rcnn_pool5_to_fcX problem

[empty16]

Hi everyone,
I installed RCNN and trained R-CNN models on PASCAL VOC 07 datasets, when I use the command test_results = rcnn_exp_train_and_test() to get a trained model (I strictly followed every step in rcnn at github), unfortunately I cannot get the right answer, the error messages are as follows:

Error using  * 
Inputs must be 2-D, or at least one input must be scalar.
To compute elementwise TIMES, use TIMES (.*) instead.

Error in rcnn_pool5_to_fcX (line 21)
      feat = max(0, bsxfun(@plus, feat*rcnn_model.cnn.layers(i).weights{1}, ...

Error in rcnn_feature_stats (line 39)
    X = rcnn_pool5_to_fcX(X, layer, rcnn_model);

Error in rcnn_train (line 80)
opts.feat_norm_mean = rcnn_feature_stats(imdb, opts.layer, rcnn_model);

Error in rcnn_exp_train_and_test (line 19)
[rcnn_model, rcnn_k_fold_model] = ... 

Where did I go wrong? I hope someone can give a explanation of the fuction rcnn_pool5_to_fcX or help me figure out this problem, so I can get this code into my own research.
Thank you!

[empty16]

I have got the solution. In Matlab R2014a, I use the squeeze and reshape function in the rcnn_pool5_to_fcX file, and finally I got an output as described in the paper.

[mahatmawoo]

Hi @empty16 ,

I'm also facing a similar issue on Matlab R2014b, just like you did.

The output :

Error using *
Inner matrix dimensions must agree.

Error in rcnn_pool5_to_fcX (line 30)
feat = max(0, bsxfun(@plus, feat*rcnn_model.cnn.layers(i).weights{1}, ...

Error in rcnn_feature_stats (line 42)
X = rcnn_pool5_to_fcX(X, layer, rcnn_model);

Error in rcnn_train_v2 (line 78)
opts.feat_norm_mean = rcnn_feature_stats(imdb, opts.layer, rcnn_model);

Error in rcnn_exp_train_and_test_v2 (line 19)
[rcnn_model, rcnn_k_fold_model] = ...

My interpretation of the problem:
After checking the size of feat = { 256 x 36 } and rcnn_model.cnn.layers(i).weights{1} ={ 9216 x 4096}, I conclude that 9216 must be reshaped into 36 for the matrix multiplication.

Proposed solution:
I'm thinking of applying the following reshaping in rcnn_pool5_to_fcX :

test_var=reshape(rcnn_model.cnn.layers(i).weights{1}, 36,[]);
feat = max(0, bsxfun(@plus, feat*test_var, ...
rcnn_model.cnn.layers(i).weights{2}'));

Question:
Is this where and how you performed the squeeze and reshape function?

Thank you in advance!

[mahatmawoo]

After applying

test_var=reshape(rcnn_model.cnn.layers(i).weights{1}, 36,[]);
feat = max(0, bsxfun(@plus, feat*test_var, ...
rcnn_model.cnn.layers(i).weights{2}'));

I got the following output:

Error using bsxfun
Non-singleton dimensions of the two input arrays must match each other.

Error in rcnn_pool5_to_fcX (line 31)
feat = max(0, bsxfun(@plus, feat*test_var, ...

Error in rcnn_feature_stats (line 42)
X = rcnn_pool5_to_fcX(X, layer, rcnn_model);

Error in rcnn_train_v2 (line 78)
opts.feat_norm_mean = rcnn_feature_stats(imdb, opts.layer, rcnn_model);

Error in rcnn_exp_train_and_test_v2 (line 19)
[rcnn_model, rcnn_k_fold_model] = ...

My interpretation of the problem:
The size of the test_var {36 X 1048576} is different from size of rcnn_model.cnn.layers(i).weights{2} { 4096 x 1}

Proposed solution:
Applied squeeze to test_var.

test_var=squeeze(test_var);

And the output gave me the same error:

Error using bsxfun
Non-singleton dimensions of the two input arrays must match each other.

Error in rcnn_pool5_to_fcX (line 33)
feat = max(0, bsxfun(@plus, feat*test_var, ...

My interpretation of the problem:
The size of the test_var before squeeze {36 X 1048576} and after squeeze {36 X 1048576} remains the same.

I'm not sure how to proceed.
Please feel free to share how you solve your issue. Thank you!

[empty16]

Hi @mahatmawoo ，sorry，I got this message today.
Here is the solution in my rcnn_pool5_to_fcX.m file
feat = max(0, bsxfun(@plus, feat*squeeze(rcnn_model.cnn.layers(i).weights{1}), ... reshape(rcnn_model.cnn.layers(i).weights{2},... size(rcnn_model.cnn.layers(i).weights{2},3),... size(rcnn_model.cnn.layers(i).weights{2},4))));
because the dims of feat and weight are different ,so I changed the weights dim.
Hope this can help you.

[mahatmawoo]

Hi @empty16 , thanks again for your reply.
I'll have a look at your solution soon.
It may help solve my issues.