cypw / dpns Goto Github PK

hi, when i run score.py, there is a such mistake:
Traceback (most recent call last):
File "D:/PycharmProjects/DPNs-master/score.py", line 110, in
speed = score(metrics=metrics, **vars(args))
File "D:/PycharmProjects/DPNs-master/score.py", line 42, in score
inter_method=2 # bicubic
File "D:\Python35\lib\site-packages\mxnet\io.py", line 725, in creator
ctypes.byref(iter_handle)))
OSError: exception: access violation reading 0x0000000000000090
can you help me ,thanks a lot.

How to trian my dataset by DPNs? Do you have some detailed description?
Thank you for help.

line 30 in DPNs/settings/symbol_dpn.py

c1x1_w = BN_AC_Conv( data=data_in, num_filter=(num_1x1_c+2*inc), kernel=( 1, 1), stride=(key_stride, key_stride), name=('%s_c1x1-w(s/%d)' %(name, key_stride)), pad=(0, 0))

How to train a model to detect objects and get results？ May I use the pretrained models？

"Please let me know if any other resarchers have proposed exactly the same technique."

you may want to search about "mix pooling". some examples are:

[1] "Mixed Pooling for Convolutional Neural Networks"- Dingjun Yu, Hanli Wang, Peiqiu Chen, and Zhihua Wei

[2] "Generalizing Pooling Functions in Convolutional Neural Networks: Mixed, Gated, and Tree" - Chen-Yu Lee, Patrick W. Gallagher, Zhuowen Tu

Can you share your train log about DPNs on imagenet

I have seen the code that the validated datavaridat is set by the parameter of ‘--data-val'. And the default value is '/tmp/val.rec‘. But I don't know what and where this file is. And if I want to test a image, how should I do?
Thanks very much!

Hi,
Thank for your sharing. But I want to use DPn-98 or Dpn-92 pretrained model on Place365-Standard dataset,
Can you give me a site?
Thank you very much

I try to train DPNs by modifying the score.py, but it doesn't work, the Train-accuracy is always 0.

Hi，I have attempted many times to replace augmenter by your image_aug_torch.cc in new mxnet source.
But ,there are some error 。
May you help me change the code to adapt to new mxnet,when you have free time .
Thank you very much.

hi, I have check the json file of your model, and I found that there is a slice-axis op,but I can't find the implementation of this operation in mxnet/src/operator

I downloaded dpn-68-5k and dpn-92-5k models, where softmax output number is 4786. But I found that there are 4998 semantic classes in train and valid lists. Could you tell me the relation between label and semantic class:
e.g.
label class
0 n12934174
1 n12789054

I noticed that in model json files, there are not "moving_mean" and "moving_variance" in BatchNorm layers. Can you explain why? Thx.

Dear cypw,
After i download the pre-trained model name dpn68-5k_2017_08_18.tar.gz.I find the result is wrong if i substrated mean rgb, Need substrated mean RGB?

Hi, there,
Do you have a pretrained DPN model based on Keras? I have check these two repos, https://github.com/cypw/DPNs and https://github.com/titu1994/Keras-DualPathNetworks, but cann't find any links to a pretrained model.
Please help.

Hello,
Your implementation uses the Micro-block as: BN-Act-Conv2d.
However, the ResNeXt uses the micro-block structure: Conv2d-BN-Act.
So between the two implementations, the Conv2d is missing at the first block.

Reading your paper, if I understand correctly, the implementation should follow the ResNeXt style
, such as implemented by Titu1994

Can you help to clarify the difference (if any)? Thanks for your help.

Hi,

You write that if one​ have seen mean-max pooling, then let you know :)

Such pooling was proposed in
Generalizing Pooling Functions in Convolutional Neural Networks: Mixed, Gated, and Tree
Chen-Yu Lee, Patrick W. Gallagher, Zhuowen Tu
AISTATS 2016
http://arxiv.org/abs/1509.08985

I have tested it in my benchmark https://github.com/ducha-aiki/caffenet-benchmark/blob/master/Pooling.md

The only difference is that max-average pooling wasn't applied to the last layer.

Hi, yunpeng. I am trying to prepare ImageNet-5k training data by your provided train.lst.
I have prepared the ImageNet-10k, and I found that many images which in your train.lst are not included in ImageNet-10k dataset.
Such as:
IOError: [Errno 2] No such file or directory: '/home/datasets/Dataset/imagenet10k/n02399000/n02399000_5702.JPEG'

Would you mind sharing more informations about the preparation of ImageNet-5k data?

I use the new mxnet/tools/im2rec.py to produce .rec file,
when I run the mxnet in your DPNs, error"Segmentation fault(core dumped)" is showed.

This is my problem:

Can you help me?Thanks@cypw

Hi，
Does DPNs has the trained model which based on Caffe?

Hi, I am trying to finetune DPN-107 on a new dataset. I use latest mxnet and add scale=0.0167 in image iter. However, the training accuracy is very low. The resnext 101 model can reach 80+ while dpn only 40+. I have verified that using latest mxnet and scale=0.0167 can get TOP-1 ~95.0 on imagenet validation. So it's very stange why finetuning DPN on new dataset is not working well. I also tried to fix all layers except the last fc for classification. The performance is also very low. Do you have any comment on how to finetune DPN on new dataset? Thanks.

Hi! Thanks for your impressive work! I'm trying to remake your results. Would you share more information about your hyper parameters, especially the steps for learning rate which I can not find any information about the value in detail.

In scope.py, the shape of input image is [3, 320, 320] during inference step, i want to know how did you do it?
I saw you use ImageRecorderIter to preprocess the input images, does it just resize the image into [3, 320, 320], or some other operations?
Thx for your help!~~