让我从一开始就开始。我在tensorflow 1.14中实现了基于not official Keras implementation的局部卷积层以进行图像修复（我已经对其进行了测试，并且可以在我的数据集上使用）。
此体系结构使用预训练的（图像网络）VGG16计算一些损耗项。令人遗憾的是，在tensorflow中实现的VGG无法正常工作（我已经尝试过使用this one)作为keras应用程序中的那个。因此，我使用了class将keras应用程序VGG16合并到了我的tensorflow 1.14中。代码。
一切正常，但随后我将混合精度训练（documentation）合并到我的代码中，而VGG16部分给出了以下错误：
Instructions for updating:
Call initializer instance with the dtype argument instead of passing it to the constructor
ERROR:tensorflow:==================================
Object was never used (type <class 'tensorflow.python.framework.ops.Tensor'>):
<tf.Tensor 'VGG16/model/IsVariableInitialized_3:0' shape=() dtype=bool>
If you want to mark it as used call its "mark_used()" method.
It was originally created here:
File "main.py",line 131,in <module>
psi_gt,psi_out,psi_comp,I_comp,layers = model.build_vgg(data_gt,unet_pconv,data_mask) File "/workspace/model.py",line 52,in build_vgg
vgg = vgg16.VGG16(image_shape=gt.shape,input_tensor=gt) File "/workspace/vgg.py",line
17,in __init__
self._build_graph(input_tensor) File "/workspace/vgg.py",line 35,in _build_graph
self.vgg16 = tf.keras.applications.VGG16(weights='imagenet',include_top=False,input_tensor=img) File "/usr/local/lib/python3.6/dist-
packages/tensorflow/python/keras/applications/__init__.py",line 70,in wrapper
return base_fun(*args,**kwargs) File "/usr/local/lib/python3.6/dist-
packages/tensorflow/python/keras/applications/vgg16.py",line 32,in VGG16
return vgg16.VGG16(*args,**kwargs) File "/usr/local/lib/python3.6/dist-
packages/keras_applications/vgg16.py",line 210,in VGG16
model.load_weights(weights_path) File "/usr/local/lib/python3.6/dist-
packages/tensorflow/python/keras/engine/training.py",line 162,in load_weights
return super(Model,self).load_weights(filepath,by_name) File
"/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/network.py",line
1424,in load_weights
saving.load_weights_from_hdf5_group(f,self.layers) File "/usr/local/lib/python3.6/dist-
packages/tensorflow/python/keras/saving/hdf5_format.py",line 759,in
load_weights_from_hdf5_group
K.batch_set_value(weight_value_tuples) File "/usr/local/lib/python3.6/dist-
packages/tensorflow/python/keras/backend.py",line 3071,in batch_set_value
get_session().run(assign_ops,feed_dict=feed_dict) File "/usr/local/lib/python3.6/dist-
packages/tensorflow/python/keras/backend.py",line 462,in get_session
_initialize_variables(session) File "/usr/local/lib/python3.6/dist-
packages/tensorflow/python/keras/backend.py",line 879,in _initialize_variables
[variables_module.is_variable_initialized(v) for v in candidate_vars]) File
"/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/backend.py",in
<listcomp>
[variables_module.is_variable_initialized(v) for v in candidate_vars]) File
"/usr/local/lib/python3.6/dist-packages/tensorflow/python/util/tf_should_use.py",line 193,in wrapped
return _add_should_use_warning(fn(*args,**kwargs))
==================================
ERROR:tensorflow:==================================
Object was never used (type <class 'tensorflow.python.framework.ops.Tensor'>):
<tf.Tensor 'VGG16/model/IsVariableInitialized_2:0' shape=() dtype=bool>
If you want to mark it as used call its "mark_used()" method.
It was originally created here:
File "main.py",data_mask)
File "/workspace/model.py",line 17,**kwargs))
==================================
ERROR:tensorflow:==================================
Object was never used (type <class 'tensorflow.python.framework.ops.Tensor'>):
<tf.Tensor 'VGG16/model/IsVariableInitialized_1:0' shape=() dtype=bool>
If you want to mark it as used call its "mark_used()" method.
It was originally created here:
File "main.py",**kwargs))
==================================
ERROR:tensorflow:==================================
Object was never used (type <class 'tensorflow.python.framework.ops.Tensor'>):
<tf.Tensor 'VGG16/model/IsVariableInitialized:0' shape=() dtype=bool>
If you want to mark it as used call its "mark_used()" method.
It was originally created here:
File "main.py",**kwargs))
==================================
Traceback (most recent call last):
File "main.py",data_mask)
File "/workspace/model.py",input_tensor=gt)
File "/workspace/vgg.py",in __init__
self._build_graph(input_tensor)
File "/workspace/vgg.py",input_tensor=img)
File "/usr/local/lib/python3.6/dist-
packages/tensorflow/python/keras/applications/__init__.py",**kwargs)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/applications/vgg16.py",**kwargs)
File "/usr/local/lib/python3.6/dist-packages/keras_applications/vgg16.py",in VGG16
model.load_weights(weights_path)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py",by_name)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/network.py",line 1424,self.layers)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/saving/hdf5_format.py",in load_weights_from_hdf5_group
K.batch_set_value(weight_value_tuples)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/backend.py",feed_dict=feed_dict)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/backend.py",in get_session
_initialize_variables(session)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/backend.py",in _initialize_variables
[variables_module.is_variable_initialized(v) for v in candidate_vars])
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/backend.py",in <listcomp>
[variables_module.is_variable_initialized(v) for v in candidate_vars])
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/util/tf_should_use.py",**kwargs))
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/variables.py",line 3083,in is_variable_initialized
return state_ops.is_variable_initialized(variable)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/state_ops.py",line 133,in is_variable_initialized
return ref.is_initialized(name=name)
AttributeError: 'Tensor' object has no attribute 'is_initialized'
对于混合精度培训，我使用的是NVIDIA泊坞窗。
在这里使用VGG16获得3张图像的特征图：
def build_vgg(gt,y_pred,mask):
vgg_layer = ['block1_pool','block2_pool','block3_pool']
vgg = vgg16.VGG16(image_shape=gt.shape,input_tensor=gt)
psi_gt = {}
psi_gt[vgg_layer[0]] = tf.identity(vgg[vgg_layer[0]],name='gt_vgg0')
psi_gt[vgg_layer[1]] = tf.identity(vgg[vgg_layer[1]],name='gt_vgg1')
psi_gt[vgg_layer[2]] = tf.identity(vgg[vgg_layer[2]],name='gt_vgg2')
vgg = vgg16.VGG16(image_shape=y_pred.shape,input_tensor=y_pred)
psi_out = {}
psi_out[vgg_layer[0]] = tf.identity(vgg[vgg_layer[0]],name='out_vgg0')
psi_out[vgg_layer[1]] = tf.identity(vgg[vgg_layer[1]],name='out_vgg1')
psi_out[vgg_layer[2]] = tf.identity(vgg[vgg_layer[2]],name='out_vgg2')
I_comp = (mask * gt) + ((1-mask) * y_pred)
vgg = vgg16.VGG16(image_shape=I_comp.shape,input_tensor=I_comp)
psi_comp = {}
psi_comp[vgg_layer[0]] = tf.identity(vgg[vgg_layer[0]],name='comp_vgg0')
psi_comp[vgg_layer[1]] = tf.identity(vgg[vgg_layer[1]],name='comp_vgg1')
psi_comp[vgg_layer[2]] = tf.identity(vgg[vgg_layer[2]],name='comp_vgg2')
return psi_gt,vgg_layer
在主脚本中使用的上一个功能：
import tensorflow as tf
import PConv
import model
import layers
import math
import os
import data
import utils
import numpy as np
import datetime
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
# Mixed precision training variable storage
def float32_variable_storage_getter(getter,name,shape=None,dtype=None,initializer=None,regularizer=None,trainable=True,*args,**kwargs):
storage_dtype = tf.float32 if trainable else dtype
variable = getter(name,shape,dtype=storage_dtype,initializer=initializer,regularizer=regularizer,trainable=trainable,**kwargs)
if trainable and dtype != tf.float32:
variable = tf.cast(variable,dtype)
return variable
# ==============================================================================
# SETTINGS
# ==============================================================================
path_ =''
batch_size = 16
best_val = math.inf
best_val_epoch = 0
patience = 0
stop = 300
epochs = 2000
steps_train = 25
steps_val = 8
template = '{},Epoch {},train_loss: {:.4f} - val_loss: {:.4f}'
path = path_ + 'tmp/'
if not os.path.isdir(path):
os.mkdir(path)
# ==============================================================================
# DATA
# ==============================================================================
X_train,m_train,y_train = data.get_filenames()
X_val,m_val,y_val = data.get_filenames(train=False)
# ==============================================================================
# DATASET
# ==============================================================================
train_dataset = tf.data.Dataset.from_tensor_slices((X_train,y_train))#(images,mask,gt))
train_dataset = train_dataset.map(data.load,num_parallel_calls=tf.data.experimental.AUTOTUNE)
train_dataset = train_dataset.batch(batch_size)
train_dataset = train_dataset.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
val_dataset = tf.data.Dataset.from_tensor_slices((X_val,y_val))#(images,gt))
val_dataset = val_dataset.map(data.load,num_parallel_calls=tf.data.experimental.AUTOTUNE)
val_dataset = val_dataset.batch(batch_size)
val_dataset = val_dataset.prefetch(buffer_size=1)
iterator = tf.data.Iterator.from_structure(train_dataset.output_types,train_dataset.output_shapes)
data_im,data_mask,data_gt = iterator.get_next()
# create the initialization operations
train_init_op = iterator.make_initializer(train_dataset)
val_init_op = iterator.make_initializer(val_dataset)
# ==============================================================================
# MODEL
# ==============================================================================
data_im = tf.cast(data_im,tf.float16)
data_mask = tf.cast(data_mask,tf.float16)
with tf.variable_scope('fp32_vars',custom_getter=float32_variable_storage_getter):
unet_pconv = model.pconv_unet(data_im,data_mask)
unet_pconv = tf.cast(unet_pconv,tf.float32)
data_mask = tf.cast(data_mask,tf.float32)
psi_gt,data_mask)
I_comp = tf.cast(I_comp,tf.float32)
# # ==============================================================================
# # LOSS
# # ==============================================================================
loss = utils.get_total_loss(unet_pconv,data_gt,psi_gt,layers)
lr = 0.0002
optimizer = utils.optimize(loss,lr)
saver = tf.train.Saver()
# # ==============================================================================
# # TRAINING
# # ==============================================================================
output_summary = tf.summary.image(name='output',tensor=unet_pconv)
merged = tf.summary.merge_all()
with tf.Session() as sess:
sess.run(tf.local_variables_initializer())
sess.run(tf.global_variables_initializer())
writer = tf.summary.FileWriter('graphs',sess.graph)
train_loss_,val_loss_ = [],[]
for epoch in range(epochs):
pred_time = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
tl,vl = [],[]
#Initialize iterator with training data
sess.run(train_init_op)
try:
for step in range (steps_train):
_,train_loss,summ = sess.run([optimizer,loss,merged])
writer.add_summary(summ,epoch)
tl.append(train_loss)
mean_train = utils.list_mean(tl)
train_loss_.append(mean_train)
except tf.errors.OutOfRangeError:
pass
if (epoch+1) % 1 == 0:
sess.run(val_init_op)
try:
for step in range (steps_val):
val_loss = sess.run([loss])
vl.append(val_loss)
mean_val = utils.list_mean(vl)
val_loss_.append(mean_val)
except tf.errors.OutOfRangeError:
pass
print(template.format(pred_time,epoch,mean_train,mean_val))
# early stopping
if mean_val < best_val:
print('Saving on epoch {0}'.format(epoch))
best_val = mean_val
patience = 0
best_val_epoch = epoch
saver.save(sess,path+'best_model')
else:
patience += 1
if patience == stop:
print('Early stopping at epoch: {}'.format(best_val_epoch))
break
# # ==============================================================================
# # SAVE CURVES
# # ==============================================================================
np.save(path_+'loss.npy',train_loss_)
np.save(path_+'val_loss.npy',val_loss_)
其优化过程如下：
def optimize(loss,learning_rate=1e-4):
U_vars = [var for var in tf.trainable_variables() if 'UNET' in var.name]
opt = tf.train.AdamOptimizer(learning_rate=learning_rate)
opt = tf.train.experimental.enable_mixed_precision_graph_rewrite(opt,loss_scale=128.0)
train_opt = opt.minimize(loss,var_list=U_vars)
return train_opt
我已经尝试修复了一段时间，但仍然不明白为什么在实施混合精度训练时它不起作用。随时询问更多细节。
如果您能伸出援手，那就太好了！预先谢谢你。