TF1迁移TF2触发str与VariableScope拼接TypeError求解

问题背景 #

• 项目原本基于TensorFlow V1开发，迁移到TensorFlow V2环境时已修复绝大多数代码兼容问题，运行时抛出上述类型错误，未检索到可行解决方案。

相关代码片段 #

wdmSystem.py 入口调用代码 #

sess = tf.compat.v1.Session()
sess, outMetrics = cft.train(sess, tf.compat.v1.train.AdamOptimizer, loss, metricsDict, trainingParam, feedDictFun, debug=True)

training.py 核心 train 函数 #

def train(sess, optimizer, loss, metricsDict, trainingParam, feedDictFun, debug=False):

    optimizer = optimizer(learning_rate=trainingParam.learningRate)
    meanMetricOpsDict, updateOps, resetOps = create_mean_metrics(metricsDict)
    trainStep, accumulateOps, zeroOps = accumulatedOptimizer(optimizer, loss, tf.compat.v1.trainable_variables(), trainingParam.nMiniBatches)

    init = tf.compat.v1.global_variables_initializer()

    if trainingParam.summaries:        
        s = [tf.compat.v1.summary.scalar(name, metric) for name,metric in meanMetricOpsDict.items()]
        metricSummaries = tf.compat.v1.summary.merge(s)

        summaries_dir = os.path.join(trainingParam.path, 'tboard', trainingParam.summaryString)
        os.makedirs(summaries_dir, exist_ok=True)

        trainWriter = tf.compat.v1.summary.FileWriter(summaries_dir + '/train', sess.graph)
    else:
        trainWriter = None
    
    sess.run(init)

    saver = tf.compat.v1.train.Saver()
    checkpoint_path = os.path.join(trainingParam.path,'checkpoint',trainingParam.filename,'best')
    if not os.path.exists(checkpoint_path):
        os.makedirs(checkpoint_path)
    else:
        print("Restoring...", flush=True)
        saver.restore(sess=sess,save_path=checkpoint_path)

    bestLoss = 100000
    bestAcc = 0
    lastImprovement = 0

    sess.run([resetOps, zeroOps])

    for epoch in range(1, trainingParam.nEpochs+1):
        sess.run(resetOps)
        
        for batch in range(1,trainingParam.nBatches+1):
            sess.run(zeroOps)
            if debug:
                print('batch: ', batch, end=' | ', flush=True)
                print('miniBatch: ', end='', flush=True)
            for miniBatch in range(1,trainingParam.nMiniBatches+1):
                if debug:
                    print(miniBatch, end=', ')
                feedDict = feedDictFun(trainingParam)
                sess.run([accumulateOps, updateOps], feed_dict=feedDict)
            if debug:
                print('', flush=True)
            sess.run(trainStep)
                
        outMetrics = sess.run(list(meanMetricOpsDict.values()), feed_dict=feedDict)
        outMetrics = { key:val for key,val in zip(list(meanMetricOpsDict.keys()), outMetrics) }
        
        if trainingParam.summaries:
            outMetricSummaries = sess.run(metricSummaries, feed_dict=feedDict)
            trainWriter.add_summary(outMetricSummaries, epoch)
        
        earlyStoppingMetric = outMetrics[trainingParam.earlyStoppingMetric]
        if earlyStoppingMetric < bestLoss:
            bestLoss = earlyStoppingMetric
            lastImprovement = epoch
            saver.save(sess=sess, save_path=checkpoint_path)
            
        if epoch%trainingParam.displayStep == 0:
            outString = 'epoch: {:04d}'.format(epoch)
            for key, value in outMetrics.items():
                outString += ' - {}: {:.4f}'.format(key, value)
            print(outString, flush=True)

    saver.restore(sess=sess,save_path=checkpoint_path)

    sess.run(resetOps)
    for _ in range(trainingParam.evalBatches):
        feedDict = feedDictFun(trainingParam)
        sess.run(updateOps, feed_dict=feedDict)
    
    outMetrics = sess.run(list(meanMetricOpsDict.values()), feed_dict=feedDict)
    outMetrics = { key:val for key,val in zip(list(meanMetricOpsDict.keys()), outMetrics) }
    
    return sess, outMetrics

training.py create_mean_metrics 函数 #

def create_mean_metrics(metricsDict):
    meanMetricOpsDict = {}
    updateOps = []
    resetOps = []
    for name, tensor in metricsDict.items():
        lossOp, updateOp, resetOp = create_reset_metric(tf.compat.v1.metrics.mean, name, tensor)
        meanMetricOpsDict[name] = lossOp
        updateOps.append(updateOp)
        resetOps.append(resetOp)    
    return meanMetricOpsDict, updateOps, resetOps

training.py create_reset_metric 函数（报错位置） #

def create_reset_metric(metric, scope='reset_metrics', *metric_args, **metric_kwargs):
    """
        指标重置逻辑参考官方issue实现
    """
    with tf.compat.v1.variable_scope(scope) as scope:
        metric_op, update_op = metric(*metric_args, **metric_kwargs)
        vars = tf.compat.v1.get_variable(scope, collections=tf.compat.v1.GraphKeys.LOCAL_VARIABLES)
        reset_op = tf.compat.v1.variables_initializer(vars)
    return metric_op, update_op, reset_op

完整报错追踪栈 #

Traceback (most recent call last):
  File "/home/osama/PycharmProjects/claude-master/tf_wdmSystem-learning.py", line 222, in <module>
    sess, outMetrics = cft.train(sess, tf.compat.v1.train.AdamOptimizer, loss, metricsDict, trainingParam, feedDictFun, debug=True)
  File "/home/osama/PycharmProjects/claude-master/claude/claudeflow/training.py", line 38, in train
    meanMetricOpsDict, updateOps, resetOps = create_mean_metrics(metricsDict)
  File "/home/osama/PycharmProjects/claude-master/claude/claudeflow/training.py", line 20, in create_mean_metrics
    lossOp, updateOp, resetOp = create_reset_metric(tf.compat.v1.metrics.mean, name, tensor)
  File "/home/osama/PycharmProjects/claude-master/claude/claudeflow/training.py", line 11, in create_reset_metric
    vars = tf.compat.v1.get_variable(scope, collections=tf.compat.v1.GraphKeys.LOCAL_VARIABLES)
  File "/home/osama/PycharmProjects/claude-master/venv/lib/python3.10/site-packages/tensorflow/python/ops/variable_scope.py", line 1616, in get_variable
    return get_variable_scope().get_variable(
  File "/home/osama/PycharmProjects/claude-master/venv/lib/python3.10/site-packages/tensorflow/python/ops/variable_scope.py", line 1308, in get_variable
    full_name = self.name + "/" + name if self.name else name
TypeError: can only concatenate str (not "VariableScope") to str

Process finished with exit code 1

问题根因与修复方案 #

• 根因：with tf.compat.v1.variable_scope(scope) as scope 语句会将原本字符串类型的入参scope覆盖为VariableScope上下文对象，后续调用tf.compat.v1.get_variable时第一个参数要求传入字符串类型的变量名，实际传入了VariableScope实例，触发字符串拼接类型错误。同时原代码获取指标对应变量的逻辑也存在错误，tf.compat.v1.get_variable是用于创建/获取单个指定名称变量的接口，不能直接传入作用域对象获取作用域下的所有变量。
• 修复方法：
  1. 修改变量作用域上下文的赋值变量名，避免覆盖原字符串类型的scope入参
  2. 改用tf.compat.v1.get_collection接口获取当前指标作用域下的所有本地变量，用于生成重置操作

修复后的create_reset_metric代码如下：

def create_reset_metric(metric, scope='reset_metrics', *metric_args, **metric_kwargs):
    # 上下文返回值使用独立变量名，不覆盖原scope字符串参数
    with tf.compat.v1.variable_scope(scope) as var_scope:
        metric_op, update_op = metric(*metric_args, **metric_kwargs)
        # 拉取当前作用域下所有属于本地变量集合的指标变量
        metric_vars = tf.compat.v1.get_collection(
            tf.compat.v1.GraphKeys.LOCAL_VARIABLES,
            scope=var_scope.name
        )
        reset_op = tf.compat.v1.variables_initializer(metric_vars)
    return metric_op, update_op, reset_op

内容的提问来源于stack exchange，提问作者Osama Ismail