About install error under torch 1.4 or torch 1.5

[sunshinnnn]

when I use torch 1.4, I can install NMR successfully by pip install external/multiperson/neural_renderer
meanwhile, i could pip install http://jasonyzhang.com/phosa_data/neural_renderer_pytorch-1.1.3-cp36-cp36m-linux_x86_64.whl,

But I will encounter the error

AttributeError Traceback (most recent call last)
in ()
8 # import neural_renderer as nr
9 import detectron2
---> 10 from detectron2.data import MetadataCatalog
11 from detectron2.structures import Boxes, BitMasks
12 from detectron2.utils.visualizer import Visualizer as PointRendVisualizer

~/anaconda3/envs/phosa2/lib/python3.6/site-packages/detectron2/data/init.py in ()
2 from . import transforms # isort:skip
3
----> 4 from .build import (
5 build_batch_data_loader,
6 build_detection_test_loader,

~/anaconda3/envs/phosa2/lib/python3.6/site-packages/detectron2/data/build.py in ()
10 from termcolor import colored
11
---> 12 from detectron2.structures import BoxMode
13 from detectron2.utils.comm import get_world_size
14 from detectron2.utils.env import seed_all_rng

~/anaconda3/envs/phosa2/lib/python3.6/site-packages/detectron2/structures/init.py in ()
4
5 from .instances import Instances
----> 6 from .keypoints import Keypoints, heatmaps_to_keypoints
7 from .masks import BitMasks, PolygonMasks, rasterize_polygons_within_box, polygons_to_bitmask
8 from .rotated_boxes import RotatedBoxes

~/anaconda3/envs/phosa2/lib/python3.6/site-packages/detectron2/structures/keypoints.py in ()
4 import torch
5
----> 6 from detectron2.layers import interpolate
7
8

~/anaconda3/envs/phosa2/lib/python3.6/site-packages/detectron2/layers/init.py in ()
9 from .wrappers import BatchNorm2d, Conv2d, ConvTranspose2d, cat, interpolate, Linear, nonzero_tuple
10 from .blocks import CNNBlockBase
---> 11 from .aspp import ASPP
12
13 all = [k for k in globals().keys() if not k.startswith("_")]

~/anaconda3/envs/phosa2/lib/python3.6/site-packages/detectron2/layers/aspp.py in ()
1 # Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
2
----> 3 import fvcore.nn.weight_init as weight_init
4 import torch
5 from torch import nn

~/anaconda3/envs/phosa2/lib/python3.6/site-packages/fvcore/nn/init.py in ()
1 # Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
----> 2 from .activation_count import activation_count
3 from .flop_count import flop_count
4 from .focal_loss import (
5 sigmoid_focal_loss,

~/anaconda3/envs/phosa2/lib/python3.6/site-packages/fvcore/nn/activation_count.py in ()
7 import torch.nn as nn
8
----> 9 from .jit_handles import generic_activation_jit, get_jit_model_analysis
10
11

~/anaconda3/envs/phosa2/lib/python3.6/site-packages/fvcore/nn/jit_handles.py in ()
72 # pyre-fixme[24]: Generic type typing.Callable expects 2 type parameters.
73 ops_handles: typing.Dict[str, typing.Callable],
---> 74 ) -> typing.Tuple[typing.Counter[str], typing.Counter[str]]:
75 """
76 Given a model, the inputs and the handles for each operation, return the

AttributeError: module 'typing' has no attribute 'Counter'

When I use torch 1.5,
import detectron2
from detectron2.data import MetadataCatalog
from detectron2.structures import Boxes, BitMasks
from detectron2.utils.visualizer import Visualizer as PointRendVisualizer
these import are all ok,
however, i can not install NMR successfully anymore.

Is there any way to solve such problems, whatever the import detectron2 error in torch 1.4 or NMR install error in torch 1.5

[sunshinnnn]

I've sloved install NMR problem under torch 1.5 by change AT_CHECK to TORCH_CHECK in all .cpp.

[monajalal]

@sunshinnnn should we use PyTorch 1.5 instead of 1.4?

[jasonyzhang]

For people in the future who have this problem,

• NMR out of the box only works with torch <=1.4. detectron 0.2.1 is the last version with torch 1.4 support. I've pinned these versions in the colab notebook.
• If you want to use torch >1.4, the changes listed by @sunshinnnn will allow NMR to work with later versions of torch. This fix should work for future releases of detectron