luyao777/hbp-pytorch

Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition reimplement in Pytorch.

Code is cheap, Show me the 🍌. XD

🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌 🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌 🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌 🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌 🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌🍌

🍎: Who Am I

• Name: Yao Lu
• Email: [email protected]
• Location: Shanghai in China
• Educational Experience：
  • Undergraduate(2014.09-2018.06): Computer Science and Technology, University of Electronic Science and Technology of China
  • Postgraduate(2018.9-2021.06): Computer Application Technology, Renmin University of China
• Employment Experience:
  1. Tencent Holdings Limited
  2. Pinduoduo Inc.

🐍 Commit Snake

⏰ Coding Time

From: 14 September 2021 - To: 03 May 2024

Total Time: 1,236 hrs 23 mins

Python              755 hrs 42 mins ███████████████▒░░░░░░░░░   61.12 %
JSON                142 hrs 23 mins ███░░░░░░░░░░░░░░░░░░░░░░   11.52 %
SQL                 117 hrs 5 mins  ██▒░░░░░░░░░░░░░░░░░░░░░░   09.47 %
Other               68 hrs 33 mins  █▒░░░░░░░░░░░░░░░░░░░░░░░   05.55 %
Text                63 hrs 11 mins  █▒░░░░░░░░░░░░░░░░░░░░░░░   05.11 %

🧮 Github Stats

📑 Recent Blog

• NameError: name 'tf' is not defined - NoMornings - 2023-03-29
• ValueError: The two structures don't have the same sequence length. Input structure has length 4, while shallow structure has length 3. - NoMornings - 2023-03-14
• 量化交易基础 - 012 - 检验中的假设条件 - NoMornings - 2023-03-11
• Tf2 ValueError: A merge layer should be called on a list of inputs - NoMornings - 2023-03-02
• 量化交易基础 - 011 - 样本外检验 - NoMornings - 2023-03-01
• 量化交易基础 - 010 - 拟合 - NoMornings - 2023-02-26
• 量化交易基础 - 009 - 检验 （2） - NoMornings - 2023-02-24
• 量化交易基础 - 007 - 研究 - NoMornings - 2023-02-23
• 量化交易基础 - 008 - 检验 （1） - NoMornings - 2023-02-22
• 量化交易基础 - 006 - 数据类型、来源和清洗 - NoMornings - 2023-02-22

⛰️ Streak Stats

hbp-pytorch's Issues

X = proj_1 * proj_2 it may be wrong

def hbp(self,conv1,conv2):
N = conv1.size()[0]
proj_1 = self.bilinear_proj(conv1)
proj_2 = self.bilinear_proj(conv2)
assert(proj_1.size() == (N,8192,28,28))
X = proj_1 * proj_2
assert(X.size() == (N,8192,28,28))
X = torch.sum(X.view(X.size()[0],X.size()[1],-1),dim = 2)
X = X.view(N, 8192)
X = torch.sqrt(X + 1e-5)
X = torch.nn.functional.normalize(X)
return X

in HBP , this line :X = proj_1 * proj_2. should be. self.bap = BAP(pool='GAP')?
because this is just dot, it should be binear pooling ，You also have to multiply between the channels。
I think that's the right way to think about it，Thank you for the code 👍

File "/torch_yolov4/HBP-pytorch/cub200.py", line 89, in __init__

os.path.join(self._root, 'processed/test.pkl'), 'rb'),encoding='iso-8859-1')

_pickle.UnpicklingError: pickle data was truncated
how can i solve it???thanks

Single projection for all feature maps?

I think you should use three individual linear transforms to project Conv5_1, Conv5_2 and Conv5_3 respectively according to the original paper.

A question about the architecture

whether need to add another linear transformation after the concatenate the bilinear feature?Because the formulation in the paper is z=P^T concat(U^T x * V^T y, ...)

HBP定义中的代码我有点不理解

    proj_1 = self.bilinear_proj(conv1)
    proj_2 = self.bilinear_proj(conv2)
    assert(proj_1.size() == (N,8192,28,28))
    X = proj_1 * proj_2
    assert(X.size() == (N,8192,28,28)) 

这里不应该是两个卷积层外积操作吗？出来不应该是[N,8192,8192]吗？为什么是(N,8192,784))

关于代码运行的错误

resnet50

how can I change pre-train model with resnet50

帅哥 加油啊 早日达到论文精度 我们为你喝彩

The matrix U and V in implementation

Thanks for implementing the paper! I am confused with the matrix U and V. These two matrices in the paper seem to be two different matrices. But in the implementation, they are initilialized as the same matrix as:
self.bilinear_proj = torch.nn.Sequential(torch.nn.Conv2d(512,8192,kernel_size=1,bias=False),
torch.nn.BatchNorm2d(8192),
torch.nn.ReLU(inplace=True))

proj_1 = self.bilinear_proj(conv1)
proj_2 = self.bilinear_proj(conv2)

NameError: name 'options' is not defined

---

main()

File "HBP_fc.py", line 270, in main
manager = HBPManager(options, path)
File "HBP_fc.py", line 94, in init
self._options = options
NameError: name 'options' is not defined