unet_training_workflows.py

# Copyright 2020 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 training 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 training:

- 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` during training.
- 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.
- Wrap Dataset with `DistributedSampler`, and disable the `shuffle` in DataLoader.
  Instead, `SupervisedTrainer` shuffles data by `train_sampler.set_epoch(epoch)` before every epoch.
- Add `StatsHandler` and `CheckpointHandler` to the master process which is `idist.get_rank() == 0`.
- ignite can automatically reduce metrics for distributed training, 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_training_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 SupervisedTrainer
from monai.handlers import CheckpointSaver, LrScheduleHandler, StatsHandler, from_engine
from monai.inferers import SimpleInferer
from monai.transforms import (
    Activationsd,
    AsChannelFirstd,
    AsDiscreted,
    Compose,
    KeepLargestConnectedComponentd,
    LoadImaged,
    RandCropByPosNegLabeld,
    RandRotate90d,
    ScaleIntensityd,
)


def train(args):
    logging.basicConfig(stream=sys.stdout, level=logging.INFO)
    # initialize the distributed training 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 40 random image, mask paris for training
        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(40):
            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")))
    train_files = [{"image": img, "label": seg} for img, seg in zip(images, segs)]

    # define transforms for image and segmentation
    train_transforms = Compose(
        [
            LoadImaged(keys=["image", "label"]),
            AsChannelFirstd(keys=["image", "label"], channel_dim=-1),
            ScaleIntensityd(keys="image"),
            RandCropByPosNegLabeld(
                keys=["image", "label"], label_key="label", spatial_size=[96, 96, 96], pos=1, neg=1, num_samples=4
            ),
            RandRotate90d(keys=["image", "label"], prob=0.5, spatial_axes=[0, 2]),
        ]
    )

    # create a training data loader
    train_ds = Dataset(data=train_files, transform=train_transforms)
    # create a training data sampler
    train_sampler = DistributedSampler(train_ds)
    # use batch_size=2 to load images and use RandCropByPosNegLabeld to generate 2 x 4 images for network training
    train_loader = DataLoader(
        train_ds,
        batch_size=2,
        shuffle=False,
        num_workers=2,
        pin_memory=True,
        sampler=train_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)
    loss = monai.losses.DiceLoss(sigmoid=True)
    opt = torch.optim.Adam(net.parameters(), 1e-3)
    lr_scheduler = torch.optim.lr_scheduler.StepLR(opt, step_size=2, gamma=0.1)
    # wrap the model with DistributedDataParallel module
    net = DistributedDataParallel(net, device_ids=[device])

    train_post_transforms = Compose(
        [
            Activationsd(keys="pred", sigmoid=True),
            AsDiscreted(keys="pred", threshold=0.5),
            KeepLargestConnectedComponentd(keys="pred", applied_labels=[1]),
        ]
    )
    train_handlers = [
        LrScheduleHandler(lr_scheduler=lr_scheduler, print_lr=True),
    ]
    if idist.get_rank() == 0:
        train_handlers.extend(
            [
                StatsHandler(tag_name="train_loss", output_transform=from_engine(["loss"], first=True)),
                CheckpointSaver(save_dir="./runs/", save_dict={"net": net, "opt": opt}, save_interval=2),
            ]
        )

    trainer = SupervisedTrainer(
        device=device,
        max_epochs=5,
        train_data_loader=train_loader,
        network=net,
        optimizer=opt,
        loss_function=loss,
        inferer=SimpleInferer(),
        # if no FP16 support in GPU or PyTorch version < 1.6, will not enable AMP evaluation
        amp=True if monai.utils.get_torch_version_tuple() >= (1, 6) else False,
        postprocessing=train_post_transforms,
        key_train_metric={"train_acc": Accuracy(output_transform=from_engine(["pred", "label"]), device=device)},
        train_handlers=train_handlers,
    )
    trainer.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()

    train(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_training_workflows.py -d DIR_OF_TESTDATA

if __name__ == "__main__":
    main()