoadImaged(keys=["image", "label"]),
EnsureChannelFirstd(keys=["image", "label"], channel_dim=-1),
ScaleIntensityd(keys="image"),
]
)
# create a evaluation data loader
val_ds = Dataset(data=val_files, transform=val_transforms)
# create a evaluation data sampler
val_sampler = DistributedSampler(val_ds, shuffle=False)
# sliding window inference need to input 1 image in every iteration
val_loader = DataLoader(val_ds, batch_size=1, shuffle=False, num_workers=2, pin_memory=True, sampler=val_sampler)
# create UNet, DiceLoss and Adam optimizer
device = torch.device(f"cuda:{idist.get_local_rank()}")
torch.cuda.set_device(device)
net = monai.networks.nets.UNet(
spatial_dims=3,
in_channels=1,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(device)
# wrap the model with DistributedDataParallel module
net = DistributedDataParallel(net, device_ids=[device])
val_post_transforms = Compose(
[
Activationsd(keys="pred", sigmoid=True),
AsDiscreted(keys="pred", threshold=0.5),
KeepLargestConnectedComponentd(keys="pred", applied_labels=[1]),
SaveImaged(keys="pred", meta_keys="image_meta_dict", output_dir="./runs/"),
]
)
val_handlers = [
CheckpointLoader(
load_path="./runs/checkpoint_epoch=4.pt",
load_dict={"net": net},
# config mapping to expected GPU device
map_location={"cuda:0": f"cuda:{idist.get_local_rank()}"},
),
]
if idist.get_rank() == 0:
val_handlers.extend(
[
StatsHandler(output_transform=lambda x: None),
]
)
evaluator = SupervisedEvaluator(
device=device,
val_data_loader=val_loader,
network=net,
inferer=SlidingWindowInferer(roi_size=(96, 96, 96), sw_batch_size=4, overlap=0.5),
postprocessing=val_post_transforms,
key_val_metric={
"val_mean_dice": MeanDice(
include_background=True,
output_transform=from_engine(["pred", "label"]),
)
},
additional_metrics={"val_acc": Accuracy(output_transform=from_engine(["pred", "label"]), device=device)},
val_handlers=val_handlers,
amp=True,
)
evaluator.run()
dist.destroy_process_group()
def main():
parser = argparse.ArgumentParser()
parser.add_argument("-d", "--dir", default="./testdata", type=str, help="directory to create random data")
args = parser.parse_args()
evaluate(args=args)
# python -m torch.distributed.launch --nproc_per_node=NUM_GPUS_PER_NODE
# --nnodes=NUM_NODES --node_rank=INDEX_CURRENT_NODE
# --master_addr="192.168.1.1" --master_port=1234
# --use_env
# unet_evaluation_workflows.py -d DIR_OF_TESTDATA
if __name__ == "__main__":
main()