我惊讶地发现，当针对同一个深度学习问题（即著名的 MNIST 手写数字识别）使用不同的后端时，在速度和模型准确性方面存在显着的性能差异。下面的代码具有完全不同的输出，具体取决于我使用的后端。由于我的 MacBook 配备了 AMD Radeon Pro 560x GPU，因此我使用 PlaidML 后端进行基于 GPU 的训练。然后我将后端切换回 Tensorflow CPU，速度和准确率都大幅下降。
#!/usr/bin/env python
import os
#os.environ["KERAS_BACKEND"] = "tensorflow"
#import keras
# import tensorflow
# from tensorflow import keras
# from tensorflow.keras.datasets import mnist
# from tensorflow.keras.models import Sequential
# from tensorflow.keras.layers import Dense, Dropout, Flatten
# from tensorflow.keras.layers import Conv2D, MaxPooling2D
# from tensorflow.keras import backend as K
os.environ["KERAS_BACKEND"] = "plaidml.keras.backend"
import keras
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten, Conv2D, MaxPooling2D
from keras import backend as K
batch_size = 128
num_classes = 10
epochs = 12
# input image dimensions
img_rows, img_cols = 28, 28
# the data, split between train and test sets
(x_train, y_train), (x_test, y_test) = mnist.load_data()
if K.image_data_format() == 'channels_first':
x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
input_shape = (1, img_rows, img_cols)
else:
x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
input_shape = (img_rows, img_cols, 1)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
# convert class vectors to binary class matrices
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3),
activation='relu',
input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=keras.optimizers.Adadelta(),
metrics=['accuracy'])
model.fit(x_train, y_train,
batch_size=batch_size,
epochs=epochs,
verbose=1,
validation_data=(x_test, y_test))
score = model.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])
PlaidML 后端的结果如下：
x_train shape: (60000, 28, 28, 1)
60000 train samples
10000 test samples
INFO:plaidml:Opening device "metal_amd_radeon_pro_560x.0"
Train on 60000 samples, validate on 10000 samples
Epoch 1/12
60000/60000 [==============================] - 21s 356us/step - loss: 0.2686 - acc: 0.9160 - val_loss: 0.0551 - val_acc: 0.9826
Epoch 2/12
60000/60000 [==============================] - 17s 290us/step - loss: 0.0900 - acc: 0.9732 - val_loss: 0.0538 - val_acc: 0.9828
Epoch 3/12
60000/60000 [==============================] - 18s 295us/step - loss: 0.0678 - acc: 0.9796 - val_loss: 0.0357 - val_acc: 0.9887
Epoch 4/12
60000/60000 [==============================] - 17s 288us/step - loss: 0.0554 - acc: 0.9830 - val_loss: 0.0462 - val_acc: 0.9853
Epoch 5/12
60000/60000 [==============================] - 18s 294us/step - loss: 0.0466 - acc: 0.9854 - val_loss: 0.0312 - val_acc: 0.9899
Epoch 6/12
60000/60000 [==============================] - 18s 296us/step - loss: 0.0415 - acc: 0.9877 - val_loss: 0.0299 - val_acc: 0.9893
Epoch 7/12
60000/60000 [==============================] - 17s 285us/step - loss: 0.0362 - acc: 0.9889 - val_loss: 0.0310 - val_acc: 0.9904
Epoch 8/12
60000/60000 [==============================] - 17s 290us/step - loss: 0.0337 - acc: 0.9900 - val_loss: 0.0254 - val_acc: 0.9920
Epoch 9/12
60000/60000 [==============================] - 17s 287us/step - loss: 0.0314 - acc: 0.9905 - val_loss: 0.0284 - val_acc: 0.9911
Epoch 10/12
60000/60000 [==============================] - 38s 635us/step - loss: 0.0288 - acc: 0.9911 - val_loss: 0.0282 - val_acc: 0.9909
Epoch 11/12
60000/60000 [==============================] - 28s 466us/step - loss: 0.0270 - acc: 0.9918 - val_loss: 0.0267 - val_acc: 0.9915
Epoch 12/12
60000/60000 [==============================] - 17s 291us/step - loss: 0.0234 - acc: 0.9929 - val_loss: 0.0249 - val_acc: 0.9919
Test loss: 0.024859706979990005
Test accuracy: 0.9919
Tensorflow 后端的结果如下：
x_train shape: (60000, 28, 28, 1)
60000 train samples
10000 test samples
2020-06-26 11:19:40.480676: I tensorflow/core/platform/cpu_feature_guard.cc:143] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2020-06-26 11:19:40.522835: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x7fee96fa2510 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2020-06-26 11:19:40.522857: I tensorflow/compiler/xla/service/service.cc:176] StreamExecutor device (0): Host, Default Version
Epoch 1/12
469/469 [==============================] - 52s 110ms/step - loss: 2.2788 - accuracy: 0.1488 - val_loss: 2.2448 - val_accuracy: 0.3032
Epoch 2/12
469/469 [==============================] - 53s 112ms/step - loss: 2.2231 - accuracy: 0.2634 - val_loss: 2.1760 - val_accuracy: 0.4383
Epoch 3/12
469/469 [==============================] - 62s 133ms/step - loss: 2.1507 - accuracy: 0.3570 - val_loss: 2.0862 - val_accuracy: 0.5263
Epoch 4/12
469/469 [==============================] - 69s 147ms/step - loss: 2.0570 - accuracy: 0.4280 - val_loss: 1.9673 - val_accuracy: 0.5936
Epoch 5/12
469/469 [==============================] - 62s 133ms/step - loss: 1.9348 - accuracy: 0.4843 - val_loss: 1.8131 - val_accuracy: 0.6555
Epoch 6/12
469/469 [==============================] - 56s 120ms/step - loss: 1.7855 - accuracy: 0.5313 - val_loss: 1.6273 - val_accuracy: 0.7145
Epoch 7/12
469/469 [==============================] - 58s 125ms/step - loss: 1.6176 - accuracy: 0.5739 - val_loss: 1.4250 - val_accuracy: 0.7579
Epoch 8/12
469/469 [==============================] - 61s 131ms/step - loss: 1.4518 - accuracy: 0.6086 - val_loss: 1.2300 - val_accuracy: 0.7858
Epoch 9/12
469/469 [==============================] - 63s 134ms/step - loss: 1.3029 - accuracy: 0.6394 - val_loss: 1.0629 - val_accuracy: 0.8057
Epoch 10/12
469/469 [==============================] - 59s 125ms/step - loss: 1.1806 - accuracy: 0.6632 - val_loss: 0.9307 - val_accuracy: 0.8176
Epoch 11/12
469/469 [==============================] - 69s 147ms/step - loss: 1.0802 - accuracy: 0.6850 - val_loss: 0.8292 - val_accuracy: 0.8270
Epoch 12/12
469/469 [==============================] - 63s 135ms/step - loss: 1.0001 - accuracy: 0.7041 - val_loss: 0.7502 - val_accuracy: 0.8357
Test loss: 0.7502195835113525
Test accuracy: 0.8356999754905701
real 12m15.867s
user 59m51.727s
sys 30m20.694s
我尝试在启用了 GPU 的 Google Colab Notebook 上运行相同的代码，结果如下：
CPU times: user 2 µs, sys: 1e+03 ns, total: 3 µs
Wall time: 5.72 µs
x_train shape: (60000, 28, 28, 1)
60000 train samples
10000 test samples
Epoch 1/24
235/235 [==============================] - 6s 25ms/step - loss: 2.2962 - accuracy: 0.1233 - val_loss: 2.2759 - val_accuracy: 0.1710
Epoch 2/24
235/235 [==============================] - 6s 24ms/step - loss: 2.2674 - accuracy: 0.1693 - val_loss: 2.2421 - val_accuracy: 0.2451
Epoch 3/24
235/235 [==============================] - 6s 24ms/step - loss: 2.2344 - accuracy: 0.2293 - val_loss: 2.2036 - val_accuracy: 0.3520
Epoch 4/24
235/235 [==============================] - 6s 24ms/step - loss: 2.1973 - accuracy: 0.2848 - val_loss: 2.1584 - val_accuracy: 0.4723
Epoch 5/24
235/235 [==============================] - 6s 24ms/step - loss: 2.1523 - accuracy: 0.3398 - val_loss: 2.1037 - val_accuracy: 0.5635
Epoch 6/24
235/235 [==============================] - 6s 24ms/step - loss: 2.0966 - accuracy: 0.4012 - val_loss: 2.0366 - val_accuracy: 0.6309
Epoch 7/24
235/235 [==============================] - 6s 24ms/step - loss: 2.0308 - accuracy: 0.4533 - val_loss: 1.9543 - val_accuracy: 0.6766
Epoch 8/24
235/235 [==============================] - 6s 24ms/step - loss: 1.9489 - accuracy: 0.4958 - val_loss: 1.8547 - val_accuracy: 0.7109
Epoch 9/24
235/235 [==============================] - 6s 25ms/step - loss: 1.8527 - accuracy: 0.5337 - val_loss: 1.7375 - val_accuracy: 0.7357
Epoch 10/24
235/235 [==============================] - 6s 24ms/step - loss: 1.7465 - accuracy: 0.5607 - val_loss: 1.6061 - val_accuracy: 0.7532
Epoch 11/24
235/235 [==============================] - 6s 24ms/step - loss: 1.6292 - accuracy: 0.5872 - val_loss: 1.4659 - val_accuracy: 0.7702
Epoch 12/24
235/235 [==============================] - 6s 24ms/step - loss: 1.5146 - accuracy: 0.6075 - val_loss: 1.3261 - val_accuracy: 0.7851
Epoch 13/24
235/235 [==============================] - 6s 24ms/step - loss: 1.4024 - accuracy: 0.6272 - val_loss: 1.1938 - val_accuracy: 0.7976
Epoch 14/24
235/235 [==============================] - 6s 24ms/step - loss: 1.3001 - accuracy: 0.6442 - val_loss: 1.0753 - val_accuracy: 0.8071
Epoch 15/24
235/235 [==============================] - 6s 24ms/step - loss: 1.2117 - accuracy: 0.6586 - val_loss: 0.9745 - val_accuracy: 0.8172
Epoch 16/24
235/235 [==============================] - 6s 24ms/step - loss: 1.1344 - accuracy: 0.6747 - val_loss: 0.8900 - val_accuracy: 0.8249
Epoch 17/24
235/235 [==============================] - 6s 24ms/step - loss: 1.0698 - accuracy: 0.6881 - val_loss: 0.8203 - val_accuracy: 0.8292
Epoch 18/24
235/235 [==============================] - 6s 24ms/step - loss: 1.0108 - accuracy: 0.7033 - val_loss: 0.7624 - val_accuracy: 0.8370
Epoch 19/24
235/235 [==============================] - 6s 24ms/step - loss: 0.9621 - accuracy: 0.7141 - val_loss: 0.7140 - val_accuracy: 0.8441
Epoch 20/24
235/235 [==============================] - 6s 24ms/step - loss: 0.9267 - accuracy: 0.7212 - val_loss: 0.6742 - val_accuracy: 0.8498
Epoch 21/24
235/235 [==============================] - 6s 24ms/step - loss: 0.8904 - accuracy: 0.7322 - val_loss: 0.6397 - val_accuracy: 0.8543
Epoch 22/24
235/235 [==============================] - 6s 24ms/step - loss: 0.8588 - accuracy: 0.7395 - val_loss: 0.6105 - val_accuracy: 0.8579
Epoch 23/24
235/235 [==============================] - 6s 24ms/step - loss: 0.8309 - accuracy: 0.7464 - val_loss: 0.5850 - val_accuracy: 0.8607
Epoch 24/24
235/235 [==============================] - 6s 24ms/step - loss: 0.8031 - accuracy: 0.7561 - val_loss: 0.5624 - val_accuracy: 0.8643
Test loss: 0.5624315142631531
Test accuracy: 0.864300012588501
如您所见，差异很大。PlaidML 后端的准确率高于 99%，而在 TensorFlow 上则低于 80%。使用 GPU 和 CPU 设备的 TensorFlow 后端的执行时间也更长。我在这里做错了什么？如何在我的 MacBook 上高效地运行 Tensorflow？
不幸的是，Keras 2.3 版是最后一个支持多后端的版本。因此，在未来，Keras API 也将可用于 TensorFlow。所以，我希望将来也能在我的 MacBook 上高效地使用 Keras。
设备和软件规格：
Python version: 3.7.6
Tensorflow version: 2.2.0
Keras version: 2.2.4
Tensroflow.keras version: 2.3.-tf
Radeon Pro 560X:
Chipset Model: Radeon Pro 560X
Type: GPU
Bus: PCIe
PCIe Lane Width: x8
VRAM (Total): 4 GB
Vendor: AMD (0x1002)
gMux Version: 5.0.0
Metal: Supported, feature set macOS GPUFamily2 v1
CPU:
Processor Name: 6-Core Intel Core i7
Processor Speed: 2,6 GHz