How can I try FCN example in Windows, VS2013 environment

[Kim Han-eol]

Hello, I have started to study caffe recently.

I tested on Windows + Visual Studio 2013 + CUDA 7.5 environment.

Based on studying on MNIST and ImageNet examples, I am able to test my learning model using AlexNet, GoogleNet, LeNet, and so forth.

However, there are problems when I am trying to implement FCN example,

(https://github.com/shelhamer/fcn.berkeleyvision.org)

The problems are as below:

1) I don't know how to make lmdb data set

From my knowledge, I made a list of training data which containing filename of images and label number.

and execute "convert_imageset.exe" to make them into lmdb format with appropriate size.

[train.txt]

f01.jpg 0

f02.jpg 0

g01.jpg 1

g02.jpg 1

...

(comment: there are 7 labels totally)

 

[command]

%CAFFE_BIN%convert_imageset.exe -resize_height=256 -resize_width=256 -shuffle %TRAIN_DATA_ROOT% %TRAIN_DATA_TXT% train_lmdb -gray=true

In this case, I am not sure that I can define lmdb like image classification examples.

I hope that there are appropriate way to define lmdb for FCN example.

2) I don't understand an error message

When I tried to run my training model using caffe.exe

In this experiment, I use a set of files as below:

[Training mode]

https://github.com/shelhamer/fcn.berkeleyvision.org/blob/master/voc-fcn32s/train.prototxt

I changed num_output parameter of two last layers, "score_fr" and "upscore", from 21 to 7.

Number 7 means the number of labels of my model

Moreover, I replaced "data" layer with mine

name: "SkyNet"

layer {

 name: "sky"

 type: "Data"

 top: "data"

 top: "label"

 include {

   phase: TRAIN

 }

 data_param {

   source: "../data/train_lmdb"

   batch_size: 1

   backend: LMDB

 }

}

layer {

 name: "sky"

 type: "Data"

 top: "data"

 top: "label"

 include {

   phase: TEST

 }

 data_param {

   source: "../data/test_lmdb"

   batch_size: 1

   backend: LMDB

 }

}

...

layer {

 name: "score_fr"

 type: "Convolution"

 bottom: "fc7"

 top: "score_fr"

 param {

   lr_mult: 1

   decay_mult: 1

 }

 param {

   lr_mult: 2

   decay_mult: 0

 }

 convolution_param {

   num_output: 7

   pad: 0

   kernel_size: 1

 }

}

layer {

 name: "upscore"

 type: "Deconvolution"

 bottom: "score_fr"

 top: "upscore"

 param {

   lr_mult: 0

 }

 convolution_param {

   num_output: 7

   bias_term: false

   kernel_size: 64

   stride: 32

 }

}

layer {

 name: "score"

 type: "Crop"

 bottom: "upscore"

 bottom: "data"

 top: "score"

 crop_param {

   axis: 2

   offset: 19

 }

}

layer {

 name: "loss"

 type: "SoftmaxWithLoss"

 bottom: "score"

 bottom: "label"

 top: "loss"

 loss_param {

   ignore_label: 255

   normalize: false

 }

}

[Solver]

https://github.com/shelhamer/fcn.berkeleyvision.org/blob/master/voc-fcn32s/solver.prototxt

However, when I run a command %CAFFE_BIN%caffe.exe train -solver=solver.prototxt, 

I faced to an error as below:

I1010 20:26:53.162600 30112 layer_factory.hpp:77] Creating layer upscore

I1010 20:26:53.162600 30112 net.cpp:100] Creating Layer upscore

I1010 20:26:53.163599 30112 net.cpp:444] upscore <- score_fr

I1010 20:26:53.163599 30112 net.cpp:418] upscore -> upscore

I1010 20:26:53.163599 30112 net.cpp:150] Setting up upscore

I1010 20:26:53.163599 30112 net.cpp:157] Top shape: 1 7 64 64 (28672)

I1010 20:26:53.163599 30112 net.cpp:165] Memory required for data: 110689996

I1010 20:26:53.164597 30112 layer_factory.hpp:77] Creating layer score

I1010 20:26:53.164597 30112 net.cpp:100] Creating Layer score

I1010 20:26:53.164597 30112 net.cpp:444] score <- upscore

I1010 20:26:53.164597 30112 net.cpp:444] score <- data_sky_0_split_1

I1010 20:26:53.164597 30112 net.cpp:418] score -> score

I1010 20:26:53.164597 30112 net.cpp:150] Setting up score

I1010 20:26:53.164597 30112 net.cpp:157] Top shape: 1 7 21 21 (3087)

I1010 20:26:53.164597 30112 net.cpp:165] Memory required for data: 110702344

I1010 20:26:53.164597 30112 layer_factory.hpp:77] Creating layer loss

I1010 20:26:53.164597 30112 net.cpp:100] Creating Layer loss

I1010 20:26:53.164597 30112 net.cpp:444] loss <- score

I1010 20:26:53.165596 30112 net.cpp:444] loss <- label

I1010 20:26:53.165596 30112 net.cpp:418] loss -> loss

I1010 20:26:53.165596 30112 layer_factory.hpp:77] Creating layer loss

F1010 20:26:53.165596 30112 softmax_loss_layer.cpp:47] Check failed: outer_num_

* inner_num_ == bottom[1]->count() (441 vs. 1) Number of labels must match numbe

r of predictions; e.g., if softmax axis == 1 and prediction shape is (N, C, H, W

), label count (number of labels) must be N*H*W, with integer values in {0, 1, .

.., C-1}.

*** Check failure stack trace: ***

How can I solve these problems?

Please give me any comments.

Thanks in advance.

-Sky Kim

[zhe...@gmail.com]

Hi Kim:

      Your problems have been solved? I also meet some problems when I try to duplicate this FCN experiment. What my problems are:

              1.python.exe stops work; 

              2.when I run the solve.py on jupyter notebook,it shows the error: caffe_device(int(sys.argv[1])):list index out of range. 

      Those problems have trouble me 1 week..........

      Do U succeed? Can U give me some suggestions about it or show me the tutorial about FCN in caffe on windows. I am a new man......

      Thanks.

                                                     ____zhengvh

                                  

在 2016年10月10日星期一 UTC+8下午7:50:20，Kim Han-eol写道：