我之前曾问过这个问题，但未能找到帮助。我有一个使用Tensorflow训练的模型，正在尝试加载。我可以加载仅使用tf.nn.xxxx函数编写的模型，但是我不能使用tf.layers编写的任何东西都可以加载。我花了几个月的时间试图找到解决办法，但似乎没有人在网上遇到相同的问题，这使我相信我在做一些非常愚蠢的事情。非常感谢您在此阶段提供的帮助
当我尝试在推理时间内通过tf.layers模型运行数据时，出现以下错误消息：
FailedPreconditionError (see above for traceback): Attempting to use uninitialized value Level3Decoding/conv3/batch_normalization/beta_1
[[Node: Level3Decoding/conv3/batch_normalization/beta_1/read = Identity[T=DT_FLOAT, _class=["loc:@Level3Decoding/conv3/batch_normalization/cond/FusedBatchNorm_3/Switch_2"], _device="/job:localhost/replica:0/task:0/device:GPU:0"](Level3Decoding/conv3/batch_normalization/beta_1)]]
[[Node: Level1Decoding/conv2d/BiasAdd_1/_811 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device_incarnation=1, tensor_name="edge_1214_Level1Decoding/conv2d/BiasAdd_1", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:CPU:0"]()]]
我正在训练一个用于细分的网络，其定义如下：
def uNet2D(x, REGULARIZER, KERNEL_SIZE, is_training):
regularizer = tf.contrib.layers.l2_regularizer(scale=REGULARIZER)
#L1 encode
with tf.variable_scope('Level1Encoding'):
with tf.variable_scope('conv1'):
conv1=tf.layers.conv2d(x,64,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same', name='conv1')
conv1 = tf.layers.batch_normalization(
inputs=conv1,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv1 = tf.nn.relu(conv1)
with tf.variable_scope('conv2'):
conv2=tf.layers.conv2d(conv1,64,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv2 = tf.layers.batch_normalization(
inputs=conv2,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv2 = tf.nn.relu(conv2)
with tf.variable_scope('conv3'):
conv3=tf.layers.conv2d(conv2,64,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv3 = tf.layers.batch_normalization(
inputs=conv3,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv3 = tf.nn.relu(conv3)
conv3mp=tf.layers.max_pooling2d(conv3,2,2,padding='same')
with tf.variable_scope('Level2Encoding'):
with tf.variable_scope('conv1'):
conv4=tf.layers.conv2d(conv3mp,128,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv4 = tf.layers.batch_normalization(
inputs=conv4,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv4 = tf.nn.relu(conv4)
with tf.variable_scope('conv2'):
conv5=tf.layers.conv2d(conv4,128,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv5 = tf.layers.batch_normalization(
inputs=conv5,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv5 = tf.nn.relu(conv5)
with tf.variable_scope('conv3'):
conv6=tf.layers.conv2d(conv5,128,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv6 = tf.layers.batch_normalization(
inputs=conv6,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv6 = tf.nn.relu(conv6)
conv6mp=tf.layers.max_pooling2d(conv6,2,2,padding='same')
with tf.variable_scope('Level3Encoding'):
with tf.variable_scope('conv1'):
conv7=tf.layers.conv2d(conv6mp,256,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv7 = tf.layers.batch_normalization(
inputs=conv7,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv7 = tf.nn.relu(conv7)
with tf.variable_scope('conv2'):
conv8=tf.layers.conv2d(conv7,256,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv8 = tf.layers.batch_normalization(
inputs=conv8,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv8 = tf.nn.relu(conv8)
with tf.variable_scope('conv3'):
conv9=tf.layers.conv2d(conv8,256,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv9 = tf.layers.batch_normalization(
inputs=conv9,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv9 = tf.nn.relu(conv9)
conv9mp=tf.layers.max_pooling2d(conv9,2,2,padding='same')
with tf.variable_scope('Level4Encoding'):
with tf.variable_scope('conv1'):
conv10=tf.layers.conv2d(conv9mp,512,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv10 = tf.layers.batch_normalization(
inputs=conv10,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv10 = tf.nn.relu(conv10)
with tf.variable_scope('conv2'):
conv11=tf.layers.conv2d(conv10,512,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv11 = tf.layers.batch_normalization(
inputs=conv11,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv11 = tf.nn.relu(conv11)
with tf.variable_scope('conv3'):
conv12=tf.layers.conv2d(conv11,512,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv12 = tf.layers.batch_normalization(
inputs=conv12,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv12 = tf.nn.relu(conv12)
conv12mp=tf.layers.max_pooling2d(conv12,2,2,padding='same')
with tf.variable_scope('Level5'):
with tf.variable_scope('conv1'):
conv13=tf.layers.conv2d(conv12mp,1024,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv13 = tf.layers.batch_normalization(
inputs=conv13,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv13 = tf.nn.relu(conv13)
with tf.variable_scope('conv2'):
conv14=tf.layers.conv2d(conv13,1024,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv14 = tf.layers.batch_normalization(
inputs=conv14,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv14 = tf.nn.relu(conv14)
with tf.variable_scope('conv3'):
conv15=tf.layers.conv2d(conv14,1024,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv15 = tf.layers.batch_normalization(
inputs=conv15,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv15 = tf.nn.relu(conv15)
conv15=tf.layers.conv2d_transpose(conv15,512,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, strides=(2,2),padding='same')
with tf.variable_scope('Level4Decoding'):
inp = tf.concat([conv12,conv15],3)
with tf.variable_scope('conv1'):
conv16 = tf.layers.conv2d(inp,256,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv16 = tf.layers.batch_normalization(
inputs=conv16,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv16 = tf.nn.relu(conv16)
with tf.variable_scope('conv2'):
conv17=tf.layers.conv2d(conv16,256,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv17 = tf.layers.batch_normalization(
inputs=conv17,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv17 = tf.nn.relu(conv17)
with tf.variable_scope('conv3'):
conv18=tf.layers.conv2d(conv17,256,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv18 = tf.layers.batch_normalization(
inputs=conv18,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv18 = tf.nn.relu(conv18)
conv18=tf.layers.conv2d_transpose(conv18,256,(KERNEL_SIZE, KERNEL_SIZE),strides=(2,2), kernel_regularizer=regularizer, padding='same')
with tf.variable_scope('Level3Decoding'):
inp = tf.concat([conv9,conv18],3)
with tf.variable_scope('conv1'):
conv19 = tf.layers.conv2d(inp,128,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv19 = tf.layers.batch_normalization(
inputs=conv19,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv19 = tf.nn.relu(conv19)
with tf.variable_scope('conv2'):
conv20=tf.layers.conv2d(conv19,128,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv20 = tf.layers.batch_normalization(
inputs=conv20,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv20 = tf.nn.relu(conv20)
with tf.variable_scope('conv3'):
conv21=tf.layers.conv2d(conv20,128,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv21 = tf.layers.batch_normalization(
inputs=conv21,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv21 = tf.nn.relu(conv21)
conv21=tf.layers.conv2d_transpose(conv21,128,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, strides=(2,2),padding='same')
with tf.variable_scope('Level2Decoding'):
inp = tf.concat([conv6,conv21],3)
with tf.variable_scope('conv1'):
conv22 = tf.layers.conv2d(inp,64,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv22 = tf.layers.batch_normalization(
inputs=conv22,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv22 = tf.nn.relu(conv22)
with tf.variable_scope('conv2'):
conv23 = tf.layers.conv2d(conv22,128,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv23 = tf.layers.batch_normalization(
inputs=conv23,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv23 = tf.nn.relu(conv23)
with tf.variable_scope('conv3'):
conv24=tf.layers.conv2d(conv23,128,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv24 = tf.layers.batch_normalization(
inputs=conv24,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv24 = tf.nn.relu(conv24)
conv24=tf.layers.conv2d_transpose(conv24,64,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, strides=(2,2),padding='same')
with tf.variable_scope('Level1Decoding'):
inp = tf.concat([conv3,conv24],3)
with tf.variable_scope('conv1'):
conv25 = tf.layers.conv2d(inp,64,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv25 = tf.layers.batch_normalization(
inputs=conv25,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv25 = tf.nn.relu(conv25)
with tf.variable_scope('conv2'):
conv26 = tf.layers.conv2d(conv25,128,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv26 = tf.layers.batch_normalization(
inputs=conv25,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv26 = tf.nn.relu(conv26)
with tf.variable_scope('conv3'):
conv27 = tf.layers.conv2d(conv26,128,(KERNEL_SIZE, KERNEL_SIZE), kernel_regularizer=regularizer, padding='same')
conv27 = tf.layers.batch_normalization(
inputs=conv26,
axis=-1,
momentum=0.999,
epsilon=0.001,
center=True,
scale=True,
training = is_training)
conv27=tf.nn.relu(conv27)
convOUT = tf.layers.conv2d(conv27,1,(1,1), kernel_regularizer=regularizer, padding='same')
return convOUT
我正在使用以下脚本对其进行训练：
os.chdir(launch)
timestr = time.strftime("%Y%m%d-%H%M%S")
outDir = './TumorOUT_no_core/' + timestr
os.mkdir(outDir)
graphDir = './graphs/' + timestr
os.mkdir(graphDir)
os.mkdir(graphDir + '/training/')
os.mkdir(graphDir + '/testing/')
with open(outDir + '/hyperparams.txt', 'w+') as f:
for key, value in hypers.items():
f.write('%s:%s\n' % (key, value))
X = tf.placeholder(tf.float32, shape=[None, None, None, NUM_CHANNELS], name='X') #input
Y = tf.placeholder(tf.float32, shape = [None, None, None, 1], name='Y') #'labels'
is_training = tf.placeholder(tf.bool, name='is_training')
def run_model():
valCosts=[]
GLOBAL_STEP = 0
minLoss=10000000000
pred = uNet2D(X, BETA, KERNEL_SIZE, is_training)
cost = tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.reshape(Y,[-1]),logits=tf.reshape(pred,[-1])))
#with tf.variable_scope('prediction') as scope:
# t_pred = uNet2D(X, BETA, KERNEL_SIZE)
# t_cost = tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.reshape(Y,[-1]),logits=tf.reshape(t_pred,[-1])))
# scope.reuse_variables()
# v_pred = uNet2D(X, BETA, KERNEL_SIZE)
# v_cost = tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.reshape(Y,[-1]),logits=tf.reshape(v_pred,[-1])))
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
# this with statement updates the BN parameters if BN is being used
optimizer = tf.train.AdamOptimizer(learning_rate=LR).minimize(cost)
with tf.name_scope("training"):
tf.summary.scalar("training_cost", cost, collections=['training'])
for var in tf.trainable_variables():
tf.summary.histogram(var.op.name , var, collections=['training'])
with tf.name_scope("validation"):
tf.summary.scalar("validation_cost", cost, collections=['validation'])
#tf.summary.image("VALIDATION_X",X, collections=['validation'])
#tf.summary.image("VALIDATION_Y",Y, collections=['validation'])
#tf.summary.image("VALIDATION_PRED", v_pred, collections=['validation'])
saver = tf.train.Saver()
with tf.Session() as sess:
train_merge = tf.summary.merge_all(key='training')
validation_merge = tf.summary.merge_all(key='validation')
train_writer = tf.summary.FileWriter( graphDir + '/training/', sess.graph)
validation_writer = tf.summary.FileWriter( graphDir + '/testing/', sess.graph)
print('Beginning Session!')
sess.run(tf.global_variables_initializer())
print('Running Model!')
while GLOBAL_STEP <= MAX_ITER:
if GLOBAL_STEP % NUM_STEPS != 0:
x,y=training.drawBatch(BATCH_SIZE)
y=np.expand_dims(y,-1)
y[y>0]=1
flip = random.uniform(0,1)
if flip>0.5:
x = np.flip(x,2)
y = np.flip(y,2)
_, acc, c = sess.run([optimizer, train_merge, cost], feed_dict = {X: x, Y: y, is_training: 1})
train_writer.add_summary(acc, GLOBAL_STEP)
#train_writer.add_summary(summary, GLOBAL_STEP)
else:
x,y=validation.drawBatch(BATCH_SIZE)
y=np.expand_dims(y,-1)
y[y>0]=1
acc, c = sess.run([validation_merge,cost], feed_dict = {X: x, Y: y, is_training: 0})
validation_writer.add_summary(acc, GLOBAL_STEP)
#validation_writer.add_summary(summary, GLOBAL_STEP)
#save_path=saver.save(sess, outDir + '/model')
print(c)
if c < minLoss:
save_path=saver.save(sess, outDir + '/model')
minLoss=c
valCosts.append(c)
GLOBAL_STEP+=1
g= open(graphDir+'/val.pickle', 'w+')
pickle.dump([valCosts], g)
run_model()
我尝试使用以下代码加载它，加载成功，但是似乎我无法正确初始化变量：
launch=os.getcwd()
tf.reset_default_graph()
os.environ["CUDA_VISIBLE_DEVICES"]="0"
launch=os.getcwd()
testDir=launch + '/in_data/no_core/Validation/'
os.chdir(testDir)
testList = glob.glob('*mat')
os.chdir(launch)
sess = tf.Session()
new_saver = tf.train.import_meta_graph(launch + '/TumorOUT_no_core/20180829-142646/model.meta')
new_saver.restore(sess,tf.train.latest_checkpoint(launch + '/TumorOUT_no_core/20180829-142646/'))
X = tf.placeholder(tf.float32, shape=[None, None, None, 2], name='X') #input
Y = tf.placeholder(tf.float32, shape = [None, None, None, 1], name='Y') #'labels'
is_training = tf.placeholder(tf.bool, name='is_training')
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
os.chdir(testDir)
pred = uNet2D(X, .1, 3,is_training)
print("Loaded Weights")
for i in testList:
print(i)
data=loadmat(i)
mri=data['data']
mri=mri[:,:,:,0:2]
outt=np.zeros((mri.shape[0],mri.shape[1],mri.shape[2]))
mri=(mri-np.mean(mri))/np.var(mri)
mask=mri[:,:,:,0]>15
mri=np.expand_dims(mri,0)
roi=data['roi_mat']
for z in range(mri.shape[3]):
b1=sess.run(pred,feed_dict={X: mri[:,:,:,z,:],is_training:False})
b=b1[0,:,:,0]
b=sess.run(tf.nn.sigmoid(b))
r=roi[:,:,z]
r=r.astype(dtype=np.float32)
outt[:,:,z]=b
out={}
out['out']= outt
out['roi']= roi
savemat('./matlab/'+i,out)
print('done :)')