yukang2017 / stitcher Goto Github PK

I have successfully installed everything without error following INSTALL.md. But when I run the train_net.py, it raises an error

Traceback (most recent call last): Traceback (most recent call last): File "tools/train_net.py", line 15, in <module> File "tools/train_net.py", line 15, in <module> from maskrcnn_benchmark.data import make_data_loader File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/data/__init__.py", line 2, in <module> from maskrcnn_benchmark.data import make_data_loader File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/data/__init__.py", line 2, in <module> from .build import make_data_loader from .build import make_data_loader File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/data/build.py", line 11, in <module> File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/data/build.py", line 11, in <module> from . import datasets as D from . import datasets as D File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/data/datasets/__init__.py", line 3, in <module> File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/data/datasets/__init__.py", line 3, in <module> from .coco import COCODataset File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/data/datasets/coco.py", line 6, in <module> from .coco import COCODataset File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/data/datasets/coco.py", line 6, in <module> from maskrcnn_benchmark.structures.segmentation_mask import SegmentationMask File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/structures/segmentation_mask.py", line 5, in <module> from maskrcnn_benchmark.structures.segmentation_mask import SegmentationMask File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/structures/segmentation_mask.py", line 5, in <module> from maskrcnn_benchmark.layers.misc import interpolate File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/layers/__init__.py", line 10, in <module> from maskrcnn_benchmark.layers.misc import interpolate File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/layers/__init__.py", line 10, in <module> from .nms import nms File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/layers/nms.py", line 3, in <module> from .nms import nms File "/mnt/xfs1/home/liangtian/project/maskrcnn-benchmark/maskrcnn_benchmark/layers/nms.py", line 3, in <module> from maskrcnn_benchmark import _C ImportError: libcudart.so.9.0: cannot open shared object file: No such file or directory

It seems like that I need cuda9.0?
But I have successfully installed
pytorch-nightly 1.0.0.dev20190328 py3.6_cuda10.0.130_cudnn7.4.2_0
pytorch 1.4.0 py3.6_cuda10.0.130_cudnn7.6.3_0
cudatoolkit 10.0.130 0
`

nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2018 NVIDIA Corporation
Built on Sat_Aug_25_21:08:01_CDT_2018
Cuda compilation tools, release 10.0, V10.0.130
`

All of the information about cuda version is CUDA10.0, I can't understand why the maskrcnn_benchmark has to find the libcudart.so.9.0, does it only support CUDA9.0?

As coco's definition, small objects are those whose area is smaller than 32*32.
If you set 1024 as the threshold, what are the others' thresholds, like medium and large? I'm curious about it for the Figure 5 in your paper.

你好，在你的代码里好像没做分布式训练的多卡ratio_small合并计算的代码。

看文中对coco数据集的small，medium，large目标的标注个数及其在图像上的分布，想知道怎么做到的

FOCS has been accepted by maskrcnn_benchmark, but I dont know how to use stitcher in FCOS. Can I just copy FCOS's config to this project?

作者的思路确实清奇。采用双dataloader实现，根据loss设定的反馈规则来确定下一个batch用正常数据还是stitch数据。mark一下，顺便给作者点个赞！！！