Now let us first try to predict the classes of our test images, without retraining. First, we clear the default graph and define a placeholder for images:
tf.reset_default_graph()x_p = tf.placeholder(shape=(None,image_height, image_width,3), dtype=tf.float32,name='x_p')
The shape of placeholder x_p is (?, 224, 224, 3). Next, load the vgg16 model:
with slim.arg_scope(vgg.vgg_arg_scope()): logits,_ = vgg.vgg_16(x_p,num_classes=inet.n_classes, is_training=False)
Add the softmax layer for producing probabilities over the classes:
probabilities = tf.nn.softmax(logits)
Define the initialization function to restore the variables, such as weights and biases from the checkpoint file.
init ...