Significantly higher testing accuracy on mnist with keras than tensorflow.keras - tensorflow

Question

I was verifying with a basic example my TensorFlow (v2.2.0), Cuda (10.1), and cudnn (libcudnn7-dev_7.6.5.32-1+cuda10.1_amd64.deb) and I'm getting weird results...
When running the following example in Keras as shown in https://keras.io/examples/mnist_cnn/ I get the ~99% acc #validation. When I adapt the imports run via the TensorFlow I get only 86%.
I might be forgetting something.
To run using tensorflow:
from __future__ import print_function
import tensorflow as tf
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
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 = tf.keras.utils.to_categorical(y_train, num_classes)
y_test = tf.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=tf.keras.losses.categorical_crossentropy,
optimizer=tf.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])
Sadly, I get the following output:
Epoch 2/12
469/469 [==============================] - 3s 6ms/step - loss: 2.2245 - accuracy: 0.2633 - val_loss: 2.1755 - val_accuracy: 0.4447
Epoch 3/12
469/469 [==============================] - 3s 7ms/step - loss: 2.1485 - accuracy: 0.3533 - val_loss: 2.0787 - val_accuracy: 0.5147
Epoch 4/12
469/469 [==============================] - 3s 6ms/step - loss: 2.0489 - accuracy: 0.4214 - val_loss: 1.9538 - val_accuracy: 0.6021
Epoch 5/12
469/469 [==============================] - 3s 6ms/step - loss: 1.9224 - accuracy: 0.4845 - val_loss: 1.7981 - val_accuracy: 0.6611
Epoch 6/12
469/469 [==============================] - 3s 6ms/step - loss: 1.7748 - accuracy: 0.5376 - val_loss: 1.6182 - val_accuracy: 0.7039
Epoch 7/12
469/469 [==============================] - 3s 6ms/step - loss: 1.6184 - accuracy: 0.5750 - val_loss: 1.4296 - val_accuracy: 0.7475
Epoch 8/12
469/469 [==============================] - 3s 7ms/step - loss: 1.4612 - accuracy: 0.6107 - val_loss: 1.2484 - val_accuracy: 0.7719
Epoch 9/12
469/469 [==============================] - 3s 6ms/step - loss: 1.3204 - accuracy: 0.6402 - val_loss: 1.0895 - val_accuracy: 0.7945
Epoch 10/12
469/469 [==============================] - 3s 6ms/step - loss: 1.2019 - accuracy: 0.6650 - val_loss: 0.9586 - val_accuracy: 0.8097
Epoch 11/12
469/469 [==============================] - 3s 7ms/step - loss: 1.1050 - accuracy: 0.6840 - val_loss: 0.8552 - val_accuracy: 0.8216
Epoch 12/12
469/469 [==============================] - 3s 7ms/step - loss: 1.0253 - accuracy: 0.7013 - val_loss: 0.7734 - val_accuracy: 0.8337
Test loss: 0.7734305262565613
Test accuracy: 0.8337000012397766
Nowhere near 99.25% as when I import Keras.
What am I missing?

Answer 1

Discrepancy in optimiser parameters between keras and tensorflow.keras
So the crux of the issue lies in the different default parameters for the Adadelta optimisers in Keras and Tensorflow. Specifically, the different learning rates. We can see this with a simple check. Using the Keras version of the code, print(keras.optimizers.Adadelta().get_config()) outpus
{'learning_rate': 1.0, 'rho': 0.95, 'decay': 0.0, 'epsilon': 1e-07}
And in the Tensorflow version, print(tf.optimizers.Adadelta().get_config() gives us
{'name': 'Adadelta', 'learning_rate': 0.001, 'decay': 0.0, 'rho': 0.95, 'epsilon': 1e-07}
As we can see, there is a discrepancy between the learning rates for the Adadelta optimisers. Keras has a default learning rate of 1.0 while Tensorflow has a default learning rate of 0.001 (consistent with their other optimisers).
Effects of a higher learning rate
Since the Keras version of the Adadelta optimiser has a larger learning rate, it converges much faster and achieves a high accuracy within 12 epochs, while the Tensorflow Adadelta optimiser requires a longer training time. If you increased the number of training epochs, the Tensorflow model could potentially achieve a 99% accuracy as well.
The fix
But instead of increasing the training time, we can simply initialise the Tensorflow model to behave in a similar way to the Keras model by changing the learning rate of Adadelta to 1.0. i.e.
model.compile(
loss=tf.keras.losses.categorical_crossentropy,
optimizer=tf.optimizers.Adadelta(learning_rate=1.0), # Note the new learning rate
metrics=['accuracy'])
Making this change, we get the following performance running on Tensorflow:
Epoch 12/12
60000/60000 [==============================] - 102s 2ms/sample - loss: 0.0287 - accuracy: 0.9911 - val_loss: 0.0291 - val_accuracy: 0.9907
Test loss: 0.029134796149221757
Test accuracy: 0.9907
which is close to the desired 99.25% accuracy.
p.s. Incidentally, it seems that the different default parameters between Keras and Tensorflow is a known issue that was fixed but then reverted:
https://github.com/keras-team/keras/pull/12841 software development is hard.