How to correctly modify https://github.com/eriklindernoren/PyTorch-YOLOv3 to use https://github.com/huawei-noah/AdderNet ?

The following colab ipynb notebook is what I have so far with the helps of others:

https://colab.research.google.com/drive/1VCafwykgNKAO6144LssBFFy0TmruDNSE#scrollTo=W3e-WcVxnKfs

How to solve the error on this models.py file ?

Namespace(batch_size=8, class_path='data/coco.names', conf_thres=0.001, data_config='config/coco.data', img_size=416, iou_thres=0.5, model_def='config/yolov3.cfg', n_cpu=8, nms_thres=0.5, weights_path='weights/yolov3.weights')
Traceback (most recent call last):
  File "test.py", line 84, in <module>
    model.load_darknet_weights(opt.weights_path)
  File "/content/PyTorch-YOLOv3/models.py", line 321, in load_darknet_weights
    num_w = conv_layer.weight.numel()
  File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 576, in __getattr__
    type(self).__name__, name))
AttributeError: 'adder2d' object has no attribute 'weight'

Hi, I am trying to train your addernet, but it returns me one runtime error, which I supposed attributes to .continuous() function or some other uncommon operations used in your adder.py.

Could you help to solve this issue?

I try to train the addernet using resnet18 for ImageNet from scratch, with 4 1080Ti cards, but it just occupies too much memory that i could only set the batch_size to 16, and it's also too too slow..

For comparision, I have tired to replace the adder filters with normal conv filters and the 4 gpu cards could load 128 batch size. Did i setup wrong, or is that the normal case currently for addernet?

Have you guys tried to train with ImageNet?

Hi, Hanting Chen,

I tried the same training setting (Poly LR schedule with 0.9 power; 400 epochs; 256 batch size), and I can only get 89.16% accuracy when training resnet20 on CIFAR-10, while the reported accuracy is 91.84%.

Could you provide more training details and help to explain the gap? thanks.

Hi, I would like to know why you defined the L2-distance as in Equation (14) appendix. Doesn't L2-distance need a square root outside the summations? And also, I would like to know how the corresponding partial derivative of L2-distance in Equation (5) comes? Thanks.

I replace adder.adder2d with torch.nn.Conv2d, and replace torch.cdist with my_cdist. train my network, the speed of the new model slower than the old at least 6 times. I'm confused.

@torch.jit.script
def my_cdist(x1, x2):
    x1_norm = x1.pow(2).sum(dim=-1, keepdim=True)
    x2_norm = x2.pow(2).sum(dim=-1, keepdim=True)
    res = torch.addmm(x2_norm.transpose(-2, -1), x1, x2.transpose(-2, -1), alpha=-2).add_(x1_norm)
    res = res.clamp_min_(1e-30).sqrt_()

    return res

您好，冒昧打扰。 之前拜读了您团队的Pretrained ImageProcessingTransformer， 受益匪浅。在尝试复现您的论文，但是在Dataloader和Datasets的写过程中遇到了问题。不知如何裁剪图片， 如何对数据进行预处理。论文中提到Overlap10Pix， 但是边缘部分如何处理！
若能分享解惑，不甚感激。 感谢感谢，期待回复。

Hi,

We recently noticed that you update the comparison with ShiftAddNet in your readme.

However, the chosen accuracy of ShiftAddNet (i.e., 85.10%) is wrongly from our adaption experiments, for which we only use half the CIFAR-10 dataset to pre-train. Our accuracy (ResNet-20 on CIFAR-10) should be 89.32%, which is also trained with a different setting and different implementation (e.g., only 160 epochs).

Could you help to update the comparison results?

Best regards, Haoran You

您好 我们参考 https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html 利用自己的数据集对网络进行微调，我们遇到了困难 普通的resnet通过fc层输出in_featrues，进行微调训练，但是addernet中resnet的fc是conv2d，我不知道如何得到fc层的输入，或者说avgpool的输出。 期待您的回答，非常感谢。

First I really appreciate your work . I met some problems when I try to run python main.py. File "main.py", line 117, in main() File "main.py", line 112, in main train_and_test(e) File "main.py", line 105, in train_and_test train(epoch) File "main.py", line 80, in train loss.backward() File "/home/anaconda3/lib/python3.6/site-packages/torch/tensor.py", line 102, in backward torch.autograd.backward(self, gradient, retain_graph, create_graph) File "/home/anaconda3/lib/python3.6/site-packages/torch/autograd/init.py", line 90, in backward allow_unreachable=True) # allow_unreachable flag RuntimeError: function adderBackward returned an incorrect number of gradients (expected 2, got 1)

你好，我想固定加法核进行加法替代卷积的操作。 我把生成的核 tensor 直接放在 output = adder2d_function(x, self.adder, self.stride, self.padding) 的 self.adder 中，发现效果和固定卷积核操作相差甚远，不知道是因为加法核和卷积核的分布差异导致还是其他原因，想咨询一下作者有固定加法核进行加法操作的方法吗？谢谢

Hi. I have a CNN architecture that I trained on CIFAR-10 with and without AdderNet. I could reach an accuracy of over %80 without AdderNet but when I used AdderNet it got stuck at %10 accuracy. Is there anything wrong with my implementation? All I did was to replace nn.conv2D with adder.adder2d. Isn't it supposed to work like this? How do you suggest I should migrate from a simple CNN architecture to AdderNet? Thank you!

您好，最近我想试试部署adder到mobilenet v2网络里，但是遇到了点问题，请问可以帮我解答吗。

我用这里的源码（训练参数对齐）跑resnet，精度和report里差不多。但是当我把mobilenet v2的pointwise层（depthwise要改为群卷积，准备等pointwise没问题再改）替换成adder-conv2d后，精度会有2%的下降。

我没有替换开头和末尾两层。请问精度下降2%会是什么原因呢？ 会不会是残差的问题，需要每一层都有残差吗（mobilenet v2若是stride=2或维度变化，就没有残差了）？

你好，我在复现的过程中，有几个问题想请教：

1. 由于论文中提到需要使用shortcut来实现恒等映射，所以网络中是不是没有独立的加法层（每个加法层都包含于一个残差块）,因此是否意味着原始EDSR中的一些独立卷积层，都要用一个加法残差块来替代。
2. 在您的论文中提到：the above function （power activation function） can be easily embedded into the conventional ReLU in any SISR models. 这个具体实现思路是什么？我可以直接在ReLU后面接一个power activation function吗？