Below is my code for the ANN in Python. It doesn't seem to want to train however.
As you can see by the code, the image conversion is a complicated mess.
I put lots of comments in the code.
Any ideas?
The answer is to set up the network using these functions:
ann = cv2.ml.ANN_MLP_create()
ann.setLayerSizes(np.array([num_input_neurons, num_hidden_neurons, num_output_neurons], dtype=np.int64))
ann.setActivationFunction(cv2.ml.ANN_MLP_SIGMOID_SYM)
ann.setTermCriteria((cv2.TERM_CRITERIA_COUNT | cv2.TERM_CRITERIA_EPS, 1, 0.000001 ))
ann.setTrainMethod(cv2.ml.ANN_MLP_BACKPROP, 0.001)
#ann.setBackpropMomentumScale(0.00001)
ann.setBackpropWeightScale(0.00001)
I swear I tried this approach before, but I guess not. Thanks to @berak for the help again.
Asked: 2018-04-27 17:17:10 -0600
Seen: 439 times
Last updated: Apr 28 '18
64x64 is not unreasonable, nor is 512x512. I am going to try to get more input images, but 4 classifications shouldn't be too little... I am thinking of the python opencv XOR ANN, which only has 4 classifications. I just did that the other day: https://github.com/sjhalayka/python_o...
yea, but that will fool you forever. it's not at all a realistic assumption / problem.
I also tried the same code (as much as possible) in C++, and it works:
https://github.com/sjhalayka/opencv_i...