dingxiaoh / repmlp Goto Github PK
View Code? Open in Web Editor NEWRepMLPNet: Hierarchical Vision MLP with Re-parameterized Locality (CVPR 2022)
License: MIT License
RepMLPNet: Hierarchical Vision MLP with Re-parameterized Locality (CVPR 2022)
License: MIT License
您好,请问下,您们的RepMLP用于人脸识别时是否加CosFace、ArcFace之类的?输入只用了9696,是否用过112112?
谢谢!
https://github.com/DingXiaoH/RepMLP/blob/main/repmlpnet.py#L49
def forward(self, inputs):
x = F.adaptive_avg_pool2d(inputs, output_size=(1, 1))
x = self.fc1(x)
according to the paper, it may should be (h, w)?
Can you add requirements ?!
关于在单位阵上做卷积,单位阵里有很多0啊,局部信息不会丢失嘛,(还是我理解错了)
比如这段代码里:
Line 107 in 55c7677
在这个上面做卷积,I的形状是(9,1,3,3),每个(3,3)中只有一个值不为0,卷积后reshape回去,也只有对角元上不为0,这样做(9,9)x(9,1)的矩阵乘的话,相当与给(3,3)里的每一个元素乘了一个单独的值,也不是卷积吧。
self.conv_embedding = conv_bn_relu(in_channels, channels[0], kernel_size=patch_size, stride=patch_size, padding=0)
The first embedding layer is still CNN without conversion to MLP finally, have you tried to convert this embedding layer to MLP as well?
This design of repmlp-resnet is different from the lastest repmlpnet, and it shows great face recognition accuracy.
why not keep repmlp-resnet in this repo?
把merge CNN into FC是完全等价的吗?
还是像formulation部分说的一种输入输出维度相同的替换?
如果是完全等价,可否在训练阶段也去掉CNN?
您好,想问下你们的RepMLP pretrain模型 训练的时候是跑了多久,用了几张卡呢?谢谢!
Can you provide the implementation for sementaic segmentation? I am wodering how you split the feature map to fit MLP.
thank you.
丁学长您好,之前看到您在文章中写到使用Cifar数据集来进行一致性的验证。
我想复现一下这个结果,但是Readme中好像并没有告知关于Cifar训练集和测试集的切分细节,以及它的输入维度。
这个能告知一下吗?
Thank you very much for proposing an excellent model and sharing it publicly. Also congratulations on the publication of your results in CVPR. Since I want the RepMLP model should be on one-dimensional data, that is, the input is only [B, C, H]. Would like to ask if it is possible to provide a RepMLP model for such one-dimensional data?
Light Block is not fast as the paper says
def test(network, p=True):
x = torch.ones(128, 3, 224, 224).cuda()
model = network.cuda()
if p: print(model)
model.eval()
with torch.no_grad():
# warm iters
for i in range(20):
y = model(x)
# inference test
iters = 50
start = time.time()
for i in range(iters):
y = model(x)
end = time.time()
print((end-start)/iters, 's')
print(y.shape)
if __name__ == "__main__":
torch.backends.cudnn.benchmark=True
test(create_RepMLPRes50_Base_224(deploy=True), False)
test(create_RepMLPRes50_Light_224(deploy=True), False)
test(create_RepMLPRes50_Bottleneck_224(deploy=True), False)
with Titan XP
Base: 17.1 ms
Light Block: 16.9 ms
Bottleneck: 18.6 ms
Hi, I try python repmlpnet.py
, and set the size of x at [100, 3, 224, 224]. But I find the deviation between the new_y and origin_y is 0.0019, which is not negligible. I wonder why it happens, thank you in advance.
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.