def run(self):
self.train()
with tf.Session() as sess:
tf.global_variables_initializer().run()
dataset = Imagenet.Cifar()
train_data, train_label, test_data, test_label = dataset.getdata()
train_label = de_onehot(train_label)
test_label = de_onehot(test_label)
path = "Equal/32-64-128-256-375-08-05"
def run(epochs):
with tf.Session() as sess:
tf.global_variables_initializer().run()
dataset = Imagenet.Cifar()
trX, trY, teX, teY = dataset.getdata()
filetime = datetime.now().strftime("%Y_%m_%d_%H_%M_%S")
#path = "Networkfile/" + "convAENN_noise"
path = "Networkfile/convAENN" + filetime
saver = NNutils.save(path, sess)
writer, merged = NNutils.graph(path, sess)
test_indices = np.arange(len(teX))
np.random.shuffle(test_indices)
test_indices = test_indices[0:(batch_size)]
st_time = datetime.now()
for i in range(epochs):
print(i, st_time)
for start, end in zip(range(0, len(trX), batch_size),
range(batch_size, len(trX), batch_size)):
summary, _, loss_nn, loss_ae, learning_rate, step = sess.run(
[merged, trainop, cost_NN, cost_AE, lr, global_step],
feed_dict={
X: trX[start:end],
Y: trY[start:end],
dropout_conv: 0.5,
dropout_fc: 0.5
})
if step % 50 == 0:
writer.add_summary(summary, step)
print(step, datetime.now(), loss_nn, loss_ae,
learning_rate)
loss_nn, loss_ae, accuracy = sess.run(
[cost_NN, cost_AE, acc_op],
feed_dict={
X: teX[test_indices],
Y: teY[test_indices],
dropout_conv: 1.0,
dropout_fc: 1.0
})
print("test results : ", accuracy, loss_nn, loss_ae)
saver.save(sess, path + "/model.ckpt", step)
# im = im.astype('uint8')
# im = Image.fromarray(im[0])
# im.save('convAENN.jpg')
end_time = datetime.now()
print("걸린 시간 = ", end_time - st_time)
def run(epochs):
with tf.Session() as sess:
tf.global_variables_initializer().run()
dataset = Imagenet.Cifar()
trX, trY, teX, teY = dataset.getdata()
filetime = datetime.now().strftime("%Y_%m_%d_%H_%M_%S")
#path = "Networkfile/convAE_sep" + "2017_03_27_20_32"
path = "Networkfile/convAENN_sep" + filetime
saver = NNutils.save(path, sess)
writer, merged = NNutils.graph(path, sess)
test_indices = np.arange(len(teX))
np.random.shuffle(test_indices)
test_indices = test_indices[0:(batch_size)]
#run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
#run_metadata = tf.RunMetadata()
st_time = datetime.now()
for i in range(epochs):
print(i, st_time)
for start, end in zip(range(0, len(trX), batch_size),
range(batch_size, len(trX), batch_size)):
summary, _, _, cost_ae, cost_nn, step = sess.run(
[
merged, trainop_unsuper, trainop_super, cost_unsuper,
cost_super, global_step
],
feed_dict={
x: trX[start:end],
y: trY[start:end],
dropout_conv: 1.0,
dropout_fc: 0.8
})
if step % 50 == 0:
writer.add_summary(summary, step)
print(step, datetime.now(), cost_nn, cost_ae)
accuracy, cost_nn, cost_ae = sess.run(
[acc_op, cost_super, cost_unsuper],
feed_dict={
x: teX[test_indices],
y: teY[test_indices],
dropout_conv: 1.0,
dropout_fc: 1.0
})
saver.save(sess, path + "/model.ckpt", step)
print("test results : ", accuracy, cost_nn, cost_ae)
# _, loss_super, step = sess.run([trainop_super, cost_super, global_step], feed_dict={y: trY[start:end]
# ,dropout_conv: 0.8, dropout_fc : 0.6})
#tf.graph_util.convert_variables_to_constants(sess, )
# if step % 50 == 0:
# writer.add_summary(summary, step)
# print(step, loss_super, loss_unsuper)
#print(np.shape(trX))
#summary, accuracy, loss = sess.run([merged, acc_op, cost], feed_dict={ X: teX[test_indices], Y: teY[test_indices]})
#print(step, datetime.now(), loss_unsuper, loss_super, learning_rate)
# loss_su, loss_un, accuracy, step = sess.run([cost_super, cost_unsuper, acc_op, global_step], feed_dict={X: teX[test_indices], Y: teY[test_indices],
# dropout_conv : 1.0, dropout_fc : 1.0})
# print("test results : ", accuracy, loss_super, loss_unsuper)
# saver.save(sess, path + "/model.ckpt", step)
end_time = datetime.now()
print("걸린 시간 = ", end_time - st_time)
with tf.name_scope(
"accuracy"): #accuracy는 tf를 이용하는게 좋다. tensorboard로 볼 수 있기 때문
predict_op = tf.equal(tf.argmax(Y, 1), tf.argmax(py_x, 1))
acc_op = tf.reduce_mean(tf.cast(predict_op, "float"))
tf.summary.scalar("accuracy", acc_op)
#네트워크 저장 관련
#saver = tf.train.Saver()
#global_step = tf.Variable(0, name='global_step', trainable=False)
#ckpt_dir = "./Networkfile/conv"
#if not os.path.exists(ckpt_dir):
# os.makedirs(ckpt_dir)
#이미지넷
imagenet = Imagenet.Imagenet()
train_num, test_num = imagenet.getnum()
with tf.Session() as sess:
#초기 설정
#tf.initialize_all_variables().run()
tf.global_variables_initializer().run()
test_indices = np.arange(test_num)
np.random.shuffle(test_indices)
test_indices = test_indices[0:batch_size]
#그래프 관련
graphPath = "./logs/conv_logs"
if not os.path.exists(graphPath):
os.makedirs(graphPath)
shutil.rmtree(graphPath)
def run(epochs):
with tf.Session() as sess:
tf.global_variables_initializer().run()
dataset = Imagenet.Cifar()
trX, trY, teX, teY = dataset.getdata()
train_gray, test_gray = dataset.getgray()
#train_gray = train_gray.reshape(-1, 32, 32, 1)
#test_gray = test_gray.reshape(-1, 32, 32, 1)
filetime = datetime.now().strftime("%Y_%m_%d_%H_%M")
path = "ConvAE/" + "rgb2gray"
#path = "Networkfile/convAE" + filetime
saver = NNutils.save(path, sess)
writer, writer_test, merged = NNutils.graph(path, sess)
test_indices = np.arange(len(teX))
np.random.shuffle(test_indices)
test_indices = test_indices[0:(batch_size)] #deconv때문에 batchsize보다 크면 문제가 발생한다.
st_time = datetime.now()
for i in range(epochs):
print(i, st_time)
for start, end in zip(range(0, len(trX), batch_size), range(batch_size, len(trX), batch_size)):
summary,\
_, loss, \
learning_rate, step = sess.run([merged,
trainop, cost,
# trainop_nn, cost_nn,
lr, global_step],
feed_dict={ X: trX[start:end],
# Y: trY[start:end],
x_gray: train_gray[start:end],
dropout_conv : 0.8, dropout_fc : 1.0})
#print(loss)
if step % 50 == 0:
writer.add_summary(summary, step)
print(step, datetime.now(), loss, learning_rate)
loss, arr = sess.run([cost, X_],
feed_dict={X: teX,
# Y: teY[0:batch_size],
x_gray: test_gray,
dropout_conv : 1.0, dropout_fc : 1.0})
# arr_uint = arr.astype('uint8')
# #arr_uint = arr_uint.reshape(-1, 32 * 32 * 1)
# im = Image.fromarray(arr_uint[0])
# im.show()
print("test results : ", loss)
saver.save(sess, path + "/model.ckpt", step)
end_time = datetime.now()
print("걸린 시간 = ", end_time - st_time)
# 평가
# SVM 학습
x_train = sess.run(Z, feed_dict={X: trX,
x_gray: test_gray,
dropout_conv: 1.0,
dropout_fc: 1.0,
})
x_train = x_train.reshape(len(x_train), -1)
y_train = np.argmax(trY, 1)
print(x_train.shape)
# SVM 예측 정확도 계산
x_test = sess.run(Z, feed_dict={X: teX,
x_gray: test_gray,
dropout_conv: 1.0,
dropout_fc: 1.0,
})
x_test = x_test.reshape(len(x_test), -1)
y_test = np.argmax(teY, 1)
accuracy = 0
iteration = 50
for i in range(iteration):
print(i)
clf = svm.LinearSVC(max_iter=200)
clf.fit(x_train, y_train)
acc = clf.score(x_test, y_test)
print(acc)
accuracy += acc
accuracy /= iteration
print(accuracy)
with tf.name_scope("cost"):
nn_Y = NN(X)
cost = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=nn_Y, labels=Y))
train_op = tf.train.AdamOptimizer(0.001).minimize(cost, global_step)
tf.summary.scalar("cost", cost)
predict_op = tf.equal(tf.arg_max(Y, 1), tf.arg_max(nn_Y, 1))
acc_op = tf.reduce_mean(tf.cast(predict_op, "float"))
#데이터 관련
# mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
# trX, trY, teX, teY = mnist.train.images, mnist.train.labels, mnist.test.images, mnist.test.labels
dataset = Imagenet.Cifar()
trX, trY, teX, teY = dataset.getdata()
trX = trX.reshape(-1, 32 * 32 * 3)
teX = teX.reshape(-1, 32 * 32 * 3)
print(tf.shape(cost))
with tf.Session() as sess:
tf.global_variables_initializer().run()
batch_size = 128
st_time = datetime.now()
print(st_time)
savepath = "Networkfile/NN"
saver = NNutils.save(savepath, sess)
writer, merge = NNutils.graph("Networkfile/NN", sess)
def run(epochs):
with tf.Session() as sess:
tf.global_variables_initializer().run()
dataset = Imagenet.Cifar()
trX, trY, teX, teY = dataset.getdata()
print(teY.shape)
filetime = datetime.now().strftime("%Y_%m_%d_%H_%M")
path = "convAE/" + "rgb2rgb"
#path = "Networkfile/convAE" + filetime
saver = NNutils.save(path, sess)
writer, writer_test, merged = NNutils.graph(path, sess)
test_indices = np.arange(len(teX))
np.random.shuffle(test_indices)
test_indices = test_indices[0:batch_size]
st_time = datetime.now()
for i in range(epochs):
print(i, st_time)
for start, end in zip(range(0, len(trX), batch_size), range(batch_size, len(trX), batch_size)):
summary, _,\
loss, learning_rate,\
step = sess.run([merged, trainop,
cost, lr,
global_step],
feed_dict={ X: trX[start:end],
dropout_conv : 0.8, dropout_fc : 0.5})
if step % 50 == 0:
writer.add_summary(summary, step)
print(step, datetime.now(), loss, learning_rate)
loss, results = sess.run([cost, Z], feed_dict={X: teX,
dropout_conv: 1.0,
dropout_fc: 1.0,
})
print("test results : ", loss)
saver.save(sess, path + "/model.ckpt", step)
#
# image = image.astype('uint8')
# im = Image.fromarray(image[0])
# im.show()
# image = teX.astype('uint8')
# im = Image.fromarray(image[0])
# im.show()
end_time = datetime.now()
print("걸린 시간 = ", end_time - st_time)
#평가
# SVM 학습
loss, x_train = sess.run([cost, Z], feed_dict={X: trX,
dropout_conv: 1.0,
dropout_fc: 1.0,
})
x_train = x_train.reshape(len(x_train), -1)
y_train = np.argmax(trY, 1)
print(x_train.shape)
clf = svm.LinearSVC(max_iter=500, random_state=2)
clf.fit(x_train, y_train)
# SVM 예측 정확도 계산
loss, x_test = sess.run([cost, Z], feed_dict={X: teX,
dropout_conv: 1.0,
dropout_fc: 1.0,
})
x_test = x_test.reshape(len(x_test), -1)
y_test = np.argmax(teY, 1)
accuracy = 0
iteration = 50
for i in range(iteration):
print(i)
clf = svm.LinearSVC(max_iter=200)
clf.fit(x_train, y_train)
acc = clf.score(x_test, y_test)
print(acc)
accuracy += acc
accuracy /= iteration
print(accuracy)
def run(self, epochs):
self.train_unsuper()
#self.train_super()
with tf.Session() as sess:
tf.global_variables_initializer().run()
dataset = Imagenet.Cifar()
trX, trY, teX, teY = dataset.getdata()
filetime = datetime.now().strftime("%Y_%m_%d_%H_%M")
path = "Networkfile/convAE_sep_class"
#path = "Networkfile/convAENN_sep" + filetime
saver = NNutils.save(path, sess)
writer, merged = NNutils.graph(path, sess)
test_indices = np.arange(len(teX))
np.random.shuffle(test_indices)
test_indices = test_indices[0:(self.batch_size)]
st_time = datetime.now()
for i in range(epochs):
print(i, st_time)
for start, end in zip(
range(0, len(trX), self.batch_size),
range(self.batch_size, len(trX), self.batch_size)):
_ = sess.run(self.trainop_unsuper,
feed_dict={
self.x: trX[start:end],
self.dropout_conv: 0.8,
self.dropout_fc: 0.6
})
# graph_def = graph_pb2.GraphDef()
# output_names = ""
# tf.graph_util.convert_variables_to_constants(sess, graph_def, output_names)
#print(cost_ae)
output = sess.run(self.z,
feed_dict={
self.x: trX[start:end],
self.dropout_conv: 0.8,
self.dropout_fc: 0.6
})
z = tf.constant(output)
print(z.shape)
# cost_nn, _ = sess.run([self.cost_super, self.trainop_super],
# feed_dict={self.x_nn: self.z, self.y: trY[start:end],
# self.dropout_conv: 0.8, self.dropout_fc : 0.6})
#
#_, loss_super, step = sess.run([trainop_super, cost_super, global_step], feed_dict={x:trX[0:0], y: trY[start:end]
# ,dropout_conv: 0.8, dropout_fc : 0.6})
# if step % 50 == 0:
# writer.add_summary(summary, step)
# print(step, loss_super, loss_unsuper)
#print(np.shape(trX))
#summary, accuracy, loss = sess.run([merged, acc_op, cost], feed_dict={ X: teX[test_indices], Y: teY[test_indices]})
#print(step, datetime.now(), loss_unsuper, loss_super, learning_rate)
# loss_su, loss_un, accuracy, step = sess.run([cost_super, cost_unsuper, acc_op, global_step], feed_dict={X: teX[test_indices], Y: teY[test_indices],
# dropout_conv : 1.0, dropout_fc : 1.0})
# print("test results : ", accuracy, loss_super, loss_unsuper)
# saver.save(sess, path + "/model.ckpt", step)
end_time = datetime.now()
print("걸린 시간 = ", end_time - st_time)