Reusing The Graph

So once we’ve optimized the graph and saved it to a file how do we use it?

First load the graph from the binary protobuf file:

from graph_utils import load_graph

sess, _ = load_graph(‘your_graph.pb’)
graph = sess.graph

Once the graph is loaded we can access operations and tensors with the get_operation_by_name and get_tensor_by_name tf.Graph methods respectively. We could also retrieve all the operations with of a tf.Graph with the get_operations method, which is very useful if we are unsure which operation to use. In this case we want to pass an image as input and receive the softmax probabilities as output:

image_input = graph.get_tensor_by_name('image_input:0')
keep_prob = graph.get_tensor_by_name('keep_prob:0')
softmax = graph.get_tensor_by_name('Softmax:0')

Then we can take an image and run the computation in a session as we normally would. keep_prob is related to the dropout probability.

probs = sess.run(softmax, {image_input: img, keep_prob: 1.0})