yxlijun / pelee.pytorch Goto Github PK

I would like to train the model on higher resolution such as 604, what changes would you recommend? At the moment you have put a check to train only on 304 input size.

在denselayer中
out3 = self.branch2c(out2)
应该改为：
out2 = self.branch2c(out2)

hey thanks for providing this , going to try it out.

can you provide a license file with the repo please ?

thanks,
Atti

原文denseblock里面不应该重复transition层吧，你的循环生成denseblock里面重复了transition，在用你的代码训练目标检测的时候效果不是很好，我改yolo做的就不是很成功。不知道你对这个模块是怎么理解的？重复transition能提高精度么？

When I disable GPU in configs/Pelee_VOC.py test_cfg.cuda = false. I got RuntimeError: expected device cpu but got device cuda:0.

So some model is using GPU. But I am not able to figure which one.

Hi, first thank you your work, it is very good. I have a datasets about one thousand images, the resolution is 1920x1080. I have use your source to train the datasets about 12 hours, but the model can not find the object a little. It is a mess of bounding boxes everywhere in the image. I did not modify the train parameters except the class number and class name, can you give me some advice? thanks very much!!!

@yxlijun
hai, sir, thank you your good job!
how many iters did you use when you trained the voc dataset, and get Pelee_VOC.pth???
1、160000? like ./configs/Pelee_VOC.py ?
2、and with 8 num_works, 32 per_batch_size???
3、other hyper-parameter is the same with ./configs/Pelee_VOC.py???

hi~~~
i am a new man in this. your repository is amazing for me, but i have a problem with the dataset.maybe its not a big problem for u.
i already download the Pelee_VOC.pth and put it to weights folder, and i follow the steps to download the dataset VOC2007, and it looks like this.

and i run the command
python demo.py --dataset VOC\VOC2007 --config ./configs/Pelee_VOC.py --trained_model ./weights/Pelee_VOC.pth --show
but it comes out the

i dont whats wrong? could u help me? thx~~~

hello，i test the FPS on nvidia-p100, but it only get about 30FPS，it is too low, so i want to ask that whether it is normal?

Can anyone download the pre training weights in the link？

class _DenseLayer(nn.Module):
"""docstring for _DenseLayer"""

def __init__(self, num_input_features, growth_rate, bottleneck_width, drop_rate):
    super(_DenseLayer, self).__init__()
    growth_rate = growth_rate // 2
    inter_channel = int(growth_rate * bottleneck_width / 4) * 4

    if inter_channel > num_input_features / 2:
        inter_channel = int(num_input_features / 8) * 4
        print('adjust inter_channel to ', inter_channel)

    self.branch1a = conv_bn_relu(
        num_input_features, inter_channel, kernel_size=1)
    self.branch1b = conv_bn_relu(
        inter_channel, growth_rate, kernel_size=3, padding=1)

    self.branch2a = conv_bn_relu(
        num_input_features, inter_channel, kernel_size=1)
    self.branch2b = conv_bn_relu(
        inter_channel, growth_rate, kernel_size=3, padding=1)
    self.branch2c = conv_bn_relu(
        growth_rate, growth_rate, kernel_size=3, padding=1)

def forward(self, x):
    out1 = self.branch1a(x)
    out1 = self.branch1b(out1)

    out2 = self.branch2a(x)
    out2 = self.branch2b(out2)
    out3 = self.branch2c(out2)

    out = torch.cat([x, out1, out2], dim=1)  // torch.cat([x, out1, out3], dim=1)
    return out

Here is the full error

The output of nvidia-smi

----------------------------------------------------------------------
|                       Pelee Training Program                       |
----------------------------------------------------------------------
free(): invalid pointer
Aborted (core dumped)

This happens when I try to train the Pelee on custom dataset (VOC format).
I think it has something to do with free(), but I'm not sure what to change.

请教下，这个怎么跑摄像头demo

What should I change?

When I train with multi gpu, there always comes the erorr:
OSError: [Errno 24] Too many open files
Do you know what's going on?

When I ran the train command, there was an error:
File "./Pelee.Pytorch/configs/CC.py", line 14, in
from addict import Dict
ImportError: No module named addict
I can't deal with this problem since I can't find the module addict. Could you please tell me what's going on? Thanks!

from . import _mask
ImportError: cannot import name '_mask'