def classification(neuro_obj = None, epoch = 100000, num_class = 10):
print 'initialize Neural Network.'
if neuro_obj: nn_obj = neuro_obj
else : nn_obj = mln.Mln().make_neuralnet([28*28, 1000, num_class], ['sigmoid', 'softmax'], 0.01) # mnist classification
# else : nn_obj = mln.Mln().make_neuralnet([28*28, 1000, num_class], ['sigmoid', 'softmax'], 0.15) # mnist classification
print "dump obj..."
dp.obj_dump(nn_obj, './default-classification.pkl')
print 'read training data and label.'
training_data = dp.obj_load_gzip('../../mnist/mnist-training_all.pkl.gz')
print 'data size : ', len(training_data)
print 'label size : ', len(training_data)
data_num = len(training_data)
print '--start--'
print '@@ Learn Character Recognition @@'
for j in range(0, epoch):
prg.show_progressxxx(j+1, epoch)
i = np.random.randint(data_num)
nn_obj.learn(training_data[i][0], training_data[i][1])
prg.end_progress()
print "dump obj..."
dp.obj_dump(nn_obj, './learn-classification.pkl')
return nn_obj
def classification(neuro_obj=None, epoch=100000, minibatch_size=1):
num_class = 10
print "initialize Neural Network."
if neuro_obj:
nn_obj = neuro_obj
else:
nn_obj = mln.Mln().make_neuralnet(
[28 * 28, 1000, num_class], ["sigmoid", "softmax"], 0.15
) # mnist classification
# else : nn_obj = mln.Mln().make_neuralnet([28*28, 1000, num_class], ['sigmoid', 'softmax'], 0.01) # mnist classification
# use weight decay.
# nn_obj.use_weight_decay(0.01) # unlearnable
# nn_obj.use_weight_decay(0.001) # learnable
nn_obj.use_weight_decay(0.0001) # learnable
# nn_obj.unuse_weight_decay()
# use momentum.
# nn_obj.use_momentum(0.1)
nn_obj.use_momentum(0.5)
# nn_obj.use_momentum(0.9)
# nn_obj.unuse_momentum()
print "dump obj..."
dp.obj_dump(nn_obj, "./default-classification-batch.pkl")
print "read training data and label."
training_data = dp.obj_load_gzip("../../mnist/mnist-training_all.pkl.gz")
print "data size : ", len(training_data)
print "label size : ", len(training_data)
print "minibatch size : ", minibatch_size
data_num = len(training_data)
print "--start--"
print "@@ Learn Character Recognition @@"
mb_size = minibatch_size
for i in range(0, epoch):
prg.show_progressxxx(i + 1, epoch)
xxx = random.sample(training_data, data_num)
x = []
d = []
for i in range(0, mb_size):
x.append(xxx[i][0])
d.append(xxx[i][1])
nn_obj.batch_learn(x[0:mb_size], d[0:mb_size], mb_size)
prg.end_progress()
print "dump obj..."
dp.obj_dump(nn_obj, "./learn-classification-batch.pkl")
return nn_obj
def classification(neuro_obj=None, epoch=100000, num_class=10):
print "initialize Neural Network."
if neuro_obj:
nn_obj = neuro_obj
else:
nn_obj = mln.Mln().make_neuralnet(
[28 * 28, 1000, num_class], ["sigmoid", "softmax"], 0.01
) # mnist classification
# else : nn_obj = mln.Mln().make_neuralnet([28*28, 1000, num_class], ['sigmoid', 'softmax'], 0.15) # mnist classification
# use weight decay.
# nn_obj.use_weight_decay(0.01) #
# nn_obj.use_weight_decay(0.001) #
# nn_obj.use_weight_decay(0.0001) # 20.02% error, with momentum 0.5.
nn_obj.use_weight_decay(0.00001) # 8.45% error, with momentum 0.5.
# nn_obj.unuse_weight_decay()
# use momentum.
# nn_obj.use_momentum(0.1)
nn_obj.use_momentum(0.5)
# nn_obj.use_momentum(0.9)
# nn_obj.unuse_momentum()
print "dump obj..."
dp.obj_dump(nn_obj, "./default-classification.pkl")
print "read training data and label."
training_data = dp.obj_load_gzip("../../mnist/mnist-training_all.pkl.gz")
print "data size : ", len(training_data)
print "label size : ", len(training_data)
data_num = len(training_data)
print "--start--"
print "@@ Learn Character Recognition @@"
for j in range(0, epoch):
prg.show_progressxxx(j + 1, epoch)
i = np.random.randint(data_num)
nn_obj.learn(training_data[i][0], training_data[i][1])
prg.end_progress()
print "dump obj..."
dp.obj_dump(nn_obj, "./learn-classification.pkl")
return nn_obj
def test_classification():
num_class = 10
print 'initialize Neural Network.'
nn_obj = dp.obj_load('./learn-classification-batch.pkl')
print 'read test data.'
test_data = dp.obj_load_gzip('../../mnist/mnist-test_data.pkl.gz')
print 'read test label.'
test_label = dp.obj_load_gzip('../../mnist/mnist-test_label.pkl.gz')
print 'data size : ', len(test_data)
print 'label size : ', len(test_label)
data_num = len(test_data)
prediction_error = []
prediction_recog = []
print '--start--'
print '@@ Test Character Recognition @@'
for j in range(0, data_num):
prg.show_progressxxx(j+1, data_num)
num_recog, list_recog = nn_obj.test(test_data[j], test_label[j])
if test_label[j][num_recog] == 0:
prediction_error.append((test_label[j], num_recog))
prg.end_progress()
count = len(prediction_error)
for item in prediction_error:
print "Truth:", np.argmax(item[0]), ", --> But predict:", item[1]
print ''
print 'error : ', count, '/', data_num, '(', count * 1.0 / data_num * 100.0, '%)'
return
def classification(neuro_obj = None, epoch = 100000, minibatch_size = 1):
num_class = 10
print 'initialize Neural Network.'
if neuro_obj: nn_obj = neuro_obj
else : nn_obj = mln.Mln().make_neuralnet([28*28, 1000, num_class], ['sigmoid', 'softmax'], 0.15) # mnist classification
# else : nn_obj = mln.Mln().make_neuralnet([28*28, 1000, num_class], ['sigmoid', 'softmax'], 0.01) # mnist classification
print "dump obj..."
dp.obj_dump(nn_obj, './default-classification-batch.pkl')
print 'read training data and label.'
training_data = dp.obj_load_gzip('../../mnist/mnist-training_all.pkl.gz')
print 'data size : ', len(training_data)
print 'label size : ', len(training_data)
print "minibatch size : ", minibatch_size
data_num = len(training_data)
print '--start--'
print '@@ Learn Character Recognition @@'
mb_size = minibatch_size
for i in range(0, epoch):
prg.show_progressxxx(i+1, epoch)
xxx = random.sample(training_data, data_num)
x = []
d = []
for i in range(0, mb_size):
x.append(xxx[i][0])
d.append(xxx[i][1])
nn_obj.batch_learn(x[0:mb_size], d[0:mb_size], mb_size)
prg.end_progress()
print "dump obj..."
dp.obj_dump(nn_obj, './learn-classification-batch.pkl')
return nn_obj
#from sklearn.preprocessing import LabelBinarizer
start = time.time()
# 多層パーセプトロン
learning_count = 50000
neuro_obj = mln.Mln().make_neuralnet([2, 3, 1], ['sigmoid', 'sigmoid'], 0.01, solved = 'classification')
# Initialization
dp.obj_dump(neuro_obj, './default-br.dump')
# XORの入出力データ
input_data = [[0., 0.], [0., 1.], [ 1., 0.], [ 1., 1.]]
teach_data = [ [0.], [1.], [1.], [0.]]
data_num = len(input_data)
print '--start--'
print '@@ Learn Character Recognition @@'
for j in range(0, learning_count):
prg.show_progressxxx(j+1, learning_count)
for i in range(0, 4):
neuro_obj.learn(input_data[i], teach_data[i])
prg.end_progress()
print ''
dp.obj_dump(neuro_obj, './learn-br.dump')
print("elapsed_time:{0} sec.".format(time.time() - start))
training_data = \
[\
(np.array([[0.], [0.]]), np.array([0.])),\
(np.array([[0.], [1.]]), np.array([1.])),\
(np.array([[1.], [0.]]), np.array([1.])),\
(np.array([[1.], [1.]]), np.array([0.])),\
]
# online learning, such as take all training data to NN sequecially.
dp.obj_dump(nn_obj, './default-fitting.pkl')
print "-----weight weight-----:"; print [item.w for item in nn_obj.weight]
print "-----weight bios-----:"; print [item.b for item in nn_obj.weight]
for i in range(0, epoch):
prg.show_progressxxx(i+1, epoch)
for j in range(0, 4):
nn_obj.learn(training_data[j][0], training_data[j][1])
prg.end_progress()
dp.obj_dump(nn_obj, './learn-fitting.pkl')
print "@@@@@TEST@@@@@"
nn_obj.feedforward(training_data[0][0], training_data[0][1])
print nn_obj.node[-1].z
nn_obj.feedforward(training_data[1][0], training_data[1][1])
print nn_obj.node[-1].z
nn_obj.feedforward(training_data[2][0], training_data[2][1])
print nn_obj.node[-1].z
nn_obj.feedforward(training_data[3][0], training_data[3][1])
