[suggestion] remove chainermn dependency from FPN sample.

[apple2373]

I mentioned here https://github.com/chainer/chainercv/issues/735#issuecomment-479616802_ before and currently FPN detector is depending on the chainermn. Unfortunately chainermn is not easy to install for those (including me) who are not familiar with server side, so I had manually to remove the dependency....

How about make the chainermn optional?

I attach my ad-hoc coding but I think you can provide something like

if (chaienrmn is installed):
    use chainermn
else:
   do not use chainermn but still code runs.

My ad-hoc coding:

from __future__ import division

import argparse
import multiprocessing
import numpy as np

import chainer
import chainer.links as L
from chainer.optimizer_hooks import WeightDecay
from chainer import serializers
from chainer import training
from chainer.training import extensions

# import chainermn

from chainercv.chainer_experimental.datasets.sliceable import TransformDataset
from chainercv.chainer_experimental.training.extensions import make_shift
from chainercv.datasets import coco_bbox_label_names
from chainercv.datasets import COCOBboxDataset
from chainercv.links import FasterRCNNFPNResNet101
from chainercv.links import FasterRCNNFPNResNet50
from chainercv import transforms

from chainercv.links.model.fpn import head_loss_post
from chainercv.links.model.fpn import head_loss_pre
from chainercv.links.model.fpn import rpn_loss

# https://docs.chainer.org/en/stable/tips.html#my-training-process-gets-stuck-when-using-multiprocessiterator
try:
    import cv2
    cv2.setNumThreads(0)
except ImportError:
    pass

class TrainChain(chainer.Chain):

    def __init__(self, model):
        super(TrainChain, self).__init__()
        with self.init_scope():
            self.model = model

    def forward(self, imgs, bboxes, labels):
        x, scales = self.model.prepare(imgs)
        bboxes = [self.xp.array(bbox) * scale
                  for bbox, scale in zip(bboxes, scales)]
        labels = [self.xp.array(label) for label in labels]

        with chainer.using_config('train', False):
            hs = self.model.extractor(x)

        rpn_locs, rpn_confs = self.model.rpn(hs)
        anchors = self.model.rpn.anchors(h.shape[2:] for h in hs)
        rpn_loc_loss, rpn_conf_loss = rpn_loss(
            rpn_locs, rpn_confs, anchors,
            [(int(img.shape[1] * scale), int(img.shape[2] * scale))
             for img, scale in zip(imgs, scales)],
            bboxes)

        rois, roi_indices = self.model.rpn.decode(
            rpn_locs, rpn_confs, anchors, x.shape)
        rois = self.xp.vstack([rois] + bboxes)
        roi_indices = self.xp.hstack(
            [roi_indices]
            + [self.xp.array((i,) * len(bbox))
               for i, bbox in enumerate(bboxes)])
        rois, roi_indices = self.model.head.distribute(rois, roi_indices)
        rois, roi_indices, head_gt_locs, head_gt_labels = head_loss_pre(
            rois, roi_indices, self.model.head.std, bboxes, labels)
        head_locs, head_confs = self.model.head(hs, rois, roi_indices)
        head_loc_loss, head_conf_loss = head_loss_post(
            head_locs, head_confs,
            roi_indices, head_gt_locs, head_gt_labels, len(x))

        loss = rpn_loc_loss + rpn_conf_loss + head_loc_loss + head_conf_loss
        chainer.reporter.report({
            'loss': loss,
            'loss/rpn/loc': rpn_loc_loss, 'loss/rpn/conf': rpn_conf_loss,
            'loss/head/loc': head_loc_loss, 'loss/head/conf': head_conf_loss},
            self)

        return loss

def transform(in_data):
    img, bbox, label = in_data

    img, params = transforms.random_flip(
        img, x_random=True, return_param=True)
    bbox = transforms.flip_bbox(
        bbox, img.shape[1:], x_flip=params['x_flip'])

    return img, bbox, label

def converter(batch, device=None):
    # do not send data to gpu (device is ignored)
    return tuple(list(v) for v in zip(*batch))

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        '--model',
        choices=('faster_rcnn_fpn_resnet50', 'faster_rcnn_fpn_resnet101'),
        default='faster_rcnn_fpn_resnet50')
    parser.add_argument('--batchsize', type=int, default=16)
    parser.add_argument('--iteration', type=int, default=90000)
    parser.add_argument('--step', type=int, nargs='*', default=[60000, 80000])
    parser.add_argument('--out', default='result')
    parser.add_argument('--resume')
    args = parser.parse_args()

    # # https://docs.chainer.org/en/stable/chainermn/tutorial/tips_faqs.html#using-multiprocessiterator
    # if hasattr(multiprocessing, 'set_start_method'):
    #     multiprocessing.set_start_method('forkserver')
    #     p = multiprocessing.Process()
    #     p.start()
    #     p.join()

    # comm = chainermn.create_communicator()
    # device = comm.intra_rank
    device = 0

    if args.model == 'faster_rcnn_fpn_resnet50':
        model = FasterRCNNFPNResNet50(
            n_fg_class=len(coco_bbox_label_names), pretrained_model='imagenet')
    elif args.model == 'faster_rcnn_fpn_resnet101':
        model = FasterRCNNFPNResNet101(
            n_fg_class=len(coco_bbox_label_names), pretrained_model='imagenet')

    model.use_preset('evaluate')
    train_chain = TrainChain(model)
    chainer.cuda.get_device_from_id(device).use()
    train_chain.to_gpu()

    train = TransformDataset(
        COCOBboxDataset(year='2017', split='train'),
        ('img', 'bbox', 'label'), transform)

    # if comm.rank == 0:
    #     indices = np.arange(len(train))
    # else:
    #     indices = None
    # indices = chainermn.scatter_dataset(indices, comm, shuffle=True)
    # train = train.slice[indices]

    train_iter = chainer.iterators.MultithreadIterator(
        train, args.batchsize)

    optimizer = chainer.optimizers.MomentumSGD()
    optimizer.setup(train_chain)
    optimizer.add_hook(WeightDecay(0.0001))

    model.extractor.base.conv1.disable_update()
    model.extractor.base.res2.disable_update()
    for link in model.links():
        if isinstance(link, L.BatchNormalization):
            link.disable_update()

    updater = training.updaters.StandardUpdater(
        train_iter, optimizer, converter=converter, device=device)
    trainer = training.Trainer(
        updater, (args.iteration * 16 / args.batchsize, 'iteration'), args.out)

    @make_shift('lr')
    def lr_schedule(trainer):
        base_lr = 0.02 * args.batchsize / 16
        warm_up_duration = 500
        warm_up_rate = 1 / 3

        iteration = trainer.updater.iteration
        if iteration < warm_up_duration:
            rate = warm_up_rate \
                + (1 - warm_up_rate) * iteration / warm_up_duration
        else:
            rate = 1
            for step in args.step:
                if iteration >= step * 16 / args.batchsize:
                    rate *= 0.1

        return base_lr * rate

    trainer.extend(lr_schedule)

    log_interval = 10, 'iteration'
    trainer.extend(extensions.LogReport(trigger=log_interval))
    trainer.extend(extensions.observe_lr(), trigger=log_interval)
    trainer.extend(extensions.PrintReport(
        ['epoch', 'iteration', 'lr', 'main/loss',
         'main/loss/rpn/loc', 'main/loss/rpn/conf',
         'main/loss/head/loc', 'main/loss/head/conf']),
        trigger=log_interval)
    trainer.extend(extensions.ProgressBar(update_interval=10))

    trainer.extend(extensions.snapshot(), trigger=(10000, 'iteration'))
    trainer.extend(
        extensions.snapshot_object(
            model, 'model_iter_{.updater.iteration}'),
        trigger=(90000 * 16 / args.batchsize, 'iteration'))

    if args.resume:
        serializers.load_npz(args.resume, trainer, strict=False)

    trainer.run()

if __name__ == '__main__':
    main()

and then

CUDA_VISIBLE_DEVICES=2 python train_multi.py --model faster_rcnn_fpn_resnet50 --batchsize 3

[kuenishi]

What operating system are you using? Most Linux distributions has OpenMPI available from package manager. For example, in Ubuntu it's quite easy like apt-get install openmpi-bin and pip install mpi4py. Then you can run mpirun -np 2 python train_multi.py --model faster_rcnn_fpn_resnet50 if you have two GPUs in your machine. Hopefully I'd like to know what makes you think it difficult to install chainermn.

[apple2373]

I use a university server, and don't have root. I simply can't run apt-get.

[kuenishi]

Some of us are using mpienv. Hope this helps.

[apple2373]

Thanks! I'll try later.

FIY, I actually tried conda install openmpi before but didn't work.

Well, to be clear, I am NOT asking help for installing MPI or setting up chainermn. I would do that in chainermn issue or chainer slack if I ever wanted to do that. This issue is to suggest removing chainermn requirements, because that's not essential for FPN training.

Also, the reason why I don't want to use MPI is not only because I can't set it up. I know I could do that if I spend more time and compile from source. It's because I will not be able to use multiple GPUs most of the time anyway, due to the limited number of GPUs in my lab. Then, MPI will just introduce unnecessary overload when used with one gpu.

Anyway, if the chainercv team decides to keep chainermn dependency, that's fine for me, and you can close the issue. This is just a suggestion from a point of view.

[Hakuyume]

For some examples, we provide both w/o ChainerMN and w/ ChainerMN versions (e.g. examples/ssd/train.py vs examples/ssd/train_multi.py). In the case of FPN, my concern is that we can not get enough batchsize with single GPU and the performance will be worse.

I think we have two options.

1. Provide script w/o ChainerMN with the original batchsize (batchsize=16). This script will reproduce the score reported in Detectron theoretically. However, most users have to reduce the batchsize practically due to the GPU memory. By changing the batchsize manually, users can be aware that the setting is different from that of the original paper.
2. Provide script w/o ChainerMN with small batchsize (batchsize=1 or 2). This script can work without any modification. However, the performance will be lower than Detectron. Some users may think "This example should have a bug".

Note that we face the same problem even if we provide a unified script that supports both w/o ChainerMN and w/ ChainerMN.

[apple2373]

Thanks for the comment! I am in favor of option 1. Not everyone has the same environment, and to me, it's acceptable that users have to adjust command line arguments (but not the code) depending on their own situation. Also, I think GPU memory will increase in the future, so the problem will be solved in the long run.

[apple2373]

I'd like to suggest another option. Maybe you can call option 3.

How about using gradient accomulation to emulate a large batch size for the single GPU case? I asked at the chainer slack and confirmed that it's possible.

https://chainer.slack.com/archives/C0LC5A6C9/p1555395952007300 https://chainer.slack.com/archives/C0LC5A6C9/p1555396176007900