def custom_conv2d(bottom_data, weights, bias):
input = bottom_data.copy()
shape = bottom_data.shape
if len(shape) > 3:
input = input.reshape((shape[1], shape[2], shape[3]))
input = input.transpose((1, 2, 0))
#weights = weights.transpose((0, 1, 2, 3)) wrong
#weights = weights.transpose((0, 1, 3, 2)) wrong
weights = weights.transpose((0, 2, 3, 1))
#weights = weights.transpose((0, 2, 1, 3)) wrong
#weights = weights.transpose((0, 3, 1, 2))
#weights = weights.transpose((0, 3, 2, 1))
conv_result = numpycnn.conv(input, weights)
for bias_index in range(len(bias)):
conv_result[:, :,
bias_index] = conv_result[:, :,
bias_index] + bias[bias_index]
return conv_result
data = fd.data
y = fd.label
length = len(data)
# get Synapse
syn0 = open("syn0.pickle", "rb")
syn0 = pickle.load(syn0)
syn1 = open("syn1.pickle", "rb")
syn1 = pickle.load(syn1)
# Testing and Counting truth Rate
epoch = length
benar = 0
runn = 0
for i in range(epoch):
ri = np.random.randint(length)
# singleData = np.reshape(data[ri], (-1, 28)) # reshape into 28 x 28
singleData = data[ri]
l1_feature_map = numpycnn.conv(singleData, filters.filter)
# ReLu layer
l1_feature_map_relu = numpycnn.relu(l1_feature_map)
# Pooling Layer
l1_feature_map_relu_pool = numpycnn.pooling(l1_feature_map_relu, 2, 2)
# Forward Propagation
final = np.array([l1_feature_map_relu_pool.ravel()])
print(testing(final.ravel(), y[ri]))
# benar += testing(final.ravel(),y[ri])
# runn += 1
# print('process : ',round(i/epoch*100,2),'% truth rate : ', round(benar / runn * 100,2), '%')
# Converting the image into gray.
img = skimage.color.rgb2gray(img)
# First conv layer
#l1_filter = numpy.random.rand(2,7,7)*20 # Preparing the filters randomly.
l1_filter = numpy.zeros((2,3,3))
l1_filter[0, :, :] = numpy.array([[[-1, 0, 1],
[-1, 0, 1],
[-1, 0, 1]]])
l1_filter[1, :, :] = numpy.array([[[1, 1, 1],
[0, 0, 0],
[-1, -1, -1]]])
print("\n**Working with conv layer 1**")
l1_feature_map = numpycnn.conv(img, l1_filter)
print("\n**ReLU**")
l1_feature_map_relu = numpycnn.relu(l1_feature_map)
print("\n**Pooling**")
l1_feature_map_relu_pool = numpycnn.pooling(l1_feature_map_relu, 2, 2)
print("**End of conv layer 1**\n")
# Second conv layer
l2_filter = numpy.random.rand(3, 5, 5, l1_feature_map_relu_pool.shape[-1])
print("\n**Working with conv layer 2**")
l2_feature_map = numpycnn.conv(l1_feature_map_relu_pool, l2_filter)
print("\n**ReLU**")
l2_feature_map_relu = numpycnn.relu(l2_feature_map)
print("\n**Pooling**")
l2_feature_map_relu_pool = numpycnn.pooling(l2_feature_map_relu, 2, 2)
print("**End of conv layer 2**\n")
np.random.seed(1)
# synapse
syn0 = 2 * np.random.random((4608, 100)) - 1
syn1 = 2 * np.random.random((100, 5)) - 1
length = len(data)
print(length)
epoch = 1 * length
for j in range(1):
# print(j)
ri = np.random.randint(length)
# convoluting layer
# singleData = np.reshape(data[ri], (-1, 28)) # reshape into 28 x 28 pnly for MNIST
singleData = data[ri] / 255
print(np.amax(singleData), np.amin(singleData))
l1_feature_map = numpycnn.conv(singleData, filters.filter1)
l1_feature_map_relu = numpycnn.relu(l1_feature_map)
l1_feature_map_relu_pool = numpycnn.pooling(l1_feature_map_relu, 2, 2)
print(l1_feature_map_relu_pool.shape, np.amax(l1_feature_map_relu_pool),
np.amin(l1_feature_map_relu_pool))
feature_input = []
for in2, conv2 in enumerate(l1_feature_map_relu_pool.T):
l2_feature_map = numpycnn.conv(conv2, filters.filter2)
l2_feature_map_relu = numpycnn.relu(l2_feature_map)
l2_feature_map_relu_pool = numpycnn.pooling(l2_feature_map_relu, 2, 2)
cv2.imwrite('conv1_' + str(in2) + '.jpg', conv2 * 255)
print(l2_feature_map_relu_pool.shape,
np.amax(l2_feature_map_relu_pool),
np.amin(l2_feature_map_relu_pool))
def custom_pool(bottom_data, size, stride):
#return numpycnn.pooling(bottom_data, size, stride)
return numpycnn.pooling_new(bottom_data, size, stride)