基准测试

在这里,我们以一些其他流行的开源Mask R-CNN实现为基准,对Detectron2中Mask R-CNN的训练速度进行了基准测试。

设置

• 硬件：8个带有NVLink的NVIDIA V100。
• 软件: Python 3.7, CUDA 10.0, cuDNN 7.6.4, PyTorch 1.3.0 (链接(https://download.pytorch.org/whl/nightly/cu100/torch-1.3.0%2Bcu100-cp37-cp37m-linux_x86_64.whl)), TensorFlow 1.15.0rc2, Keras 2.2.5, MxNet 1.6.0b20190820.
• 模型：端到端R-50-FPN Mask-RCNN模型,使用与Detectron基线配置(https://github.com/facebookresearch/Detectron/blob/master/configs/12_2017_baselines/e2e_mask_rcnn_R-50-FPN_1x.yaml)相同的超参数 。
• 指标：我们使用100-500次迭代中的平均吞吐量来跳过GPU预热时间。请注意,对于R-CNN样式的模型,模型的吞吐量通常会在训练期间发生变化,因为它取决于模型的预测。因此,该指标不能直接与model zoo中的"训练速度"相比较,后者是整个训练过程的平均速度。

主要结果

工具吞吐率(img / s) Detectron259 maskrcnn-benchmark51 tensorpack 50 mmdetection41 simpledet39 Detectron19 matterport/Mask_RCNN14

每个实现的链接:

• Detectron2:https://github.com/facebookresearch/detectron2/
• maskrcnn-benchmark:https://github.com/facebookresearch/maskrcnn-benchmark/
• tensorpack:https://github.com/tensorpack/tensorpack/tree/master/examples/FasterRCNN
• mmdetection:https://github.com/open-mmlab/mmdetection/
• simpledet:https://github.com/TuSimple/simpledet/
• Detectron:https://github.com/facebookresearch/Detectron
• matterport/Mask_RCNN:https://github.com/matterport/Mask_RCNN/

每个实现的详细信息：

• Detectron2: python tools/train_net.py --config-file configs/Detectron1-Comparisons/mask_rcnn_R_50_FPN_noaug_1x.yaml --num-gpus 8
• maskrcnn-benchmark: 通过sed -i ‘s/torch.uint8/torch.bool/g’ **/*.py使用commit 0ce8f6f与使其与最新的PyTorch兼容。然后,运行 python -m torch.distributed.launch --nproc_per_node=8 tools/train_net.py --config-file configs/e2e_mask_rcnn_R_50_FPN_1x.yaml我们观察到的速度比其model zoo快,这可能是由于软件版本不同所致。
• tensorpack: 在提交caafda,export TF_CUDNN_USE_AUTOTUNE=0, 然后运行 mpirun -np 8 ./train.py --config DATA.BASEDIR=/data/coco TRAINER=horovod BACKBONE.STRIDE_1X1=True TRAIN.STEPS_PER_EPOCH=50 --load ImageNet-R50-AlignPadding.npz
• mmdetection: commit4d9a5f,应用以下diff,然后运行 ./tools/dist_train.sh configs/mask_rcnn_r50_fpn_1x.py 8 我们观察到的速度比其model zoo快,这可能是由于软件版本不同所致。 (diff使其使用相同的超参数-单击展开) diff --git i/configs/mask_rcnn_r50_fpn_1x.py w/configs/mask_rcnn_r50_fpn_1x.py index 04f6d22..ed721f2 100644 --- i/configs/mask_rcnn_r50_fpn_1x.py +++ w/configs/mask_rcnn_r50_fpn_1x.py @@ -1,14 +1,15 @@ # model settings model = dict( type='MaskRCNN', - pretrained='torchvision://resnet50', + pretrained='open-mmlab://resnet50_caffe', backbone=dict( type='ResNet', depth=50, num_stages=4, out_indices=(0, 1, 2, 3), frozen_stages=1, - style='pytorch'), + norm_cfg=dict(type="BN", requires_grad=False), + style='caffe'), neck=dict( type='FPN', in_channels=[256, 512, 1024, 2048], @@ -115,7 +116,7 @@ test_cfg = dict( dataset_type = 'CocoDataset' data_root = 'data/coco/' img_norm_cfg = dict( - mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) + mean=[123.675, 116.28, 103.53], std=[1.0, 1.0, 1.0], to_rgb=False) train_pipeline = [ dict(type='LoadImageFromFile'), dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
• SimpleDet: 在commit9187a1时运行 python detection_train.py --config config/mask_r50v1_fpn_1x.py
• Detectron: 运行 python tools/train_net.py --cfg configs/12_2017_baselines/e2e_mask_rcnn_R-50-FPN_1x.yaml请注意,它的许多操作都在CPU上运行,因此性能受到限制。
• matterport/Mask_RCNN:在commit时3deaec,应用以下diff ,export TF_CUDNN_USE_AUTOTUNE=0, 然后运行 python coco.py train --dataset=/data/coco/ --model=imagenet请注意,此实现中的许多小细节可能与Detectron的标准不同。

(diff使其使用相同的超参数-单击展开)

  diff --git i/mrcnn/model.py w/mrcnn/model.py
  index 62cb2b0..61d7779 100644
  --- i/mrcnn/model.py
  +++ w/mrcnn/model.py
  @@ -2367,8 +2367,8 @@ class MaskRCNN():
        epochs=epochs,
        steps_per_epoch=self.config.STEPS_PER_EPOCH,
        callbacks=callbacks,
  -            validation_data=val_generator,
  -            validation_steps=self.config.VALIDATION_STEPS,
  +            #validation_data=val_generator,
  +            #validation_steps=self.config.VALIDATION_STEPS,
        max_queue_size=100,
        workers=workers,
        use_multiprocessing=True,
  diff --git i/mrcnn/parallel_model.py w/mrcnn/parallel_model.py
  index d2bf53b..060172a 100644
  --- i/mrcnn/parallel_model.py
  +++ w/mrcnn/parallel_model.py
  @@ -32,6 +32,7 @@ class ParallelModel(KM.Model):
      keras_model: The Keras model to parallelize
      gpu_count: Number of GPUs. Must be > 1
      """
  +        super().__init__()
      self.inner_model = keras_model
      self.gpu_count = gpu_count
      merged_outputs = self.make_parallel()
  diff --git i/samples/coco/coco.py w/samples/coco/coco.py
  index 5d172b5..239ed75 100644
  --- i/samples/coco/coco.py
  +++ w/samples/coco/coco.py
  @@ -81,7 +81,10 @@ class CocoConfig(Config):
    IMAGES_PER_GPU = 2

    # Uncomment to train on 8 GPUs (default is 1)
  -    # GPU_COUNT = 8
  +    GPU_COUNT = 8
  +    BACKBONE = "resnet50"
  +    STEPS_PER_EPOCH = 50
  +    TRAIN_ROIS_PER_IMAGE = 512

    # Number of classes (including background)
    NUM_CLASSES = 1 + 80  # COCO has 80 classes
  @@ -496,29 +499,10 @@ if __name__ == '__main__':
      # *** This training schedule is an example. Update to your needs ***

      # Training - Stage 1
  -        print("Training network heads")
      model.train(dataset_train, dataset_val,
            learning_rate=config.LEARNING_RATE,
            epochs=40,
  -                    layers='heads',
  -                    augmentation=augmentation)
  -
  -        # Training - Stage 2
  -        # Finetune layers from ResNet stage 4 and up
  -        print("Fine tune Resnet stage 4 and up")
  -        model.train(dataset_train, dataset_val,
  -                    learning_rate=config.LEARNING_RATE,
  -                    epochs=120,
  -                    layers='4+',
  -                    augmentation=augmentation)
  -
  -        # Training - Stage 3
  -        # Fine tune all layers
  -        print("Fine tune all layers")
  -        model.train(dataset_train, dataset_val,
  -                    learning_rate=config.LEARNING_RATE / 10,
  -                    epochs=160,
  -                    layers='all',
  +                    layers='3+',
            augmentation=augmentation)

    elif args.command == "evaluate":