CODEWITHSUNDEEP
caffeconv-neural-networkmatcaffe
Alex_AMC
Alex_AMC

Reputation: 27

Error running a forward pass with matcaffe

I am doing transfer learning with a CNN. I want to train the network with my data but I got this error when doing a forward pass: 

Error using CHECK (line 4)
input data cell length must match input blob number

Error in caffe.Net/forward (line 92)
  CHECK(length(input_data) == length(self.inputs), ...

Error in main (line 79)
results= Unet.forward({data});

To take it slow and solve errors step by step I have only a data layer in my network for now. This is my train.prototxt file : 

name: 'my_phseg_v5-train'

force_backward: true

layer {top: 'image' top:'anno'   name: 'loaddata'    type: 'HDF5Data'   hdf5_data_param { source: '/home/alexandra/Documents/my-u-net/my_data.txt' batch_size: 1} include: { phase: TRAIN }}

In matlab: 

model = '/home/alexandra/Documents/my-u-net/my_phseg_v5-train.prototxt';
weights = '/home/alexandra/Documents/my-u-net/my_phseg_v5.caffemodel';

%defining the net: 
Unet = caffe.Net(model, weights, 'train'); % create net and load weights

results= Unet.forward({'image'});

I don't really understand what I have to put as an argument in forward( argument ). Could someone help me on that point ?

I have also noticed that in my Unet the dimension of input cell was 0x1 ... I guess that also a reason why it is not working.

Does someone have an idea on how to solve this problem ?

Upvotes: 0

Views: 559

Answers (1)

Alex_AMC
Alex_AMC

Reputation: 27

I found a solution to my problem:

for the input cell that had a dimension of 0x1:

I used the deploy.prototxt file instead of the train.prototxt file that I was using at the beginning. In this file the dimension of the input is defined.

I used this as an argument of the function forward : 

output = Unet.blobs('image').get_data();
results= Unet.forward({output});

It is the data (in my case the images) itself that have to be put as the input.

Upvotes: 0

Related Questions