Keras符号张量转NumPy数组报错：无法转换符号tf.Tensor

问题背景 #

使用Python 3.12.3、NumPy 1.26.4、无GPU支持的TensorFlow 2.17.0在私有环境开发，需实现调用C库的Keras自定义损失函数，计划将损失函数参数转为NumPy数组后通过PyBind11传递给C。但在损失函数中执行numpy.asarray(yPred)时触发报错：

NotImplementedError: Cannot convert a symbolic tf.Tensor (data_1:0) to a numpy array. This error may indicate that you're trying to pass a Tensor to a NumPy call, which is not supported.

启用tensorflow.config.run_functions_eagerly(True)可解决报错，但运行时间从11秒骤增至145秒，性能严重下降，寻求高效解决方案。

示例代码 #

from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import mean_squared_error
from keras.layers import Input
from keras.layers import Dense
from keras import Model
import numpy
import tensorflow
from tensorflow.python.ops import math_ops

def pinnLoss(yTrue, yPred):
    yArr = numpy.asarray(yPred) # Will be used later.
    squared_difference = math_ops.square(yTrue - yPred)
    return math_ops.mean(squared_difference, axis=-1)  # Note the `axis=-1`

x = numpy.asarray([i for i in range(-50,51)])
y = numpy.asarray([i * i for i in x])
x = x.reshape((len(x), 1))
y = y.reshape((len(y), 1))
scale_x = MinMaxScaler()
x = scale_x.fit_transform(x)
scale_y = MinMaxScaler()
y = scale_y.fit_transform(y)
inputShape = (1,)
inputs = Input(shape=inputShape)
tmp = inputs
tmp = Dense(10, activation='relu', kernel_initializer='he_uniform')(tmp)
tmp = Dense(10, activation='relu', kernel_initializer='he_uniform')(tmp)
tmp = Dense(1)(tmp)
model = Model(inputs, tmp)
model.compile(loss=pinnLoss, optimizer='adam')
model.fit(x, y, epochs=500, batch_size=10, verbose=0)

完整报错信息 #

2024-10-14 11:37:04.117027: I external/local_xla/xla/tsl/cuda/cudart_stub.cc:32] Could not find cuda drivers on your machine, GPU will not be used.
2024-10-14 11:37:04.119994: I external/local_xla/xla/tsl/cuda/cudart_stub.cc:32] Could not find cuda drivers on your machine, GPU will not be used.
2024-10-14 11:37:04.129464: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:485] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
2024-10-14 11:37:04.144761: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:8454] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
2024-10-14 11:37:04.149407: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1452] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
2024-10-14 11:37:04.975719: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
Traceback (most recent call last):
  File "/home/bamer/work/dev/optimizer/src/minimalworking2.py", line 31, in <module>
    model.fit(x, y, epochs=500, batch_size=10, verbose=0)
  File "/home/bamer/ownpython/lib/python3.12/site-packages/keras/src/utils/traceback_utils.py", line 122, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "/home/bamer/work/dev/optimizer/src/minimalworking2.py", line 11, in pinnLoss
    yArr = numpy.asarray(yPred) # Will be used later.
           ^^^^^^^^^^^^^^^^^^^^
NotImplementedError: Cannot convert a symbolic tf.Tensor (data_1:0) to a numpy array. This error may indicate that you're trying to pass a Tensor to a NumPy call, which is not supported.

高效解决方案 #

方法1：用tf.py_function封装C++调用逻辑 #

该方法在TensorFlow图模式下执行，同时支持调用Python函数（进而调用C++库），避免全Eager模式的性能损耗。

实现代码

from sklearn.preprocessing import MinMaxScaler
from keras.layers import Input
from keras.layers import Dense
from keras import Model
import numpy
import tensorflow as tf
from tensorflow.python.ops import math_ops

# 替换为你的PyBind11调用C++库的逻辑
def cpp_loss_calculation(y_true_np, y_pred_np):
    # 这里写实际调用C++库的代码
    squared_diff = (y_true_np - y_pred_np) ** 2
    return numpy.mean(squared_diff, axis=-1)

def pinnLoss(yTrue, yPred):
    # 用tf.py_function包装Python函数，指定输入输出类型
    loss = tf.py_function(
        func=cpp_loss_calculation,
        inp=[yTrue, yPred],
        Tout=tf.float32
    )
    # 保持张量形状，避免后续模型训练报错
    loss.set_shape(yTrue.get_shape())
    return loss

# 后续模型构建与训练代码不变
x = numpy.asarray([i for i in range(-50,51)])
y = numpy.asarray([i * i for i in x])
x = x.reshape((len(x), 1))
y = y.reshape((len(y), 1))
scale_x = MinMaxScaler()
x = scale_x.fit_transform(x)
scale_y = MinMaxScaler()
y = scale_y.fit_transform(y)
inputShape = (1,)
inputs = Input(shape=inputShape)
tmp = inputs
tmp = Dense(10, activation='relu', kernel_initializer='he_uniform')(tmp)
tmp = Dense(10, activation='relu', kernel_initializer='he_uniform')(tmp)
tmp = Dense(1)(tmp)
model = Model(inputs, tmp)
model.compile(loss=pinnLoss, optimizer='adam')
model.fit(x, y, epochs=500, batch_size=10, verbose=0)

方法2：实现TensorFlow自定义Op（极致性能） #

若追求最高性能，可将C++逻辑封装为TensorFlow自定义Op，完全融入TensorFlow图计算，性能与原生Op一致。

步骤

1. 用C++编写符合TensorFlow Op规范的代码，包含Op注册逻辑
2. 将代码编译为动态链接库（.so文件）
3. 在Python中加载该库，直接在损失函数中调用自定义Op

此方法门槛较高，但适合计算密集型的损失函数场景。

方法3：结合TensorFlow数据管道批量处理 #

若C++库支持批量处理，可使用tf.data管道提前将数据转为NumPy数组处理后喂给模型，适合离线预处理场景，但实时损失计算仍推荐前两种方法。

内容的提问来源于stack exchange，提问作者Balázs Bämer