def train_softmax(pooled_train_features, train_num_images, train_labels):
# train softmax regression
softmax_lambda = 1e-4
num_classes = 4
softmax_x = np.transpose(pooled_train_features, (0, 2, 3, 1))
softmax_x = softmax_x.reshape((np.size(pooled_train_features)/train_num_images, train_num_images))
softmax_y = train_labels.reshape(np.size(train_labels)) - 1
init_theta = smrt.initialize_theta(softmax_x.shape[0], num_classes)
theta = smrt.train(init_theta, softmax_x, softmax_y, num_classes, softmax_lambda)
mat = smrt.vec2mat(theta, num_classes)
return mat
def pre_train(train_data, train_labels):
global input_size, num_labels, hidden_size_l1, hidden_size_l2, ro, lbda, beta
levels_l1=[input_size, hidden_size_l1, input_size]
print 'Step 1:', time.time()
# stack 1
init_theta = initialization(levels_l1)
try:
theta = np.genfromtxt("theta_l1.txt")
print theta.shape
except:
theta, cost, info = fmin_l_bfgs_b(J, init_theta, args=(train_data, train_data, levels_l1, lbda, ro, beta), approx_grad=False, maxiter=400)
np.savetxt("theta_l1.txt", theta)
print "cost:", cost
print "info:", info
WB_l1 = vec2mat(theta, levels_l1)
sio.savemat('W_l1', {'W_l1':WB_l1[0][0]})
print 'Step 2:', time.time()
# stack 2
levels_l2 = [hidden_size_l1, hidden_size_l2, hidden_size_l1]
train_l1_a2 = feed_forward(WB_l1[0], train_data, hidden_size_l1)
init_theta = initialization(levels_l2)
try:
theta = np.genfromtxt("theta_l2.txt")
print theta.shape
except:
theta, cost, info = fmin_l_bfgs_b(J, init_theta, args=(train_l1_a2, train_l1_a2, levels_l2, lbda, ro, beta), approx_grad=False, maxiter=400)
np.savetxt("theta_l2.txt", theta)
print "cost:", cost
print "info:", info
WB_l2 = vec2mat(theta, levels_l2)
sio.savemat('W_l2', {'W_l2':WB_l2[0][0]})
print 'Step 3:', time.time()
try:
sr_theta = np.genfromtxt("sr_theta.txt")
print sr_theta.shape
except:
train_l2_a2 = feed_forward(WB_l2[0], train_l1_a2, hidden_size_l2)
sr_init_theta = sr_train.initialize_theta(hidden_size_l2, num_labels)
sr_theta = sr_train.train(sr_init_theta, train_l2_a2, train_labels, num_labels)
np.savetxt("sr_theta.txt", sr_theta)
sr_mat = sr_vec2mat(sr_theta, num_labels)
return WB_l1, WB_l2, sr_mat
def main():
global input_size, num_labels, hidden_size, ro, lbda, beta
levels = [input_size, hidden_size, input_size]
mnist_data = loader.load_train_imgs()
mnist_labels = loader.load_train_labels()
labeled_set = np.where(mnist_labels < num_labels)[0]
unlabeled_set = np.where(mnist_labels > num_labels-1)[0]
train_num = np.round(len(labeled_set)/2)
train_data = mnist_data[:, labeled_set[:train_num]]
train_labels = mnist_labels[labeled_set[:train_num]]
test_data = mnist_data[:, labeled_set[train_num:]]
test_labels = mnist_labels[labeled_set[train_num:]]
unlabeled_data = mnist_data[:, unlabeled_set]
print "train data:", train_data.shape[1]
print "test data:", test_data.shape[1]
print "unlabeled data:", unlabeled_data.shape[1]
init_theta = initialization(levels)
theta, cost, info = fmin_l_bfgs_b(J, init_theta, args=(unlabeled_data, unlabeled_data, levels, lbda, ro, beta), approx_grad=False, maxiter=400)
print "cost:", cost
print "info:", info
WB = vec2mat(theta, levels)
sio.savemat('W', {'W': WB[0][0]})
train_a2 = feed_forward(WB[0], train_data, hidden_size)
test_a2 = feed_forward(WB[0], test_data, hidden_size)
sr_init_theta = sr_train.initialize_theta(hidden_size, num_labels)
sr_theta = sr_train.train(sr_init_theta, train_a2, train_labels, num_labels)
sr_mat = sr_vec2mat(sr_theta, num_labels)
pY = sr_predict(sr_mat, test_a2)
print test_labels[:20]
print pY[:20]
miss = 0.
for i, l in enumerate(test_labels):
if l != pY[i]:
miss += 1.
print miss/len(test_labels)