函数调用堆栈：keras_scratch_graph 错误

Question

我正在重新实现一个 text2speech 项目。我正面临一个函数调用堆栈：解码器部分的keras_scratch_graph错误。网络架构来自 Deep Voice 3 论文。

我在 Google Colab 上使用来自 TF 2.0 的 keras。下面是解码器 Keras 模型的代码。

y1 = tf.ones(shape = (16, 203, 320))
def Decoder(name = "decoder"):
    # Decoder Prenet
    din = tf.concat((tf.zeros_like(y1[:, :1, -hp.mel:]), y1[:, :-1, -hp.mel:]), 1)
    keys = K.Input(shape = (180, 256), batch_size = 16, name = "keys")
    vals = K.Input(shape = (180, 256), batch_size = 16, name = "vals")
    prev_max_attentions_li = tf.ones(shape=(hp.dlayer, hp.batch_size), dtype=tf.int32)
    #prev_max_attentions_li = K.Input(tensor = prev_max_attentions_li)
    for i in range(hp.dlayer):
        dpout = K.layers.Dropout(rate = 0 if i == 0 else hp.dropout)(din)
        fc_out = K.layers.Dense(hp.char_embed, activation = 'relu')(dpout)

    print("=======================================================================================================")
    print("The FC value is ", fc_out)
    print("=======================================================================================================")

    query_pe = K.layers.Embedding(hp.Ty, hp.char_embed)(tf.tile(tf.expand_dims(tf.range(hp.Ty // hp.r), 0), [hp.batch_size, 1]))
    key_pe = K.layers.Embedding(hp.Tx, hp.char_embed)(tf.tile(tf.expand_dims(tf.range(hp.Tx), 0), [hp.batch_size, 1]))

    alignments_li, max_attentions_li = [], []
    for i in range(hp.dlayer):
        dpout = K.layers.Dropout(rate = 0)(fc_out)
        queries = K.layers.Conv1D(hp.datten_size, hp.dfilter, padding = 'causal', dilation_rate = 2**i)(dpout)
        fc_out = (queries + fc_out) * tf.math.sqrt(0.5)
        print("=======================================================================================================")
        print("The FC value is ", fc_out)
        print("=======================================================================================================")
        queries = fc_out + query_pe
        keys += key_pe

        tensor, alignments, max_attentions = Attention(name = "attention")(queries, keys, vals, prev_max_attentions_li[i])

        fc_out = (tensor + queries) * tf.math.sqrt(0.5)

        alignments_li.append(alignments)
        max_attentions_li.append(max_attentions)

    decoder_output = fc_out

    dpout = K.layers.Dropout(rate = 0)(decoder_output)
    mel_logits = K.layers.Dense(hp.mel * hp.r)(dpout)

    dpout = K.layers.Dropout(rate = 0)(fc_out)
    done_output = K.layers.Dense(2)(dpout)

    return K.Model(inputs = [keys, vals], outputs = [mel_logits, done_output, decoder_output, alignments_li, max_attentions_li], name = name)

decode = Decoder()
kin = tf.ones(shape = (16, 180, 256))
vin = tf.ones(shape = (16, 180, 256))
print(decode(kin, vin))
tf.keras.utils.plot_model(decode, to_file = "decoder.png", show_shapes = True)

当我使用一些数据进行测试时，它会显示以下错误消息。“fc_out”会出现一些问题，但我不知道如何将“fc_out”输出从第一个 for 循环传递到第二个 for 循环？任何答案将不胜感激。

File "Decoder.py", line 60, in <module>
    decode = Decoder()
  File "Decoder.py", line 33, in Decoder
    dpout = K.layers.Dropout(rate = 0)(fc_out)
  File "/Users/ydc/dl-npm/lib/python3.7/site-packages/tensorflow/python/keras/engine/base_layer.py", line 596, in __call__
    base_layer_utils.create_keras_history(inputs)
  File "/Users/ydc/dl-npm/lib/python3.7/site-packages/tensorflow/python/keras/engine/base_layer_utils.py", line 199, in create_keras_history
    _, created_layers = _create_keras_history_helper(tensors, set(), [])
  File "/Users/ydc/dl-npm/lib/python3.7/site-packages/tensorflow/python/keras/engine/base_layer_utils.py", line 245, in _create_keras_history_helper
    layer_inputs, processed_ops, created_layers)
  File "/Users/ydc/dl-npm/lib/python3.7/site-packages/tensorflow/python/keras/engine/base_layer_utils.py", line 245, in _create_keras_history_helper
    layer_inputs, processed_ops, created_layers)
  File "/Users/ydc/dl-npm/lib/python3.7/site-packages/tensorflow/python/keras/engine/base_layer_utils.py", line 245, in _create_keras_history_helper
    layer_inputs, processed_ops, created_layers)
  File "/Users/ydc/dl-npm/lib/python3.7/site-packages/tensorflow/python/keras/engine/base_layer_utils.py", line 243, in _create_keras_history_helper
    constants[i] = backend.function([], op_input)([])
  File "/Users/ydc/dl-npm/lib/python3.7/site-packages/tensorflow/python/keras/backend.py", line 3510, in __call__
    outputs = self._graph_fn(*converted_inputs)
  File "/Users/ydc/dl-npm/lib/python3.7/site-packages/tensorflow/python/eager/function.py", line 572, in __call__
    return self._call_flat(args)
  File "/Users/ydc/dl-npm/lib/python3.7/site-packages/tensorflow/python/eager/function.py", line 671, in _call_flat
    outputs = self._inference_function.call(ctx, args)
  File "/Users/ydc/dl-npm/lib/python3.7/site-packages/tensorflow/python/eager/function.py", line 445, in call
    ctx=ctx)
  File "/Users/ydc/dl-npm/lib/python3.7/site-packages/tensorflow/python/eager/execute.py", line 67, in quick_execute
    six.raise_from(core._status_to_exception(e.code, message), None)
  File "<string>", line 3, in raise_from
tensorflow.python.framework.errors_impl.FailedPreconditionError:  Error while reading resource variable _AnonymousVar19 from Container: localhost. This could mean that the variable was uninitialized. Not found: Resource localhost/_AnonymousVar19/N10tensorflow3VarE does not exist.
     [[node dense_7/BiasAdd/ReadVariableOp (defined at Decoder.py:33) ]] [Op:__inference_keras_scratch_graph_566]

Function call stack:
keras_scratch_graph

Answer 1

我的情况是 tensorflow 示例代码在 Google colab 中工作正常，但在我的机器中却没有，因为我收到了 keras_scratch_graph 错误。

然后我在开头添加了这个 Python 代码，它工作正常。

gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
    try:
        # Restrict TensorFlow to only use the fourth GPU
        tf.config.experimental.set_visible_devices(gpus[0], 'GPU')

        # Currently, memory growth needs to be the same across GPUs
        for gpu in gpus:
            tf.config.experimental.set_memory_growth(gpu, True)
        logical_gpus = tf.config.experimental.list_logical_devices('GPU')
        print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs")
    except RuntimeError as e:
        # Memory growth must be set before GPUs have been initialized
        print(e)

默认情况下，TensorFlow 映射CUDA_VISIBLE_DEVICES进程可见的所有 GPU（受制于）的几乎所有 GPU 内存。

在某些情况下，进程只需要分配可用内存的一个子集，或者只根据进程需要增加内存使用量是可取的。

例如，您想用一个 GPU 同时训练多个小模型。通过调用tf.config.experimental.set_memory_growth，它尝试仅分配运行时分配所需的 GPU 内存：它开始分配很少的内存，随着程序运行并且需要更多 GPU 内存，我们扩展了分配给 TensorFlow 的 GPU 内存区域过程。

希望能帮助到你！

Answer 2

我遇到了类似的错误。我减小了批量大小，错误消失了。我不知道为什么，但它对我有用。我猜测与过度堆叠有关。