leaplabthu / acmix Goto Github PK

你好，感谢你们杰出的工作，我想使用你们的ACmix模型来训练自己的数据，但是直接从0开始，需要大量的计算，但我在mindspore找不到对应的模型，请问能否提供一个在imagenet下训练好的模型？谢谢！

我使用的是configs/acmix_swin_tiny_patch4_window7_224.yaml

When will you release the pre-train model of ResNet?

I noticed that this part of the paper is '3N×k^2N'，but in the code is '3N×k^2'. Is there a mistake here or am I understanding it wrong?

使用YOLOv7结合ACmix，出现如下报错：

`Traceback (most recent call last):
  File "/home/liu/桌面/zwx/YOLOv7-main/train.py", line 613, in <module>
    train(hyp, opt, device, tb_writer)
  File "/home/liu/桌面/zwx/YOLOv7-main/train.py", line 415, in train
    results, maps, times = test.test(data_dict,
  File "/home/liu/桌面/zwx/YOLOv7-main/test.py", line 110, in test
    out, train_out = model(img, augment=augment)  # inference and training outputs
  File "/home/liu/anaconda3/envs/yolo-torch2/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
  File "/home/liu/桌面/zwx/YOLOv7-main/models/yolo.py", line 320, in forward
    return self.forward_once(x, profile)  # single-scale inference, train
  File "/home/liu/桌面/zwx/YOLOv7-main/models/yolo.py", line 346, in forward_once
    x = m(x)  # run
  File "/home/liu/anaconda3/envs/yolo-torch2/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
  File "/home/liu/桌面/zwx/YOLOv7-main/models/common.py", line 530, in forward
    pe = self.conv_p(position(h, w, x.is_cuda))
  File "/home/liu/anaconda3/envs/yolo-torch2/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
  File "/home/liu/anaconda3/envs/yolo-torch2/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 463, in forward
    return self._conv_forward(input, self.weight, self.bias)
  File "/home/liu/anaconda3/envs/yolo-torch2/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 459, in _conv_forward
    return F.conv2d(input, weight, bias, self.stride,
RuntimeError: Input type (float) and bias type (c10::Half) should be the `same`

RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.cuda.HalfTensor) should be the same

hello，I would like to learn the code based on the objection detection

Can I use ACmix in test_bottleneck.py to replace Conv 3*3 in other network directly?

在position coding时，输入固定float，在yolo训练val时会自动进行半精度训练，将position()输出类型转换成input就好
position(......).to(x.dtype)

请问作者可以放出关于目标检测实现的代码吗

When you will release the pre-trained models of ResNet, SAN, and PVT?

Hi！

Could you please share which tool was used to create Figure 1 in your paper ?

Thank you,

ResNet/test_bottleneck.py
line 101
original:
f_conv = f_all.permute(0, 2, 1, 3).reshape(x.shape[0], -1, x.shape[-1], x.shape[-1])
but I think it should be:
f_conv = f_all.permute(0, 2, 1, 3).reshape(x.shape[0], -1, x.shape[-2], x.shape[-1])
to maintain the shape of input height and width.
I do not know if it is correct. Looking forward to your reply. Thanks.

请问这个投影部分为什么用三个重复的1*1操作呀，卷积的过程没太看明白，可以帮我解释一下吗？谢谢

Hello, I would like to ask whether the code based on ResNet will be made public

The ACmix module consumes too much gpu resources

Hello, if I want to transform other networks, can I directly use acmix instead of large volume core？

Hi! Thanks for your great work! I think squeeze should be replaced by unsqueeze

我自己测试了一下用nn.Conv2d(16, 64, 1)，输入大小是(1, 16, 224, 224)，这个参数量只有1088，但是如果用ACmix得到的参数量是8604，这差了快8倍了，但是文章说 “同时与纯卷积或self-attention相比具有最小的计算开销”，好像没有体现，这是咋回事啊？

Why does the tensor of fp16 appear in acmix？