ahmedfgad / numpycnn Goto Github PK

Hi Ahmed,

    if len(img.shape) > 2 or len(conv_filter.shape) > 3: # Check if number of image channels matches the filter depth.
        if img.shape[-1] != conv_filter.shape[-1]:
            print("Error: Number of channels in both image and filter must match.")
            sys.exit()

The above snippet in the conv function checks for channel mismatch. But, if len(img.shape) is greater than 2 (eg - (224, 224, 3)), and the conv_filter's shape is (2,3,3), then the condition if img.shape[-1] != conv_filter.shape[-1] is False , even though the number of channels of the image, and the conv_filter, are not the same.
One possible solution is -

    if len(img.shape) != len(conv_filter.shape) - 1: # Check whether number of dimensions is the same
        print("Error: Number of dimensions in conv filter and image do not match.")  
        exit()
    if len(img.shape) > 2 or len(conv_filter.shape) > 3: # Check if number of image channels matches the filter depth.
        if img.shape[-1] != conv_filter.shape[-1]:
            print("Error: Number of channels in both image and filter must match.")
            sys.exit()

The size of input image is (512,512), and the size of conv1_out is (510,510), the size of pool_out should be (255,255). However the size of pool_out is (254,254). So I suggest that the codes in pooling() function should be changed as following:
pool_out = np.zeros((np.uint16((feature_map.shape[0]-size)/stride+1),
np.uint16((feature_map.shape[1]-size)/stride+1),
feature_map.shape[-1]))

The current implementation of the sigmoid() function can get a RuntimeWarning overflow in cnn.py at line 29 numpy.exp():
"return 1.0 / (1 + numpy.exp(-1 * sop))"
To avoid/correct the overflow, use:
"return numpy.exp(-numpy.logaddexp(0, -sop))"

就是在第二层和第三层的卷积核为什么要用随机数赋值，不是很懂这些原理

your code is wonderful!
but I have one question: how to create my own dataset use code. And how to label or training it?
thank you !

When i run this code have a error that is "operands could not be broadcast together with shapes (7,6) (7,7) "

I ran the NumPyCnn's pooling code, and find the result is not same with caffe's pooling layer, when featuremap.shape[0] or featuremap.shape[1] is odd,. And I think this is a bug.

I modify the pooling code ,and it works well in my project. The new code is below.

The current relu function actually does no work for the feature_map.
The code relu_out[r, c, map_num] = numpy.max(feature_map[r, c, map_num], 0) should be written as
relu_out[r, c, map_num] = numpy.max([feature_map[r, c, map_num], 0])
And thanks for the code you shared!

@ahmedfgad

in the pooling()
change the "feature_map.shape[0]-size+1)/stride" to "feature_map.shape[0]-size)/stride+1"

pool_out = numpy.zeros((numpy.uint16((feature_map.shape[0]-size)/stride+1), 
                         numpy.uint16((feature_map.shape[1]-size+1)/stride+1), 
                         feature_map.shape[-1])) 

in the relu()
change the "feature_map.shape[1]-size-1" to "feature_map.shape[1]-size+1" to arrive at the end.

 for r in numpy.arange(0,feature_map.shape[0]-size+1, stride): 

 for c in numpy.arange(0, feature_map.shape[1]-size+1, stride):