def loadMnist():
f = gzip.open('mnist.pkl.gz', 'rb')
train_set, valid_set, test_set = cPickle.load(f)
f.close()
trainingData = MakeData.shared(train_set)
validationData = MakeData.shared(valid_set)
testData = MakeData.shared(test_set)
return trainingData, validationData, testData
def experiment_0():
datas = MakeData.make1()
for _ in range(10):
# datas.extend(makeData(testfunc_circle, 100, 0, 2*np.pi, 10))
# datas.extend(makeData(testfunc_cubic, 100, 0, 2*np.pi, 0.1))
print(datas)
# datas.append(np.array(data))
# 転置
invdata = np.array(datas).T
print(invdata)
plt.plot(invdata[0], invdata[1])
plt.show()
model = hmm.GaussianHMM(n_components=10, covariance_type="full")
model.fit(datas)
'''
# 学習結果を表示
print("startprob_")
print(model.startprob_)
print("means_")
print(model.means_)
print("covars_")
print(model.covars_)
print("transmat_")
print(model.transmat_)
'''
# 推定．連続する同状態はカットして，繊維の様子だけ取り出す
pre = model.predict(datas)
result = []
for p in pre:
if len(result) == 0 or p != result[-1]:
result.append(p)
print(result)
# makeData し直して推定した場合の状態遷移列を比較したい
for _ in range(1000):
datas = MakeData.make1_half()
pre = model.predict(datas)
result_2 = []
for p in pre:
if len(result_2) == 0 or p != result_2[-1]:
result_2.append(p)
# print(result)
if result != result_2:
print("Different")
print(result_2)
break
def loadCifar100():
train_set, test_set = cifar100.load_data()
X, Y = train_set
X = X.reshape((X.shape[0], X.shape[1] * X.shape[2] * X.shape[3]))
Y = Y.reshape((Y.shape[0],))
trainingData = MakeData.shared((X[0:40000, :], Y[0:40000]))
validationData = MakeData.shared((X[40000:50000, :], Y[40000:50000]))
X, Y = test_set
X = X.reshape((X.shape[0], X.shape[1] * X.shape[2] * X.shape[3]))
Y = Y.reshape((Y.shape[0],))
testData = MakeData.shared((X, Y))
return trainingData, validationData, testData
def data_augmentation():
#輝度をあげる
os.system('bash ./initface.sh')
DA.sepia('sp')
DA.Shape(10, -1, 'spsp', '0') #学習画像データを水増し（データ拡張）を行う
mizumashi_data = ImageDataGenerator()
mizumashi_generator = mizumashi_data.flow_from_directory(
directory=root_dir,
target_size=(img_size, img_size),
batch_size=batch_size,
shuffle=True)
#テスト画像データを水増しする。
val_datagen = ImageDataGenerator()
val_gen = val_generator = val_datagen.flow_from_directory(
directory=root_dir,
target_size=(img_size, img_size),
batch_size=batch_size,
shuffle=False)
valX, valy = val_gen.next()
return (mizumashi_generator, val_generator, valX, valy)
def shrink(train, cv, test):
row = train.shape[0]
nums = np.sum(train + cv + test, axis=1)
locs = list(map(lambda x: x[0], np.argwhere(nums > 0)))
outTrain = train[locs]
outTest = test[locs]
outCv = cv[locs]
return outTrain, outCv, outTest
def transpose(train, cv, test):
return np.transpose(train), np.transpose(cv), np.transpose(test)
maker = MakeData.ScipyMatMaker()
train = maker.ReadMat(TRAIN_FILE)
cv = maker.ReadMat(CV_FILE)
test = maker.ReadMat(TEST_FILE)
print(train.shape)
train, cv, test = shrink(train, cv, test)
train, cv, test = transpose(train, cv, test)
train, cv, test = shrink(train, cv, test)
train, cv, test = transpose(train, cv, test)
# with open(TRAIN_FILE, 'wb') as fs:
# pickle.dump(train, fs)
# with open(TEST_FILE, 'wb') as fs:
# pickle.dump(test, fs)
# with open(CV_FILE, 'wb') as fs:
# pickle.dump(cv, fs)
print(train.shape)
result_images = "./result/out.png"
# In[579]:
#輝度をあげます
import MakeData as DA
#重複防止
# In[597]:
os.system('./initface.sh ')
# In[598]:
#水増し部分
if (args[1] == '0'):
DA.sepia('sp')
# DA.data_eraser(int(110/2))
elif (args[1] == '1'):
DA.CE_gray('red', 'red')
DA.Shape(10, -1, 'red', '0')
# DA.data_eraser(int(110/2))
elif (args[1] == '2'):
DA.CE_gray('green', 'green')
DA.Shape(10, -1, 'green', '0')
# DA.data_eraser(int(110/2))
elif (args[1] == '3'):
DA.CE_gray('blue', 'blue')
DA.Shape(10, -1, 'blue', '0')
# DA.data_eraser(int(110/2))
elif (args[1] == '4'):
DA.Shape(10, -1, 'sp', '0')
heapSizeLower = int(sys.argv[1])
heapSizeUpper = int(sys.argv[2])
heapAmountLower = int(sys.argv[3])
heapAmountUpper = int(sys.argv[4])
f = open("mlpstatistics.txt", "w")
filename = "CorrectMoves.txt"
player = PP.PerfectPlayer()
#The loop that defines the start-states from which the programs will be run.
for heapSize in range(heapSizeLower, heapSizeUpper + 1):
for heapAmount in range(heapAmountLower, heapAmountUpper + 1):
"""The Supervised learning program is run 10 times for each start-state, and plays against the perfect player each time,
to get a winrate for statistics. The time used is also written to file as statistics. The data used is the same for a given state,
but it is sorted differently for each time."""
state = [heapSize] * heapAmount
print(state)
md = MD.MakeData()
md.setup(state)
timeAvg = 0
percent = 0
for i in range(10):
#sorting and separating the data into the needed categories.
data = np.loadtxt(filename)
stateLen = len(data[0])
boards = data[::2]
moves = data[1::2]
order = list(range(np.shape(boards)[0]))
np.random.shuffle(order)
boards = boards[order, :]
