Using Tensorflow model for prediction

Question

I am new to tensorflow. I have manged to train and validate a CNN, saved the session through the Saver object into a CPKT file and loaded it back. Now I'd like to use the trained model in order to check how it performs against a photo I took myself. I can't figure out how to do it.

Answer 1

Look at this blog. Mainly you have saved operations as a part of your computational graph. So after you load your model, you can restore the session and call the predict operation that you created for training and validating your data, and run it on the new data hy feeding into the feed_dict. Make sure it is in the same format and same shape as your training data. I am adding a small snippet of code, maybe that will help.

checkpoint_file=tf.train.latest_checkpoint(checkpoint_directory)
graph=tf.Graph()

 with graph.as_default():
     session_conf = tf.ConfigProto(allow_safe_placement=True, log_device_placement =False)
     sess = tf.Session(config = session_conf)
     with sess.as_default():
          saver =     tf.train.import_meta_graph("{}.meta".format(checkpoint_file))
          saver.restore(sess,checkpoint_file)
          input = graph.get_operation_by_name("input").outputs[0]
          prediction=graph.get_operation_by_name("prediction").outputs[0]
          #newdata=put your data here
          print sess.run(prediction,feed_dict={input:newdata})

To use operation_by_name, you will have to name the operations and variables in the original model. Secondly, in the feed dict you will have to give all the parameters needed to perform the operation(prediction here), so if you are using dropout you will have to give that too. There might be more simpler ways of doing this. This is the way I did it.