houqb / dss Goto Github PK

I am using caffe in your github(https://github.com/Andrew-Qibin/caffe_dss). I followed the instructions in the DSS readme. When I executed the net = caffe.Net('deploy.prototxt', 'dss_model_released.caffemodel', caffe.TEST) in DSS-tutorial.ipynb, it reported the following error.

I0808 20:35:52.644126 30 net.cpp:255] Network initialization done.
F0808 20:35:52.754325 30 upgrade_proto.cpp:95] Check failed: ReadProtoFromBinaryFile(param_file, param) Failed to parse NetParameter file: dss_model_released.caffemodel
*** Check failure stack trace: ***
Aborted (core dumped)

It seems that the caffemodel was not loaded successfully.
Can anyone help me solve this problem?
Thanks.

Can the ground truth have pixel values other than 0 and 1 (after normalization)?
After loading the label images using your data layer, I was checking the unique pixel values. For any image the values are always 0 and 1 (normalize=true) even though the label image has values other than 0 and 255.

Does your data layer support only binary label images?

Hi, Andrew,

Thank you very much to provide the code to the paper DSS.

I followed the instructions to download the code and the pretrained model. Then I used the provided code DSS-tutorial.ipynb to run test on ECSSD dataset. The PR curve I got is even worse than DCL.

Any idea about what happened? Thank you in advance.

Hi! Could you provide the results without CRF?

When i train the model, i can't solve the problem " Unknown layer type: ImageLabelmapData ". If you know a method to solve the problem, please to help me. Thank you very much!!!

I want test a picture by your pretrained model,how should I use your code?Where is the entrance of the test picture?

Can you please upload the results of the paper in google drive link. I am unable to access Baidu drive.

layer { name: "concat-dsn3" bottom: "conv3-dsn3" bottom: "upscore-dsn6-3" bottom: "upscore-dsn5-3" top: "concat-dsn3" type: "Concat" concat_param { concat_dim: 1} }

@Andrew-Qibin Hi, I'm re-implement this network in keras, and have read your paper before.
I remember the concate in your paper present in a dense way. The conct-dsn3 should be include a bottom like 'upscore-dsn4-3', but there only include upscore-dsn5-3 and upscore-dsn6-3.

the layer 'concat_dsn3' included 'upscore-dsn4-3' or did not ?

I am getting this error at the line net.forward() during testing:
F0304 14:26:21.611594 4889 syncedmem.cpp:71] Check failed: error == cudaSuccess (2 vs. 0) out of memory
*** Check failure stack trace: ***

Process finished with exit code 134 (interrupted by signal 6: SIGABRT)

1. How to solve this.
2. I understand the default batch size of testing phase is 1. Is my understanding right?

I tried to reproduce the result,but unfortunately I got lower grades with training on the MSRA10K dataset.so i want to train on the the msra-b dataset as you do,but i only get the groundtruth of msra-b,could you provide me the msra-b dataset? thank you very much!

Hello, when I am using your code in my computer, I found this problem.
When I configured the caffe and run python run_saliency.py in the main directory, this error occured

I1018 16:44:03.186270 3995 solver.cpp:86] Creating training net from train_net file: train_val.prototxt
[libprotobuf ERROR google/protobuf/text_format.cc:245] Error parsing text-format caffe.NetParameter: 20:14: Message type "caffe.ImageDataParameter" has no field named "normalize".
F1018 16:44:03.186388 3995 upgrade_proto.cpp:934] Check failed: ReadProtoFromTextFile(param_file, param) Failed to parse NetParameter file: train_val.prototxt.

I checked the file train_val.prototxt, it had the param 'normalize' like this:
layer {
name: "data"
type: "ImageLabelmapData"
top: "data"
top: "label"
include { phase: TRAIN }
transform_param {
mirror: true
mean_value: 104.00699
mean_value: 116.66877
mean_value: 122.67892
}
image_data_param {
root_folder: "/opt/dataset/saliency/msra_b/"
source: "../../data/msra_b/train.lst"
batch_size: 1
shuffle: true
normalize: true
}
}

Any idea?

I am using caffe 1.0.0. I followed the instructions in the DSS readme, but when I try to execute the script run_saliency.py, I am getting the following error:

[libprotobuf ERROR google/protobuf/text_format.cc:245] Error parsing text-format caffe.NetParameter: 20:14: Message type "caffe.ImageDataParameter" has no field named "normalize".
F0724 17:51:04.304796 10783 upgrade_proto.cpp:88] Check failed: ReadProtoFromTextFile(param_file, param) Failed to parse NetParameter file: train_val.prototxt
*** Check failure stack trace: ***
Aborted (core dumped)

The error occurs because of line 20 of train_val.prototxt which has normalize: true.

Can you please help me in understanding why this happens? Thanks in advance.

F0919 09:43:45.324304 18941 layer_factory.hpp:81] Check failed: registry.count(type) == 1 (0 vs. 1) Unknown layer type: ImageLabelmapData
How to fix this problem? hope someone can help me, thanks~

I run the code in CPU_ONLY mode. It seems that my computer doesn't have enough memory (4G), How to reduce the use of memory?

I download your result data, code and model in order to reproduce your result, and I really done
But I use the ECSSD groudtruth in your result data could not reproduce your result, so I change to the original ECSSD dataset's groudtruth and I reproduced.
There are some differences between your gt and original gt dataset, such as data type and files' storage space.
So can you check the gt and give the right one ?
I found difference in SOD, PASCAL-S and ECSSD, thank you!

When I try to run cell #3 of the DSS tutorial notebook, I get the following error:

I0529 12:11:43.640518 15105 layer_factory.hpp:77] Creating layer crop
I0529 12:11:43.640524 15105 net.cpp:91] Creating Layer crop
I0529 12:11:43.640528 15105 net.cpp:425] crop <- score-dsn6-up
I0529 12:11:43.640532 15105 net.cpp:425] crop <- data_input_0_split_1
I0529 12:11:43.640537 15105 net.cpp:399] crop -> upscore-dsn6
F0529 12:11:43.640548 15105 crop_layer.cpp:68] Check failed: bottom[0]->shape(i) - crop_offset >= bottom[1]->shape(i) (1 vs. 3) invalid crop parameters in dimension: 1
*** Check failure stack trace: ***

I'm using caffe release 1.0.
How can I get the code to run?

Thank you very much in advance.

I use my own caffe and find the loss doesn't decrease a lot. After a period of training, nan happens. The data layer is same as predict process provided by you while label lnput is normalized to 0-1. I find that there is a Crop layer, which reshape output salient map to the shape of 'data'. However, salient map has 1 channel while data has 3 channels, why we need to make the shapes same. Is there any problem with it? Also, when I use shape 500,500 as width and height of input, nan happens very quickly while I use shape 512,512. the loss doesn't decrease a lot. It quite strange.

You had created an amazing and large model based on pre-trained VGG model, but you set the same learning rate in the pre-trained models and almost all the following layers in your training. However, the scale of saliency datasets are much smaller than ImageNet dataset, and the results may suffered from overfitting problem. So, how did you solve this problem in you training ?

Does DSS support window&caffe? And can I know the requirements of cudnn version?

I0321 17:23:04.430434 8648 layer_factory.hpp:77] Creating layer loss-dsn6
I0321 17:23:04.446079 8648 net.cpp:100] Creating Layer loss-dsn6
I0321 17:23:04.446079 8648 net.cpp:444] loss-dsn6 <- upscore-dsn6_crop_0_split_0
I0321 17:23:04.446079 8648 net.cpp:444] loss-dsn6 <- label_data_1_split_0
I0321 17:23:04.446079 8648 net.cpp:418] loss-dsn6 -> loss-dsn6
F0321 17:23:04.463042 8648 sigmoid_cross_entropy_loss_layer.cpp:23] Check failed: bottom[0]->count() == bottom[1]->count() (104000 vs. 1) SIGMOID_CROSS_ENTROPY_LOSS layer inputs must have the same count.
*** Check failure stack trace: ***

How can I get the code to run?

Thank you very much in advance.

I had read many CVPR2017 paper related to saliency detection, your model and NLDF(Non-Local Deep Features for Salient Object Detection) all say the processing time == 0.08s, but the model and DL framwork(TF vs Caffe) are different, which should cause a bit different. This makes me confused.
Can you upload an simple ipynb file to explain how the 0.08s to be computed ? or just tell me your method if it's easy to implement.
Thank you very much.

Hi, can u provide google drive link for ur results (saliency map). Unable to access baidu.

I have a question, how much memory do you need to run the code? I try to apply the short connection to other network models, but there are always out of memory errors，I was looking forward to receiving your reply.

could you release the tensorflow code with training part ? Thanks!