M1 MacBook运行Apple原生TensorFlow2.4的Numpy兼容性问题咨询

问题现象 #

• 运行环境：M1 MacBook，使用苹果原生TensorFlow 2.4包，miniforge搭建的Python 3.8虚拟环境
• 代码运行背景：同一份Jupyter Notebook代码原本可在Intel架构设备、旧版TensorFlow环境下正常运行
• 已做排查：先后尝试匹配NumPy 1.20、1.18版本，完成降级操作后均未解决兼容性问题
• 触发场景：向Keras Sequential模型添加LSTM层时直接中断报错

核心报错信息 #

未实现错误：无法将符号张量（lstm_1/strided_slice:0）转换为numpy数组。该错误通常表示你正在尝试将张量传递给NumPy调用，而这种操作是不受支持的
---------------------------------------------------------------------------
未实现错误                               Traceback (most recent call last)
<ipython-input-20-73358e637fe3> in <module>
      4 model = Sequential()
      5 model.add(Embedding(vocab_size+1, W2V_SIZE, weights=[embedding_matrix], input_length=MAX_SEQUENCE_LENGTH, trainable=False))
----> 6 model.add(LSTM(500, dropout=0.2, recurrent_dropout=0.2))
      7 model.add(Dense(units = 10000, kernel_initializer = 'glorot_uniform', activation = 'relu'))
      8 model.add(Dropout(0.35))

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/training/tracking/base.py in _method_wrapper(self, *args, **kwargs)
    515     self._self_setattr_tracking = False  # pylint: disable=protected-access
    516     try:
--> 517       result = method(self, *args, **kwargs)
    518     finally:
    519       self._self_setattr_tracking = previous_value  # pylint: disable=protected-access

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/keras/engine/sequential.py in add(self, layer)
    221       # 如果模型是在输入层上方连续构建的：
    222       # 刷新其输出
--> 223       output_tensor = layer(self.outputs[0])
    224       if len(nest.flatten(output_tensor)) != 1:
    225         raise ValueError(SINGLE_LAYER_OUTPUT_ERROR_MSG)

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/keras/layers/recurrent.py in __call__(self, inputs, initial_state, constants, **kwargs)
    658 
    659     if initial_state is None and constants is None:
--> 660       return super(RNN, self).__call__(inputs, **kwargs)
    661 
    662     # 如果指定了`initial_state`或`constants`中的任意参数，且这些参数是Keras

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/keras/engine/base_layer.py in __call__(self, *args, **kwargs)
    944     # >> model = tf.keras.Model(inputs, outputs)
    945     if _in_functional_construction_mode(self, inputs, args, kwargs, input_list):
--> 946       return self._functional_construction_call(inputs, args, kwargs,
    947                                                 input_list)
    948 

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/keras/engine/base_layer.py in _functional_construction_call(self, inputs, args, kwargs, input_list)
   1083           layer=self, inputs=inputs, build_graph=True, training=training_value):
   1084         # 检查层构建后设置的输入假设，例如输入形状
-> 1085         outputs = self._keras_tensor_symbolic_call(
   1086             inputs, input_masks, args, kwargs)
   1087 

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/keras/engine/base_layer.py in _keras_tensor_symbolic_call(self, inputs, input_masks, args, kwargs)
    815       return nest.map_structure(keras_tensor.KerasTensor, output_signature)
    816     else:
--> 817       return self._infer_output_signature(inputs, args, kwargs, input_masks)
    818 

    819   def _infer_output_signature(self, inputs, args, kwargs, input_list):

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/keras/engine/base_layer.py in _infer_output_signature(self, inputs, args, kwargs, input_masks)
    856           # TODO(kaftan): 是否在此处执行maybe_build，还是已经完成构建？
    857           self._maybe_build(inputs)
--> 858           outputs = call_fn(inputs, *args, **kwargs)
    859 
    860         self._handle_activity_regularization(inputs, outputs)

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/keras/layers/recurrent_v2.py in call(self, inputs, mask, training, initial_state)
   1161     # LSTM不支持常量，处理过程中忽略该参数
   1162     orig_initial_state = initial_state
-> 1163     inputs, initial_state, _ = self._process_inputs(inputs, initial_state, None)
   1164 

   1165     if isinstance(mask, list):

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/keras/layers/recurrent.py in _process_inputs(self, inputs, initial_state, constants)
    857         initial_state = self.states
    858     elif initial_state is None:
--> 859       initial_state = self.get_initial_state(inputs)
    860 

    861     if len(initial_state) != len(self.states):

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/keras/layers/recurrent.py in get_initial_state(self, inputs)
    640     dtype = inputs.dtype
    641     if get_initial_state_fn:
--> 642       init_state = get_initial_state_fn(
    643           inputs=None, batch_size=batch_size, dtype=dtype)
    644     else:

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/keras/layers/recurrent.py in get_initial_state(self, inputs, batch_size, dtype)
   2504 
   2505   def get_initial_state(self, inputs=None, batch_size=None, dtype=None):
-> 2506     return list(_generate_zero_filled_state_for_cell(
   2507         self, inputs, batch_size, dtype))
   2508 

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/keras/layers/recurrent.py in _generate_zero_filled_state_for_cell(cell, inputs, batch_size, dtype)
   2985     batch_size = array_ops.shape(inputs)[0]
   2986     dtype = inputs.dtype
-> 2987   return _generate_zero_filled_state(batch_size, cell.state_size, dtype)
   2988 

   2989 

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/keras/layers/recurrent.py in _generate_zero_filled_state(batch_size_tensor, state_size, dtype)
   3001 
   3002   if nest.is_nested(state_size):
-> 3003     return nest.map_structure(create_zeros, state_size)
   3004   else:
   3005     return create_zeros(state_size)

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/util/nest.py in map_structure(func, *structure, **kwargs)
    657 
    658   return pack_sequence_as(
--> 659       structure[0], [func(*x) for x in entries],
    660       expand_composites=expand_composites)
    661 

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/util/nest.py in <listcomp>(.0)
    657 
    658   return pack_sequence_as(
--> 659       structure[0], [func(*x) for x in entries],
    660       expand_composites=expand_composites)
    661 

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/keras/layers/recurrent.py in create_zeros(unnested_state_size)
   2998     flat_dims = tensor_shape.TensorShape(unnested_state_size).as_list()
   2999     init_state_size = [batch_size_tensor] + flat_dims
-> 3000     return array_ops.zeros(init_state_size, dtype=dtype)
   3001 

   3002   if nest.is_nested(state_size):

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/util/dispatch.py in wrapper(*args, **kwargs)
    199     """调用目标函数，若抛出TypeError则回退到分发器逻辑。"""
    200     try:
--> 201       return target(*args, **kwargs)
    202     except (TypeError, ValueError):
    203       # 注意：convert_to_eager_tensor当前抛出的是ValueError，而非

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/ops/array_ops.py in wrapped(*args, **kwargs)
   2818   def wrapped(*args, **kwargs):
-> 2819     tensor = fun(*args, **kwargs)
   2820     tensor._is_zeros_tensor = True
   2821     return tensor

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/ops/array_ops.py in zeros(shape, dtype, name)
   2866           # 如果张量尺寸不大则创建常量，否则创建填充
   2867           # 操作以避免序列化的GraphDef过大
-> 2868           output = _constant_if_small(zero, shape, dtype, name)
   2869           if output is not None:
-> 2870             return output

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/ops/array_ops.py in _constant_if_small(value, shape, dtype, name)
   2802 def _constant_if_small(value, shape, dtype, name):
   2803   try:
-> 2804     if np.prod(shape) < 1000:
   2805       return constant(value, shape=shape, dtype=dtype, name=name)
   2806   except TypeError:
<__array_function__ internals> in prod(*args, **kwargs)

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/numpy/core/fromnumeric.py in prod(a, axis, dtype, out, keepdims, initial, where)
   3028     10
   3029     """
-> 3030     return _wrapreduction(a, np.multiply, 'prod', axis, dtype, out,
   3031                           keepdims=keepdims, initial=initial, where=where)
   3032 

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/numpy/core/fromnumeric.py in _wrapreduction(obj, ufunc, method, axis, dtype, out, **kwargs)
         85                 return reduction(axis=axis, out=out, **passkwargs)
         86 
---> 87     return ufunc.reduce(obj, axis, dtype, out, **passkwargs)
         88 

         89 

~/miniforge3/envs/tf2.4/lib/python3.8/site-packages/tensorflow/python/framework/ops.py in __array__(self)
        850 
        851   def __array__(self):
--> 852     raise NotImplementedError(
        853         "无法将符号张量（{}）转换为numpy数组。"
        854         "该错误通常表示你正在尝试将张量传递给"

问题根因 #

• 直接触发点位于TensorFlow 2.4内部的_constant_if_small函数，该函数判断零张量是否为小尺寸时，直接调用np.prod(shape)计算形状元素总数。当LSTM层初始化时传入的batch size是动态符号张量，NumPy 1.20及以上版本会拒绝将符号张量传入计算接口，触发类型转换错误。
• 降级到NumPy 1.18仍报错，大概率是两种情况：一是降级后未重启Jupyter内核，NumPy版本未在当前会话生效；二是M1原生TensorFlow 2.4的早期构建包本身存在接口适配遗漏，未完美兼容低版本NumPy的输入要求。

解决方法 #

• 先确认当前会话的NumPy版本：在Jupyter单元格执行import numpy; numpy.__version__，确认输出为1.19.2（TensorFlow 2.4官方验证兼容的版本），若不对则执行pip install numpy==1.19.2完成安装，安装后重启Jupyter内核再运行代码测试。
• 若降级后仍触发报错，可临时修改TF源码绕过冲突：找到虚拟环境路径下的tensorflow/python/ops/array_ops.py文件，定位第2804行的if np.prod(shape) < 1000:代码，将np.prod(shape)替换为tensor_shape.num_elements(shape)，保存文件后重启内核即可正常添加LSTM层。
• 长期优化建议：M1芯片适配的TensorFlow 2.4属于早期测试版本，对Keras循环层、自定义层的兼容存在较多遗留问题，可升级至适配M1的TensorFlow 2.10及以上版本，可覆盖绝大多数已知兼容问题。

内容的提问来源于stack exchange，提问作者Tom Xu