Underperforming on CIFAR10 example

I am trying out the CIFAR10 example code on keras, but I don’t get the expected results. The comment at the start of the file says:

It gets down to 0.65 test logloss in 25 epochs, and down to 0.55 after 50 epochs.

However in my case, after 25 epochs I am at 0.84, and after 50 I am at 0.75, and it doesn’t really go down any further from there :frowning:

I was wondering, has anyone else tried it, and what were the results? I am using the TF backend and Python 3. Snipped output:

x_train shape: (50000, 32, 32, 3)
50000 train samples
10000 test samples
Using real-time data augmentation.
Epoch 1/200
1562/1562 [==============================] - 14s - loss: 1.8745 - acc: 0.3141 - val_loss: 1.5766 - val_acc: 0.4326
Epoch 2/200
1562/1562 [==============================] - 12s - loss: 1.5852 - acc: 0.4179 - val_loss: 1.3793 - val_acc: 0.5049

[....]

Epoch 25/200
1562/1562 [==============================] - 12s - loss: 0.8431 - acc: 0.7100 - val_loss: 0.7394 - val_acc: 0.7456

[....]

Epoch 50/200
1562/1562 [==============================] - 12s - loss: 0.7547 - acc: 0.7427 - val_loss: 0.6530 - val_acc: 0.7824

[....]

Epoch 75/200
1562/1562 [==============================] - 12s - loss: 0.7592 - acc: 0.7484 - val_loss: 0.6522 - val_acc: 0.7924

[....]

Epoch 100/200
1562/1562 [==============================] - 12s - loss: 0.7908 - acc: 0.7396 - val_loss: 0.6784 - val_acc: 0.7798

Did anyone give this a try? (I literally just copied & pasted the code into a notebook and ran it).