ailab-cvc / unireplknet Goto Github PK

Problem with parameter in unireplknet.py line 338.
conv = nn.Conv2d(self.dwconv.in_channels, self.dwconv.out_channels, self.dwconv.kernel_size,
self.dwconv.padding, self.dwconv.groups, bias=True)

-> conv = nn.Conv2d(self.dwconv.in_channels, self.dwconv.out_channels, self.dwconv.kernel_size,
padding=self.dwconv.padding, groups=self.dwconv.groups, bias=True)

no suitable conversion function from "__half" to "float" existsfast_math.h(741): error: no suitable conversion function from "__half" to "float" exists

您好 图像分割里没有上传upernet_unireplknet_n_512_160k_ade20k.py这个文件,感觉与upernet_unireplknet_t_512_160k_ade20k.py的 paramwise_cfg={'decay_rate': 1, 'decay_type': 'layer_wise', 'num_layers': 9})有区别

Thank you for your excellent work!
when I use UniRepLKNet in my work (person reid), input_size of which is usually 384x128, it tends to be nan after stages3.
I think maybe because feature_map after the fifth downsample (downsample3) is too small (12x4), and kernel size in stages3 is 13x13.
Can you give me some possible resons and suggestions about it?
Thank you so much.

您好，请问代码中self.gamma的作用是什么？

133：enc_out = self.encoder(enc_out)
# enc_out's shape [350, 1000]
134：enc_out = enc_out.reshape(self.node_num,-1,self.d_model)
# however, node_num = 350 and d_model = 768
这导致无法转换：RunTimEerror：shape '[350,-1,768]' is invalid for input size 350000.
我发现github仓库中代码和点云代码中不同，是否是哪里需要进行改动，但是我不敢随意改动网络模型格式。
非常感谢！

是不是写错了（5，5，3，3）？

我使用了您提供的cutlass depthwise 做加速，在输入为(1536,90,180)上似乎比torch 的depthwise conv还要慢2-4倍，显卡是A100，驱动是11.8，请问是哪方面的问题？

When I execute the command

python -m torch.distributed.launch --nproc_per_node=2 main.py
--model unireplknet_xl --drop_path 0.3 --input_size 384
--batch_size 32 --lr 5e-5 --update_freq 2
--model_ema true --model_ema_eval true
--warmup_epochs 0 --epochs 20 --weight_decay 1e-8 --smoothing 0.3
--layer_decay 0.8 --head_init_scale 0.001 --cutmix 0 --mixup 0
--finetune unireplknet_xl_in22k_to_in1k_384_acc87.96.pth
--data_path UniRepLKNet/dataset
--output_dir UniRepLKNet/dataset/results

this code：
if args.layer_decay < 1.0 or args.layer_decay > 1.0:
num_layers = 12 # convnext layers divided into 12 parts, each with a different decayed lr value.
assert args.model in ['convnext_small', 'convnext_base', 'convnext_large', 'convnext_xlarge'],
"Layer Decay impl only supports convnext_small/base/large/xlarge"
assigner = LayerDecayValueAssigner(list(args.layer_decay ** (num_layers + 1 - i) for i in range(num_layers + 2)))
else:
assigner = None
run wrong！！！
Do you think it's a typo in the name that caused this?How can I resolve it? Thank you.

i use the unireplknet-s model to train my dataset and save a pth, but when i use the reparameterize.py or pytorch2onnx.py all occured the same false patameters
2023-12-02 15:19:52,935 - mmseg - WARNING - unexpected key in source state_dict: norm.weight, norm.bias, head.weight, head.bias

missing keys in source state_dict: stages.2.16.dwconv.bias, stages.2.26.dwconv.bias, stages.2.2.dwconv.bias, stages.2.5.dwconv.bias, stages.2.10.dwconv.bias, stages.2.4.dwconv.bias, stages.2.14.dwconv.bias, stages.2.8.dwconv.bias, stages.2.13.dwconv.bias, stages.2.19.dwconv.bias, stages.0.1.dwconv.bias, stages.2.23.dwconv.bias, stages.2.7.dwconv.bias, stages.2.20.dwconv.bias, stages.2.22.dwconv.bias, stages.0.2.dwconv.bias, stages.2.17.dwconv.bias, stages.0.0.dwconv.bias, stages.2.11.dwconv.bias, stages.2.1.dwconv.bias, stages.2.25.dwconv.bias

What a nice job!

I try to explore the ability of RepLK-series networks to understand point cloud with the approach of this websitePoint Cloud Understanding with UniRepLKNet. However, when I run this command bash tool/train_unireplknet.sh mv_unireplket-s ModelNet40 config/ModelNet40/multiview_UniRepLKNet-S.yaml, it reported an error cp: 无法获取'ModelNet40' 的文件状态(stat): 没有那个文件或目录.
Here is my file directory structure. I run the command at UniRepLKNet/Point.
What is the problem...?

.
├── config
│   ├── ModelNet40
│   │   └── multiview_UniRepLKNet-S.yaml
│   ├── Pretrain_ModelNet40
│   │   ├── pretrain-2048-ConvNeXt-L-1k.yaml
│   │   ├── pretrain-2048-ResNet-18.yaml
│   │   ├── pretrain-2048-ResNet-50.yaml
│   │   ├── pretrain-4096-ConvNeXt-L-1k.yaml
│   │   ├── pretrain-4096-ResNet-18.yaml
│   │   └── pretrain-4096-ResNet-50.yaml
│   ├── ScanObjectNN
│   │   ├── p2p_ConvNeXt-B-1k.yaml
│   │   ├── p2p_ConvNeXt-L-1k.yaml
│   │   ├── p2p_ConvNeXt-S-1k.yaml
│   │   ├── p2p_ConvNeXt-T-1k.yaml
│   │   ├── p2p_HorNet-L-22k-mlp.yaml
│   │   ├── p2p_ResNet-101.yaml
│   │   ├── p2p_ResNet-18.yaml
│   │   ├── p2p_ResNet-50.yaml
│   │   ├── p2p_Swin-B-1k.yaml
│   │   ├── p2p_Swin-S-1k.yaml
│   │   ├── p2p_Swin-T-1k.yaml
│   │   ├── p2p_ViT-B-1k.yaml
│   │   ├── p2p_ViT-S-1k.yaml
│   │   └── p2p_ViT-T-1k.yaml
│   └── ShapeNet
│   ├── p2p_ConvNeXt-B-1k.yaml
│   ├── p2p_ConvNeXt-L-1k.yaml
│   ├── p2p_ResNet-18.yaml
│   ├── p2p_ResNet-50.yaml
│   ├── p2p_Swin-B-1k.yaml
│   └── p2p_ViT-B-1k.yaml
├── data
│   ├── ModelNet40
│   │   ├── modelnet40_shape_names.txt
│   │   ├── modelnet40_test_8192pts_fps.dat
│   │   ├── modelnet40_test.txt
│   │   ├── modelnet40_train_8192pts_fps.dat
│   │   └── modelnet40_train.txt
│   └── ScanObjectNN
│   └── main_split
├── dataset
│   ├── data_transforms.py
│   ├── init.py
│   ├── modelnet.py
│   ├── scanobjectnn.py
│   ├── shapenet55.py
│   ├── shapenet.py
│   └── utils.py
├── Exp
│   └── config
│   └── ModelNet40
│   └── multiview_UniRepLKNet-S.yaml
│   └── mv_unireplket-s
│   ├── model
│   ├── result
│   │   ├── best
│   │   └── last
│   ├── test.py
│   ├── test.sh
│   ├── train_adaptor_lkv.py
│   └── train.sh
├── models
│   ├── deit
│   │   └── init.py
│   ├── init.py
│   ├── layers
│   │   ├── decoder.py
│   │   ├── encoder.py
│   │   ├── fusion.py
│   │   ├── head.py
│   │   ├── HorNet.py
│   │   ├── transformer.py
│   │   └── utils.py
│   ├── p2p_adaptor_lk.py
│   ├── p2p_adaptor.py
│   ├── p2p_joint.py
│   ├── p2p.py
│   └── p2p_raw.py
├── README.md
├── tool
│   ├── pretrain.py
│   ├── reproduce.sh
│   ├── test.py
│   ├── test.sh
│   ├── train_adaptor_lkv.py
│   ├── train_adaptor.py
│   ├── train.py
│   ├── train.sh
│   ├── train_unireplknet.sh
│   └── zero-shot.py
└── util
├── config.py
├── init.py
├── rotate.py
└── util.py

您好！无论是model = timm.create_model('unireplknet_l', num_classes=num_classes_of_your_task, in_22k_pretrained=True)还是您在点云示例代码中使用的model = timm.create_model('unireplknet_small')，都提示Unknown model 。我已排除网络问题，请问是什么情况？

Hi， When I try to install 19_large_depthwise_conv2d_torch_extension on the GPU service platform, I'm prompted that I don't have sufficient permissions to install it. Is there any alternative installation method or can it be installed via pip?

My train seems too slow comoared with ConvNeXt and SwinTransformer

I'm using a Windows 10 operating system, and I'm sure the system variable for CUDA_HOME is set to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.5.

But as you can see, it's always giving me an error and I can't go to the next step. I don't know what the problem is. Is it a problem with my windows system? Has anyone been able to successfully install it on windows?