2 determine which predictions were incorrect and

2. Determine which predictions were incorrect and propagate back the difference
between the prediction and the true value (backpropagation).
3. Rinse and repeat till the predictions become sufficiently accurate.
It’s quite likely that the initial iteration would have close to 0% accuracy. Repeating
the process several times, however, can yield a highly accurate model (> 90%).
The batch size defines how many images are seen by the CNN at a time. It’s important
that each batch have a good variety of images from different classes in order to pre‐
vent large fluctuations in the accuracy metric between iterations. A sufficiently large
batch size would be necessary for that. However, it’s important not to set the batch
24 | Chapter 2: Cats vs Dogs - Transfer Learning in 30 lines with Keras

size too large for a couple of reasons. First, if the batch size is too large, you could end
up crashing the program due to lack of memory. Second, the training process would
be slower. Usually, batch sizes are set as powers of 2. 64 is a good number to start with
for most problems and you can play with the number by increasing/decreasing it.
Data Augmentation
Usually, when you hear deep learning, you associate it with millions of images. But
500 images like what we have might be a low number for real-world training. Now,
these deep neural networks are powerful, a little too powerful for small quantities of
data. The danger of a limited set of training images is that the neural network might
memorize your training data, and show great prediction performance on the training
set, but bad accuracy on the validation set. In other words, the model has overtrained
and does not generalize on previously unseen images. And we don’t want that, right?
There are often cases where there’s not enough data available. Perhaps you’re working
on a niche problem and data is hard to come by. There are a few ways you can artifi‐
cially augment your dataset:
1. Rotation: In our example, we might want to rotate the 500 images randomly by
20 degrees in either direction, yielding up to 20000 possible unique images.
2. Random Shift: Shift the images slightly to the left, or to the right.
3. Zoom: Zoom in and out slightly of the image
By combining rotation, shifting and zooming, the program can generate almost infin‐
ite unique images. This important step is called data augmentation. Keras provides
ImageDataGenerator function that augments the data while it is being loaded from
the directory. Example augmentations generated by the imgaug (
aleju/imgaug ) for a sample image are shown in Figure 2-3 .
Building a Custom Classi er in Keras with Transfer Learning | 25

Figure 2-3. Possible image augmentations generated from a single image by imgaug
library
Colored images usually have 3 channels - red, green, and blue. Each channel has an
intensity value ranging from 0 to 255. To normalize it (i.e. scale down the value to
between 0 and 1), we will divide each pixel by 255.