pingolh/pytorch-hardnet

35% faster than ResNet: Harmonic DenseNet, A low memory traffic network

pytorch-hardnet's Issues

Hi, I trained on a new coco-like dataset with 7 classes but detection were all wrong

I try to detection trafficlights and trafficsigns, I only using coco config and it's labeling format just like coco, but after training the detection were all wrong.

the location loss around 3.8x, any insights on what's going wrong?

(training on coco actually works and get a very impressive results)

Misunderstanding for HarDBlock

Hi

I generated the HarDNet68ds.onnx, and visualized the network structure in Netron.
I make a similar figure, which is referenced by result of Netron, with your paper, and try to figure out the same structure.
Please see the belowing figure, it's almost similar.
Layer 8 connected with $8-2^n$ ($8/2^n$), so n=0,1,2,3
Hence, layer 8 connected with layers 7,6,4,0 (Blue line)
But I can't find the connection between "input layer" and "Next Block" (please see the orange lines).
I also trace the code, and can't find any evidence about the connection between input layer and Next block.
Can you help me figure out this problem?

Many thanks,
Tommy

论文上的疑惑

您好，我在阅读您的论文的时候有点疑惑：'CIO dominates the inference time only when the computational density, which is, the MACs over CIO (MoC) of a layer, is below a certain ratio that depends on platforms'
如果我理解的没有问题的话，这句话表示：computational density的相当于MoC低于某个ratio，而MoC的意思是 运算量/内存访问。是否意味着computational density是运算量较少的情况呢？

Could you provide the pretrained model?

Could you provide HarDNet39DS and HarDNet68DS pretrained model?

I think this function doesn't make it flatten.

Pytorch-HarDNet/hardnet.py

Line 10 in 12b140c

return x.view(x.data.size(0),-1)

Is it right to revise this function to (x.view(x.data.size(0), -1) → (x.view(-1))

How do you calculate MACs?

Hello, I am trying to replicate your results. Would you please tell me how did you calculate the GMACs?

License?

What is the license of this repository? Can FC-HarDNet-70 be used for commercial purpose?

problems in compute CIO

Nice work, thanks for the great idea of the CIO. When I compute MACs and CIO, I get MACs values the same as the paper.

but, I compute the CIO of the models as follows: hardnet39ds (9.8M), hardnet68ds (17.3M), mobilenetv2 (13.4M, same), resnet18 (4.7M, same) and resnet50 (21.8M). I am confused about these results.

Could you provide the CIO code? thanks a lot.