Error:operands could not be broadcast together with shapes (7,6) (7,7) about numpycnn

[lomenie]

I ve got the same....

[lomenie]

Thanks but still an issue
Working with conv layer 1*
('Filter ', 1)
Traceback (most recent call last):
File "example.py", line 44, in
l1_feature_map = numpycnn.conv(img, l1_filter)
File "/Users/lomenie/Documents/GitHub/NumPyCNN/numpycnn.py", line 78, in conv
conv_map = conv_(img, curr_filter)
File "/Users/lomenie/Documents/GitHub/NumPyCNN/numpycnn.py", line 38, in conv_
curr_result = curr_region * conv_filter
ValueError: operands could not be broadcast together with shapes (2,2) (3,3)

[lomenie]

no answer from the developer ?

[ahmedfgad]

Thanks for your note.
The shared code is running fine with me and some other people. This is why I got confused of it.
But after investigating the code well, I figured out the source the error and I am planning to work on the code and solve it in the next few days.
Thanks again.

[ahmedfgad]

The issue is fixed. The issue was due to incompatibility of shape between the filter and the image region being worked on. It is solved based on such lines:

curr_region = img[r-numpy.uint16(numpy.floor(filter_size/2)):r+numpy.uint16(numpy.ceil(filter_size/2)), 
                              c-numpy.uint16(numpy.floor(filter_size/2)):c+numpy.uint16(numpy.ceil(filter_size/2))]

[lomenie]

Thanks but it means that some version of python or numpy are not sensitive to the type issue right? Nicolas Lomenie

…

Le 24 mai 2018 à 15:28, Ahmed Gad ***@***.***> a écrit : The issue is fixed. The issue was due to incompatibility of shape between the filter and the image region being worked on. It is solved based on such line: curr_region = img[r-numpy.uint16(numpy.floor(filter_size/2)):r+numpy.uint16(numpy.ceil(filter_size/2)), c-numpy.uint16(numpy.floor(filter_size/2)):c+numpy.uint16(numpy.ceil(filter_size/2))] — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[ahmedfgad]

Yes this is the reason.
The code is running in my PC without error. The same code gives an error on another without changing any part.
At all, the version of Python and NumPy I used are as follows:
The project is tested using Python 3.5.2 installed inside Anaconda 4.2.0 (64-bit)
NumPy version used is 1.14.0

[lomenie]

With the first code here is the pb (so I do not think it is a uint type issue)
Working with conv layer 1
('Filter ', 1)
Filter Size
3
('Filter ', 2)
Filter Size
3

ReLU

Pooling
End of conv layer 1

Working with conv layer 2
('Filter ', 1)
Filter Size
5
Filter Size
5
('Filter ', 2)
Filter Size
5
Filter Size
5
('Filter ', 3)
Filter Size
5
Filter Size
5

ReLU

Pooling
End of conv layer 2

Working with conv layer 3
('Filter ', 1)
Filter Size
7
Traceback (most recent call last):
File "NumPyCNN.py", line 129, in
l3_feature_map = conv(l2_feature_map_relu_pool, l3_filter)
File "NumPyCNN.py", line 58, in conv
conv_map = conv_(img[:, :, 0], curr_filter[:, :, 0]) # Array holding the sum of all feature maps.
File "NumPyCNN.py", line 23, in conv_
curr_result = curr_region * conv_filter
ValueError: operands could not be broadcast together with shapes (7,6) (7,7)

[ahmedfgad]

I am very thankful for your notes.
You are right. The bug still exist.
I figured a fault that may causing such error after running such code on another device with another version of NumPy.
I figured out that the numpy.ceil was not working correctly. When I run this line,

numpy.uint16(numpy.ceil(3/2))

The expected output is 2 but it returns 1. Due to dividing two integers (3 and 2), the result will also be an integer (1). The ceil of that integer is 1.
Based on that, the region size is calculated based on the following code will be always smaller than the filter size:

            curr_region = img[r-numpy.uint16(numpy.floor(filter_size/2)):r+numpy.uint16(numpy.ceil(filter_size/2)), 
                              c-numpy.uint16(numpy.floor(filter_size/2)):c+numpy.uint16(numpy.ceil(filter_size/2))]

This bug is fixed by making one of the two numbers float. For example, numpy.ceil(filter_size/2.0). The previous code will be as follows:

curr_region = img[r-numpy.uint16(numpy.floor(filter_size/2.0)):r+numpy.uint16(numpy.ceil(filter_size/2.0)), c-numpy.uint16(numpy.floor(filter_size/2.0)):c+numpy.uint16(numpy.ceil(filter_size/2.0))]

[ahmedfgad]

The issue is closed for being solved.