我的网络有两个步骤。有一个使用Connectionist时间分类（CTC）进行特征提取和识别的预训练网络。加载预训练的模型并提取feature layer ('conv5_3')，然后通过roi_pooling裁剪部分特征并使用CTC进行识别。

代码如下。

************重新训练了预训练的模型***********************

graph = tf.Graph()
    with graph.as_default():
       global_step = slim.get_or_create_global_step()
       input_data,input_labels,input_boxes = input_train_data.input_fn()
       input_boxes = tf.reshape(input_boxes,[input_boxes.shape[0]*2,-1])
       rows = tf.expand_dims(tf.range(input_boxes.shape[0]),1)/2
       add_index = tf.concat([tf.cast(rows,tf.float32),input_boxes],-1)
       index = tf.not_equal(tf.reduce_sum(add_index[:,4:],axis=1),0)
       input_boxes = tf.gather_nd(add_index,tf.where(index))
       # build model and loss
       net = pixel_link_symbol.PixelLinkNet(input_data,is_training = False)
       masks = pixel_link.tf_decode_score_map_to_mask_in_batch(net.pixel_pos_scores,net.link_pos_scores)
       f_saver = tf.train.Saver(max_to_keep=1000,write_version=tf.train.SaverDef.V2,save_relative_paths=True)
       sess_config = tf.ConfigProto(log_device_placement = False,allow_soft_placement = True)
       if flaGS.gpu_memory_fraction < 0:
          sess_config.gpu_options.allow_growth = True
       elif flaGS.gpu_memory_fraction > 0:
          sess_config.gpu_options.per_process_gpu_memory_fraction = flaGS.gpu_memory_fraction;
       session = tf.Session(graph=graph,config=sess_config)
       coordinator = tf.train.Coordinator()
       threads = tf.train.start_queue_runners(sess=session,coord=coordinator)

       tf.logging.info('Initialize from: ' + config.train.init_checkpoint)
       f_saver.restore(session,config.train.init_checkpoint)

**********提取特征（'conv5_3'），通过roi_pooling并使用CTC进行训练***************

       #Get feature layers
       conv5_3 = net.end_points['conv5_3']
       with tf.variable_scope("Recognition"):
          #Pass through recognition net
          r_net = regnet.ConstructRecNet(conv5_3)
          conv7_7 = r_net.end_points['pool7']
          #implement ROI Pooling
          input_boxes=tf.dtypes.cast(input_boxes,tf.int32)
          prob = roi_pooling(conv7_7,input_boxes,pool_height=1,pool_width=28)
          #Get features for CTC training
          prob = tf.transpose(prob,(1,2))  # prepare for CTC
          data_length = tf.fill([tf.shape(prob)[1]],tf.shape(prob)[0])  # input seq length,batch size
          ctc = tf.py_func(CTCUtils.compute_ctc_from_labels,[input_labels],[tf.int64,tf.int64,tf.int64])
          ctc_labels = tf.to_int32(tf.SparseTensor(ctc[0],ctc[1],ctc[2]))
          predictions = tf.to_int32(tf.nn.ctc_beam_search_decoder(prob,data_length,merge_repeated=False,beam_width=10)[0][0])
          tf.sparse_tensor_to_dense(predictions,default_value=-1,name='d_predictions')
          tf.reduce_mean(tf.edit_distance(predictions,ctc_labels,normalize=False),name='error_rate')
          loss = tf.reduce_mean(tf.nn.ctc_loss(inputs=prob,labels=ctc_labels,sequence_length=data_length,ctc_merge_repeated=True),name='loss')
          learning_rate = tf.train.piecewise_constant(global_step,[150000,200000],[config.train.learning_rate,0.1 * config.train.learning_rate,0.01 * config.train.learning_rate])
          opt_loss = tf.contrib.layers.optimize_loss(loss,global_step,learning_rate,config.train.opt_type,config.train.grad_noise_scale,name='train_step')

********************初始化并训练************************* *

     tf.global_variables_initializer()
     writer = None
     if config.train.need_to_save_log:
        writer = tf.summary.FileWriter(config.model_dir,session.graph)

     graph.finalize()


     for i in range(train.steps):
         **********
         **********

我对以下内容感到困惑。 （1）f_saver仅用于恢复预训练模型，或仍可用于保存整个图形，包括预训练和CTC识别部分。 （2）要初始化CTC识别部分，请使用tf.global_variables_initializer（）。 这是正确的使用方式吗？还是只需要初始化识别范围中的变量？ （3）