情境
我正在建立一个用tensorfow进行图像识别的模型。事实上，我正在尝试保存我的模型，然后将其恢复以进行预测。
程序
我建立了一个具有这种结构的CNN
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_3 (Conv2D) (None, 204, 204, 32) 896
_________________________________________________________________
dropout_3 (Dropout) (None, 204, 204, 32) 0
_________________________________________________________________
conv2d_4 (Conv2D) (None, 204, 204, 32) 9248
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 102, 102, 32) 0
_________________________________________________________________
flatten_2 (Flatten) (None, 332928) 0
_________________________________________________________________
dense_3 (Dense) (None, 512) 170459648
_________________________________________________________________
dropout_4 (Dropout) (None, 512) 0
_________________________________________________________________
dense_4 (Dense) (None, 39) 20007
_________________________________________________________________
activation_1 (Activation) (None, 39) 0
=================================================================
Total params: 170,489,799
Trainable params: 170,489,799
Non-trainable params: 0
我将图表和变量保存在保护程序中。我有目的地命名我的占位符 X 和 Y ，并命名预测（这是一个操作）：输出。命名它们的目的是能够恢复它们并将它们用于预测。
我做了什么
以下是我使用的代码
定义＆amp;设计模型结构
def create_model(X,Y):
# Convolutional Layer #1
conv1 = tf.layers.conv2d(
inputs=X,
filters=32,
kernel_size=[3, 3],
padding="same",
activation=tf.nn.relu)
print('conv1 OUTPUT shape: ',conv1.shape)
# Dropout layer #1
dropout1 = tf.layers.dropout(
inputs=conv1, rate=0.2, training='TRAIN' == tf.estimator.ModeKeys.TRAIN)
print('drop1 OUTPUT shape: ',dropout1.shape)
# Convolutional Layer #2
conv2 = tf.layers.conv2d(
inputs=dropout1,
filters=32,
kernel_size=[3, 3],
padding="same",
activation=tf.nn.relu)
print('conv2 OUTPUT shape: ',conv2.shape)
# Pooling Layer #2
pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2],strides=2)
print('pool2 OUTPUT shape: ',pool2.shape)
pool2_flat = tf.reshape(pool2, [-1, pool2.shape[1]*pool2.shape[2]*pool2.shape[3]])
# Dense layer #3
dense3 = tf.layers.dense(inputs=pool2_flat, units=512, activation=tf.nn.relu)
print('dense3 OUTPUT shape: ',dense3.shape)
# Dropout layer #3
dropout3 = tf.layers.dropout(
inputs=dense3, rate=0.5, training='TRAIN' == tf.estimator.ModeKeys.TRAIN)
print('drop3 OUTPUT shape: ',dropout3.shape)
# Dense layer #4
Z = tf.layers.dense(inputs=dropout3, units=39, activation=tf.nn.sigmoid)
print('dense4 OUTPUT shape: ',Z.shape)
#Threshold
#Z1 = tf.to_int32(Z > 0.5)
#print('OUTPUT: shape output after threshold ',Z1.shape)
# Calculating cost
cost= tf.reduce_mean(Y * - tf.log(tf.clip_by_value(Z,1e-10,1.0)) + (1 - Y) * - tf.log(tf.clip_by_value(1 - Z,1e-10,1.0)),axis=0)
print('cost: shape of cost: ',cost.shape)
cost= tf.reshape(cost, [1, 39])
print('cost reshaped: shape of cost reshaped: ',cost.shape)
#Optimizer
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cost)
#Metric paramaters
predicted_positive = predicted_positives(Y,Z)
true_and_possible_positive = true_and_possible_positives(Y,Z)
#naming operations
output1=tf.multiply(Z,1,name='the_output_for_prediction')
cost1=tf.multiply(cost,1,name='cost')
return output1,predicted_positive,true_and_possible_positive,optimizer,cost1
创建模型＆amp;培训与培训拯救和评估第一纪元后的预测（工作正常）
# For a given (X_train,Y_train)&(X_valid,Y_train) and some hyperparam( some cannot be changed: it trains a model, saves it and returns the metric values.
#def train_model(X_train,Y_train,X_valid,Y_valid,num_epochs,minibatch_size=32):
#Reset tensorflow graph
ops.reset_default_graph()
# hyperparameters
num_epochs=1
minibatch_size=32
m=X_train.shape[0]
#Initialize metric variables
costs = []
precision=[]
recall=[]
costs_v = []
precision_v=[]
recall_v=[]
m=X_train.shape[0]
#Placeholders
X = tf.placeholder(tf.float32, [None, X_train.shape[1],X_train.shape[2],X_train.shape[3]],name="X")
Y = tf.placeholder(tf.float32, [None, Y_train.shape[1]],name="Y")
#Graph construction
output,predicted_positives,true_and_possible_positives,optimizer,cost=create_model(X,Y)
# Initializer for all the variables globally
init = tf.global_variables_initializer()
#Creating tf session and running it (i.e initialize the global variables)
sess=tf.Session()
sess.run(init)
# Create model saver
#model_saver = tf.train.Saver()
saver = tf.train.Saver()
#Training algorithm
#Running epochs
for epoch in range(num_epochs):
########################################### training ################################################
print('training epoch number :',epoch)
#initialize parameters of metrics
predicted_po=0
true_po=0
possible_po=0
#initialize cost param and mini_batches
minibatch_cost = 0.
num_minibatches = int(m / minibatch_size) # number of minibatches of size minibatch_size in the train set
minibatches = random_mini_batches(X_train, Y_train, minibatch_size)
#Running batches
for minibatch in minibatches:
#minibatch extraction
minibatch_X=X_train[minibatch,]
minibatch_Y=Y_train[minibatch,]
# IMPORTANT: The line that runs the graph on a minibatch.
# Run the session to execute the optimizer, cost and metric parameters , the feedict should contain a minibatch for (X,Y).
_ , temp_cost = sess.run([optimizer, cost], feed_dict={X:minibatch_X, Y:minibatch_Y})
predicted_p,true_p_p = sess.run([predicted_positives, true_and_possible_positives], feed_dict={X:minibatch_X, Y:minibatch_Y})
#calculate cost,predicted_true_possible positives per batch.
minibatch_cost += temp_cost / num_minibatches
predicted_po+=predicted_p
true_po+=true_p_p[0]
possible_po+=true_p_p[1]
#print batch param metric
#print('predicted_p per batch ',noi,' ',predicted_p)
#print('true_p per bartch ',noi,' ',true_p_p[0])
#print('possible_p per batch ',noi,' ',true_p_p[1])
'''
# Print the cost every epoch
# print ("Cost after epoch %i: %f" % (epoch, minibatch_cost))
print('precision_on_training epoch: ',epoch,' ',true_po/predicted_po)
print('recall on training epoch ',epoch,' ', true_po/possible_po)
'''
#appending tables of costs and metrics on training
costs.append(minibatch_cost)
precision.append(true_po/predicted_po)
recall.append(true_po/possible_po)
minibatches=None
minibatch_X=None
minibatch_Y=None
###################################### Validation ###################################################
print('validation epoch number :',epoch)
#create minibatches for validation
#initialize parameters of metrics
m_v=X_valid.shape[0]
minibatch_cost_v = 0.
num_minibatches_v = int(m_v / minibatch_size) # number of minibatches validation of size minibatch_size_v in the validation set
predicted_po_v=0
true_po_v=0
possible_po_v=0
minibatches_valid=random_mini_batches(X_valid, Y_valid, minibatch_size)
#running batches
for minibatch in minibatches_valid:
#print('batch validation number ', noi_v, ' / ',num_minibatches_v)
minibatch_X_valid=X_train[minibatch,]
minibatch_Y_valid=Y_train[minibatch,]
temp_cost_v = sess.run(cost, feed_dict={X:minibatch_X_valid, Y:minibatch_Y_valid})
predicted_p_v,true_p_p_v = sess.run([predicted_positives, true_and_possible_positives], feed_dict={X:minibatch_X_valid, Y:minibatch_Y_valid})
minibatch_cost_v += temp_cost_v / num_minibatches_v
predicted_po_v+=predicted_p_v
true_po_v+=true_p_p_v[0]
possible_po_v+=true_p_p_v[1]
'''
#Printing precision and recall on validation
print('precision_on_validation epoch: ',epoch,' ',true_po_v/predicted_po_v)
print('recall on validation epoch ',epoch,' ', true_po_v/possible_po_v)
'''
#appending tables of costs and metrics on training
costs_v.append(minibatch_cost_v)
precision_v.append(true_po_v/predicted_po_v)
recall_v.append(true_po_v/possible_po_v)
minibatches_valid=None
minibatch_X_valid=None
minibatch_Y_valid=None
#A simple test to check whether a restored model has exactly the same values (w and graph) as the original. Testing on a new output.
prediction_first_epoch=sess.run([output], feed_dict={X:X_train[[0]], Y:Y_train[[0]]})
print(np.mean(prediction_first_epoch))
#Save model
#global_step: used for checkpoints
saver.save(sess, 'ok/',global_step=1000)
##Save different training parameters.
costs_array=np.asarray(costs)
precision_array=np.asarray(precision)
recall_array=np.asarray(recall)
costs_v_array=np.asarray(costs_v)
precision_v_array=np.asarray(precision_v)
recall_v_array=np.asarray(recall_v)
np.save('costs_array', costs_array)
np.save('precision_array', precision_array)
np.save('recall_array', recall_array)
np.save('costs_v_array', costs_v_array)
np.save('precision_v_array', precision_v_array)
np.save('recall_v_array', recall_v_array)
sess.close()
恢复已保存的模型并尝试使用与之前相同的数据进行预测
ops.reset_default_graph()
sess=tf.Session()
#Load meta_graph and restore weights
saver = tf.train.import_meta_graph('ok/-1000.meta')
saver.restore(sess,tf.train.latest_checkpoint('ok/'))
#put graph in variable and load adequate tensors & operations
graph = tf.get_default_graph()
X = graph.get_tensor_by_name("X:0")
Y = graph.get_tensor_by_name("Y:0")
output=graph.get_operation_by_name("the_output_for_prediction")
output_vis= sess.run(output, feed_dict={X:X_train[[0]], Y:Y_train[[0]]})
print(output_vis)
sess.close()
结果：
INFO:tensorflow:Restoring parameters from ok/-1000
None
问题
即使我在默认图表中恢复了已保存的图表并提取了两个占位符和预测操作，我也不明白为什么我会获得无值。
答案 0 :(得分：0)
[问题解决]
我在调用这个操作，这就是为什么我得到一个无价值的原因。
相反，我需要调用与该操作相关联的张量并用输入来提供它。
所以不要这样做：
output=graph.get_operation_by_name("the_output_for_prediction")
我应该这样做：
output=graph.get_tensor_by_name("the_output_for_prediction:0")
PS：＆＃34;：0＆＃34;是如此重要。如果你写＆＃34; the_output_for_prediction＆＃34; ，这不会奏效！