def convolutionFeatureMap(data, bias, w):
featureMap = []
size = len(w)
for index in range(size):
result = w[index].dot(data) + bias[index, :]
featureMap.append(nnet.sigmoid(result))
return featureMap
def convolutionFeatureMap(data, bias, w):
featureMap = []
size = len(w)
for index in range(size):
result = w[index].dot(data) + bias[index,:]
featureMap.append(nnet.sigmoid(result))
return featureMap
def test1_sigmoid(self):
s = nnet.sigmoid(np.array([[0, 1, 2], [2, 4, -1]]))
self.assertEqual(s.shape[0], 2)
self.assertEqual(s.shape[1], 3)
self.assertAlmostEqual(s[0, 0], 0.5)
self.assertAlmostEqual(s[0, 1], 0.7310585786300049)
self.assertAlmostEqual(s[0, 2], 0.8807970779778823)
self.assertAlmostEqual(s[1, 2], 0.2689414213699951)
def test_forward_prop():
np.random.seed(1)
X = np.random.randn(3, 3)
params = init_params([3, 2, 1])
W1, W2 = params['W']
b1, b2 = params['b']
Y_pred, cache = forward_prop(X, params)
A0 = X
Z1 = np.dot(W1, A0) + b1
A1 = sigmoid(Z1)
Z2 = np.dot(W2, A1) + b2
A2 = sigmoid(Z2)
assert (Z1 == cache['Z'][0]).all()
assert (A1 == cache['A'][1]).all()
assert (Z2 == cache['Z'][1]).all()
assert (A2 == Y_pred).all()
def convolutionFeatureMapMulti(data, bias, w):
featNum = len(data) # how many L1 feature maps to be convoluted
size = len(w[0]) # how many L2 features(covolution cores) dose each L1 feature map have
# so variable w is the array that [L1-feat-nums * L2-feat-nums * each-convlution-weight]
featureMap = []
for index in range(size):
for i in range(featNum):
if i == 0:
tmp = w[i][index].dot(data[i])
else:
tmp += w[i][index].dot(data[i])
result = tmp + bias[index, :]
featureMap.append(nnet.sigmoid(result))
return featureMap
def convolutionFeatureMapMulti(data, bias, w):
featNum = len(data) #how many L1 feature maps to be convoluted
size = len(w[0]) #how many L2 features(covolution cores) dose each L1 feature map have
#so variable w is the array that [L1-feat-nums * L2-feat-nums * each-convlution-weight]
featureMap = []
for index in range(size):
for i in range(featNum):
if i == 0:
tmp = w[i][index].dot(data[i])
else:
tmp += w[i][index].dot(data[i])
result = tmp + bias[index,:]
featureMap.append(nnet.sigmoid(result))
return featureMap
#step4 plot the learned feature
fig = plt.figure(2)
for index in range(hiddenSize/10):
weight = W1[index,:]
weight = np.reshape(weight,(28,28))
ax = fig.add_subplot(5,4,1+index)
ax.imshow(weight,mpl.cm.gray)
plt.show()
#step5 extract features from test & training data
#TODO move sigmoid to miscellaneous
labeled_dataset = nnet.normalization(labeled_dataset)
test_data = nnet.normalization(test_data)
train_a1 = nnet.sigmoid(W1.dot(labeled_dataset)+b1)
test_a1 = nnet.sigmoid(W1.dot(test_data)+b1)
#step 6 softmax regression
W = softmax.softmax_regression(hiddenSize,numLabels,lmd,train_a1,labeled_labelset,100)
#step 7 testing
theta = W.reshape((numLabels, hiddenSize))
predict = (theta.dot(test_a1)).argmax(0)
print predict
print test_labels.flatten()
accuracy = (predict == test_labels.flatten())
print 'Accuracy:',accuracy.mean()
print 'done'
def predict(im, params):
mu, log_sigmasq = map(np.squeeze, gmlp(im, params))
return 2*np.pi*sigmoid(mu), 2.*np.tanh(log_sigmasq / 2.)
start = time.time()
#step2 L1 feature learning using sparse autoencoder
#TODO move normalization to miscellaneous
training_data = nnet.normalization(training_data)
#W = nnet.sparseAutoencoder(inputSize,hiddenSizeL1,sparsityParam,lmd,beta,alpha,training_data,iters=500)
W = load_data('weightL1')
W = np.array(W)
W = W.transpose()
#savedata(W,'weightL1')
W11 = np.reshape(W[:hiddenSizeL1*inputSize,], (hiddenSizeL1, inputSize))
b11 = np.reshape(W[2*hiddenSizeL1*inputSize:2*hiddenSizeL1*inputSize+hiddenSizeL1,],(hiddenSizeL1,1))
#step3 L2 feature learning using sparse autoencoder
training_a1 = nnet.sigmoid(W11.dot(training_data)+b11)
#W = nnet.sparseAutoencoder(hiddenSizeL1,hiddenSizeL2,sparsityParam,lmd,beta,0.009,training_a1,iters=500)
W = load_data('weightL2')
W = np.array(W)
W = W.transpose()
#savedata(W,'weightL2')
W21 = np.reshape(W[:hiddenSizeL2*hiddenSizeL1,], (hiddenSizeL2, hiddenSizeL1))
b21 = np.reshape(W[2*hiddenSizeL2*hiddenSizeL1:2*hiddenSizeL2*hiddenSizeL1+hiddenSizeL2,],(hiddenSizeL2,1))
#step4 plot the learned feature
#fig = plt.figure(2)
#for index in range(hiddenSizeL1/10):
# weight = W11[index,:]
# weight = np.reshape(weight,(28,28))
# #print weight.shape
def predict(im, params):
mu, log_sigmasq = map(np.squeeze, gmlp(im, params))
return 2 * np.pi * sigmoid(mu), 2. * np.tanh(log_sigmasq / 2.)
#step2 L1 feature learning using sparse autoencoder
#TODO move normalization to miscellaneous
training_data = nnet.normalization(training_data)
#W = nnet.sparseAutoencoder(inputSize,hiddenSizeL1,sparsityParam,lmd,beta,alpha,training_data,iters=500)
W = load_data('weightL1')
W = np.array(W)
W = W.transpose()
#savedata(W,'weightL1')
W11 = np.reshape(W[:hiddenSizeL1 * inputSize, ], (hiddenSizeL1, inputSize))
b11 = np.reshape(
W[2 * hiddenSizeL1 * inputSize:2 * hiddenSizeL1 * inputSize +
hiddenSizeL1, ], (hiddenSizeL1, 1))
#step3 L2 feature learning using sparse autoencoder
training_a1 = nnet.sigmoid(W11.dot(training_data) + b11)
#W = nnet.sparseAutoencoder(hiddenSizeL1,hiddenSizeL2,sparsityParam,lmd,beta,0.009,training_a1,iters=500)
W = load_data('weightL2')
W = np.array(W)
W = W.transpose()
#savedata(W,'weightL2')
W21 = np.reshape(W[:hiddenSizeL2 * hiddenSizeL1, ],
(hiddenSizeL2, hiddenSizeL1))
b21 = np.reshape(
W[2 * hiddenSizeL2 * hiddenSizeL1:2 * hiddenSizeL2 * hiddenSizeL1 +
hiddenSizeL2, ], (hiddenSizeL2, 1))
#step4 plot the learned feature
#fig = plt.figure(2)
#for index in range(hiddenSizeL1/10):
