machine-learning - FCN :Check failed: outer_num_ * inner_num_ == bottom[1]->count()

Question

我设计了一个与FCN相同的网络。输入数据是1*224*224，输入标签是1*224*224。但是我遇到错误:

 F0502 07:57:30.032742 18127 softmax_loss_layer.cpp:47] Check failed: outer_num_ * inner_num_ == bottom[1]->count() (50176 vs. 1) Number of labels must match number 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}.

这是输入结构:

layer {
  name: "data"
  type: "ImageData"
  top: "data"
  top: "label"
  image_data_param {
    ource: "/home/zhaimo/fcn-master/mo/train.txt"   
    batch_size: 1               
    shuffle: true
 }
} 

softmax 层:

layer {
name: "loss"
type: "SoftmaxWithLoss"
bottom: "upscore1"
bottom: "label"
top: "loss"
loss_param {
  ignore_label: 255
  normalize: false
  }
}

train.txt 文件:

/home/zhaimo/fcn-master/data/vessel/train/original/01.png /home/zhaimo/SegNet/data/vessel/train/label/01.png
/home/zhaimo/fcn-master/data/vessel/train/original/02.png /home/zhaimo/SegNet/data/vessel/train/label/02.png
/home/zhaimo/fcn-master/data/vessel/train/original/03.png /home/zhaimo/SegNet/data/vessel/train/label/03.png
/home/zhaimo/fcn-master/data/vessel/train/original/04.png /home/zhaimo/SegNet/data/vessel/train/label/04.png

第一个文件名是输入数据，第二个文件名是其标签。

=============================更新=================== ===================

我尝试使用两个 ImageData 层作为输入:

layer {
name: "data"
  type: "ImageData"
  top: "data"
  image_data_param {
    source: "/home/zhaimo/fcn-master/mo/train_o.txt"    
    batch_size: 1               
    shuffle: false
  }
} 

layer {
  name: "label"
  type: "ImageData"
  top: "label"
  image_data_param {
    source: "/home/zhaimo/fcn-master/mo/train_l.txt"    
    batch_size: 1               
    shuffle: false
  }
} 

但是遇到另一个错误:

I0502 08:34:46.429774 19100 layer_factory.hpp:77] Creating layer data
I0502 08:34:46.429808 19100 net.cpp:100] Creating Layer data
I0502 08:34:46.429816 19100 net.cpp:408] data -> data
F0502 08:34:46.429834 19100 layer.hpp:389] Check failed: ExactNumTopBlobs() == top.size() (2 vs. 1) ImageData Layer produces 2 top blob(s) as output.
*** Check failure stack trace: ***
Aborted (core dumped)

train_o.txt:

/home/zhaimo/fcn-master/data/vessel/train/original/01.png
/home/zhaimo/fcn-master/data/vessel/train/original/02.png
/home/zhaimo/fcn-master/data/vessel/train/original/03.png
/home/zhaimo/fcn-master/data/vessel/train/original/04.png
/home/zhaimo/fcn-master/data/vessel/train/original/05.png

train_l.txt:

/home/zhaimo/SegNet/data/vessel/train/label/01.png
/home/zhaimo/SegNet/data/vessel/train/label/02.png
/home/zhaimo/SegNet/data/vessel/train/label/03.png
/home/zhaimo/SegNet/data/vessel/train/label/04.png
/home/zhaimo/SegNet/data/vessel/train/label/05.png

==================================更新2================ ===================如果我使用两个ImageData层，如何修改deploy.prototxt？这是我编写的文件:

layer {
  name: "data"
  type: "ImageData"
  top: "data"
  top: "tmp0"
  input_param { shape: { dim: 1 dim: 1 dim: 224 dim: 224 } }
} 

和forward.py文件:

import numpy as np
from PIL import Image
caffe_root = '/home/zhaimo/' 
import sys
sys.path.insert(0, caffe_root + 'caffe-master/python')

import caffe

# load image, switch to BGR, subtract mean, and make dims C x H x W for Caffe
im = Image.open('/home/zhaimo/fcn-master/data/vessel/test/13.png')
in_ = np.array(im, dtype=np.float32)
#in_ = in_[:,:,::-1]
#in_ -= np.array((104.00698793,116.66876762,122.67891434))
#in_ = in_.transpose((2,0,1))

# load net
net = caffe.Net('/home/zhaimo/fcn-master/mo/deploy.prototxt', '/home/zhaimo/fcn-master/mo/snapshot/train/_iter_200000.caffemodel', caffe.TEST)
# shape for input (data blob is N x C x H x W), set data
net.blobs['data'].reshape(1, *in_.shape)
net.blobs['data'].data[...] = in_
# run net and take argmax for prediction
net.forward()
out = net.blobs['score'].data[0].argmax(axis=0)

plt.axis('off')
plt.savefig('/home/zhaimo/fcn-master/mo/result/13.png')   

但是我遇到了错误:

F0504 08:16:46.423981  3383 layer.hpp:389] Check failed: ExactNumTopBlobs() == top.size() (2 vs. 1) ImageData Layer produces 2 top blob(s) as output.

请问如何修改forward.py文件？

Answer 1

您的问题出在数据顶部 blob 编号上。对于两个 imagedata 层，请使用以下命令:

layer {
  name: "data"
  type: "ImageData"
  top: "data"
  top: "tmp"
  image_data_param {
    source: "/home/zhaimo/fcn-master/mo/train_o.txt"    
    batch_size: 1               
    shuffle: false
  }
} 

layer {
  name: "label"
  type: "ImageData"
  top: "label"
  top: "tmp1"
  image_data_param {
    // you probably also need
    //is_color: false
    source: "/home/zhaimo/fcn-master/mo/train_l.txt"    
    batch_size: 1               
    shuffle: false
  }
} 

在文本文件中，只需将所有标签设置为 0。您不会使用 tmp/tmp1，所以没关系。