In this example, we will use transfer learning for feature extraction and a German traffic sign dataset to develop a classifier. Used here is an AlexNet implementation by Michael Guerzhoy and Davi Frossard, and AlexNet weights are from the Berkeley vision and Learning center. The complete code and dataset can be downloaded from here.

AlexNet expects a 227 x 227 x 3 pixel image, whereas the traffic sign images are 32 x 32 x 3 pixels. In order to feed the traffic sign images into AlexNet, we'll need to resize the images to the dimensions that AlexNet expects, that is, 227 x 227 x 3:

original_image = tf.placeholder(tf.float32, (None, 32, 32, 3))
resized_image = tf.image.resize_images(original_imag, (227, 227))

We can do so with the help of the tf.image.resize_images method by TensorFlow. Another issue here is that AlexNet was trained on the ImageNet dataset, which has 1,000 classes of images. So, we will replace this layer with a 43-neuron classification layer. To do this, figure out the size of the output from the last fully connected layer; since this is a fully connected layer and so is a 2D shape, the last element will be the size of the output. fc7.get_shape().as_list()[-1] does the trick; combine this with the number of classes for the traffic sign dataset to get the shape of the final fully connected layer: shape = (fc7.get_shape().as_list()[-1], 43). The rest of the code is just the standard way to define a fully connected layer in TensorFlow. Finally, calculate the probabilities with softmax:

#Refer AlexNet implementation code, returns last fully connected layer
fc7 = AlexNet(resized, feature_extract=True)
shape = (fc7.get_shape().as_list()[-1], 43)
fc8_weight = tf.Variable(tf.truncated_normal(shape, stddev=1e-2))
fc8_b = tf.Variable(tf.zeros(43))
logits = tf.nn.xw_plus_b(fc7, fc8_weight, fc8_b)
probs = tf.nn.softmax(logits)