labsaint / spd-conv Goto Github PK

I validated complexity of your 's' the model (YOLOv5-SPD-s) mentioned in your paper, but the results show that the flops of this model are extremely huge compared with same input size of yolov5-s. the detailed information are:
YOLOv5-SPD-s: Model Summary: 277 layers, 8771389 parameters, 8771389 gradients, 33.9 GFLOPs
YOLOv5-s: Model Summary: 270 layers, 7235389 parameters, 7235389 gradients, 16.5 GFLOPs
so, I wonder what the value of your idea in paper which is intended to improve the detection performance of some small targets at the cost of increase of model computation complexity?

I tested SPD-yolov5s and yolov5s in RK3588 which is a platform using NPU，but I find that yolov5s with 30 FPS was faster than SPD-yolov5s with 7 FPS，so what's the reason or how to improve it？thank you！

请问这个支持tensorrt吗

!python detect.py --weight runs/train/yolov7-carspd9/weights/best.pt --conf 0.5 --img-size 224 --device 1 --source 9.jpg
File "detect.py", line 196, in
detect()
File "detect.py", line 39, in detect
model = TracedModel(model, device, opt.img_size)
File "/root/code/yolov7/utils/torch_utils.py", line 362, in init
traced_script_module = torch.jit.trace(self.model, rand_example, strict=False)
File "/root/anaconda3/envs/cuda11_1/lib/python3.8/site-packages/torch/jit/_trace.py", line 735, in trace
return trace_module(
File "/root/anaconda3/envs/cuda11_1/lib/python3.8/site-packages/torch/jit/_trace.py", line 952, in trace_module
module._c._create_method_from_trace(
File "/root/anaconda3/envs/cuda11_1/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
return forward_call(*input, **kwargs)
File "/root/anaconda3/envs/cuda11_1/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1039, in _slow_forward
result = self.forward(*input, **kwargs)
File "/root/code/yolov7/models/yolo.py", line 599, in forward
return self.forward_once(x, profile) # single-scale inference, train
File "/root/code/yolov7/models/yolo.py", line 625, in forward_once
x = m(x) # run
File "/root/anaconda3/envs/cuda11_1/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
return forward_call(*input, **kwargs)
File "/root/anaconda3/envs/cuda11_1/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1039, in _slow_forward
result = self.forward(*input, **kwargs)
File "/root/code/yolov7/models/common.py", line 282, in forward
return torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1)
RuntimeError: torch.cat(): Sizes of tensors must match except in dimension 1. Got 4 and 3 in dimension 2 (The offending index is 1)

Hi，i think this idea is amazing. If i want to extend it to 3d to solve some question, such as video, should i add a dimension or just use it on every feature map in all channel

Hello!
Your research is very helpful to me. I ran your code, but there is an error. Do you know what is going on?thank you for your reply

Hello, thanks for your significant contribution. I have a question about space_to_depth fuction.
When space_to_depth is used in model, the channel will be expand to 4X. For the next conv, the filter become 4X bigger.
My question is: Need I add a new conv to sqeenze the channel to 2X , because the stride=2 conv often expand channel to 2X or bigger than X,
Can you give me some advise?
Thanks again!!!

Why is there everything twice in your requirements file for the yolov5?

Hi,

It's not clear to me what is the difference between SPD-Conv and Conv with stride S and kernel size S. I.e. this is what many transformer papers (e.g. Swin) use as their "patchify" stem.

It's also used in their downsampling layer. Although, they formulate it without using Conv2D (to avoid NHWC to NCHW reshape) but it is conceptually equivalent to Conv2D with stride 2 and kernel size 2.

Can I use SPD based on loading a ImageNet-pretrained model?