dingxiaoh / repmlp Goto Github PK

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?

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！

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：

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。

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 ?!

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.

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？

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 

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！

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?

Can you provide the implementation for sementaic segmentation? I am wodering how you split the feature map to fit MLP.

thank you.

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?