默认 MaxPoolingOp 仅在使用从 GPU 移植的训练模型时在设备类型 CPU 错误上支持 NHWC

我使用 python、keras 和 tensorflow 在 PC 上开发并训练了一个模型，并使用 GPU 运行预测等。一切正常。

然后，我将模型和预测代码转移到笔记本电脑上，并使用requirements.txt重建环境，将gpu包交换为cpu包。

当我尝试运行预测代码时，我收到一个我无法理解的错误。

我的印象是张量流会透明地使用/不使用 GPU，所以我想知道它还能是什么。

    Traceback (most recent call last):
  File ".\metatrader.py", line 231, in <module>
    result = predict(ret[0] + filename)
  File ".\metatrader.py", line 104, in predict
    array = model.predict(x)
  File "C:\Users\antho\.conda\envs\tensorflow\lib\site-packages\tensorflow\python\keras\engine\training.py", line 130, in _method_wrapper
    return method(self, *args, **kwargs)
  File "C:\Users\antho\.conda\envs\tensorflow\lib\site-packages\tensorflow\python\keras\engine\training.py", line 1599, in predict
    tmp_batch_outputs = predict_function(iterator)
  File "C:\Users\antho\.conda\envs\tensorflow\lib\site-packages\tensorflow\python\eager\def_function.py", line 780, in __call__
    result = self._call(*args, **kwds)
  File "C:\Users\antho\.conda\envs\tensorflow\lib\site-packages\tensorflow\python\eager\def_function.py", line 846, in _call
    return self._concrete_stateful_fn._filtered_call(canon_args, canon_kwds)  # pylint: disable=protected-access
  File "C:\Users\antho\.conda\envs\tensorflow\lib\site-packages\tensorflow\python\eager\function.py", line 1843, in _filtered_call
    return self._call_flat(
  File "C:\Users\antho\.conda\envs\tensorflow\lib\site-packages\tensorflow\python\eager\function.py", line 1923, in _call_flat
    return self._build_call_outputs(self._inference_function.call(
  File "C:\Users\antho\.conda\envs\tensorflow\lib\site-packages\tensorflow\python\eager\function.py", line 545, in call
    outputs = execute.execute(
  File "C:\Users\antho\.conda\envs\tensorflow\lib\site-packages\tensorflow\python\eager\execute.py", line 59, in quick_execute
    tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name,
tensorflow.python.framework.errors_impl.InvalidArgumentError:  Default MaxPoolingOp only supports NHWC on device type CPU
         [[node sequential/max_pooling2d_1/MaxPool (defined at .\metatrader.py:104) ]] [Op:__inference_predict_function_445]

Function call stack:
predict_function

任何帮助，将不胜感激！！

Updated:

我使用的版本：tensorflow==2.3.1

经过进一步的摆弄，我发现训练也不能在仅 CPU 的系统上运行，但在 GPU 系统上运行良好。我预计这是与以前版本的一些不兼容性，我还没有完全理解。

import os
import sys
from tensorflow.keras import optimizers
from tensorflow.keras.layers import Dropout, Flatten, Dense, Activation, BatchNormalization
import tensorflow.keras.layers as lyrs
from tensorflow.keras.models import Sequential
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.preprocessing import image
import numpy as np
import tensorflow as tf

epochs = 100

train_data_dir = './data/train/'
validation_data_dir = './data/validate/'

nb_train_samples = sum([len(files) for r, d, files in os.walk(train_data_dir)])
nb_validation_samples = sum([len(files) for r, d, files in os.walk(validation_data_dir)])

nb_filters1 = 32
nb_filters2 = 32
nb_filters3 = 64
conv1_size = 3
conv2_size = 2
conv3_size = 5
pool_size = 2
classes_num = 2
batch_size = 128
chanDim =3

model = Sequential()
model.add(lyrs.Conv2D(nb_filters1, (conv1_size, conv1_size), input_shape=(150, 150, 3), padding='same'))
model.add(Activation('relu'))
model.add(lyrs.MaxPooling2D(pool_size=(pool_size, pool_size)))
model.add(lyrs.Conv2D(nb_filters2, (conv2_size, conv2_size), padding="same"))
model.add(Activation('relu'))
model.add(lyrs.MaxPooling2D(pool_size=(pool_size, pool_size), data_format="channels_first"))
model.add(lyrs.Conv2D(nb_filters3, (conv3_size, conv3_size), padding='same'))
model.add(Activation('relu'))
model.add(lyrs.MaxPooling2D(pool_size=(pool_size, pool_size), data_format="channels_first"))
model.add(Flatten())
model.add(Dense(1024))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(classes_num, activation='softmax'))
model.summary()

model.compile(loss='categorical_crossentropy',
                      optimizer=optimizers.RMSprop(),
                      metrics=['accuracy'])

train_datagen = ImageDataGenerator(
    horizontal_flip=False)

test_datagen = ImageDataGenerator(
    horizontal_flip=False)

train_generator = train_datagen.flow_from_directory(
    train_data_dir,
    target_size=(150, 150),
    batch_size=batch_size,
    class_mode='categorical'
)

validation_generator = test_datagen.flow_from_directory(
    validation_data_dir,
    target_size=(150, 150),
    batch_size=batch_size,
    class_mode='categorical')
    
model.fit_generator(
    train_generator,
    steps_per_epoch=nb_train_samples//batch_size,
    epochs=epochs,
    shuffle=True,
    validation_data=validation_generator,
    validation_steps=nb_validation_samples//batch_size)

model.save('./my_model.hdf5', overwrite=True)

我得到的错误：

tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name,
tensorflow.python.framework.errors_impl.InvalidArgumentError:  Default MaxPoolingOp only supports NHWC on device type CPU
         [[node sequential/max_pooling2d_1/MaxPool (defined at .\sample-training.py:72) ]] [Op:__inference_train_function_1065]

好吧，我最终找到了答案。

在tensorfol0w 2.3.1中，与模型在GPU和仅CPU系统上的训练方式相关，某些地方似乎存在一些不一致。

当在纯 CPU 系统上尝试预测时，由 GPU 系统创建和保存的模型将不起作用。

虽然上述代码在具有相同版本张量流的 GPU 系统上运行良好，但在 CPU 系统上会失败并出现上述错误，并且在用于仅 CPU 系统上预测时由 GPU PC 创建/保存的任何模型也会失败。

更改以下行可以解决问题

model.add(lyrs.MaxPooling2D(pool_size=(pool_size, pool_size), data_format="channels_first"))

to:

model.add(lyrs.MaxPooling2D(pool_size=(pool_size, pool_size), data_format="channels_last"))

我还不清楚这到底是做什么的，但它也几乎减少了可训练元素数量的 1/2。

由于我是这个领域的新手，任何有启发性的话请随时发表评论或纠正我，如果/哪里我错了。