5

I wrote neural-network using tensorflow tools. everything working and now I want to export the final weights of my neural network to make a single prediction method. How can I do this?

Or Perets
  • 389
  • 1
  • 9
  • 21

3 Answers3

3

You will need to save your model at the end of training by using the tf.train.Saver class.

While initializing the Saver object, you will need to pass a list of all the variables you wish to save. The best part is that you can use these saved variables in a different computation graph!

Create a Saver object by using, 

# Assume you want to save 2 variables `v1` and `v2`
saver = tf.train.Saver([v1, v2])

Save your variables by using the tf.Session object,

saver.save(sess, 'filename');

Of course, you can add additional details like global_step.

You can restore the variables in the future by using the restore() function. The restored variables will be initialized to these values automatically.

martianwars
  • 6,380
  • 5
  • 35
  • 44
  • 1
    Is it possible to obtain raw data of the parameters? I want to run a tensorflow-trained-model on anther platform, how can I do that? – Wesley Ranger Dec 13 '16 at 08:50
  • 2
    You can get the final value of the weights using `sess.run(weights)` and export them to a numpy array for instance – martianwars Dec 13 '16 at 08:55
  • That's what I need. Another problem: I used `tf.nn.rnn_cell.LSTMCell` in the net, how can I access the weight/bias of an `LSTMCell` object? – Wesley Ranger Dec 13 '16 at 08:59
  • 1
    If you try to use `tf.all_variables()`, you will see the names of the weight and bias of the `LSTMCell`. You can improve this search by searching for variables under the `LSTMCell` scope only – martianwars Dec 13 '16 at 09:02
  • http://stackoverflow.com/questions/36533723/tensorflow-get-all-variables-in-scope – martianwars Dec 13 '16 at 09:03
  • 1
    Thank you @martianwars . You just saved the day! – Wesley Ranger Dec 13 '16 at 09:10
1

The answer above is the standard way to save/restore session snapshot. However, if you want to export your network as a single binary file for further use with other tensorflow tools, you'll need to perform few more steps.

First, freeze the graph. TF provides the corresponding tool. I use it like this:

#!/bin/bash -x

# The script combines graph definition and trained weights into
# a single binary protobuf with constant holders for the weights.
# The resulting graph is suitable for the processing with other tools.


TF_HOME=~/tensorflow/

if [ $# -lt 4 ]; then
    echo "Usage: $0 graph_def snapshot output_nodes output.pb"
    exit 0
fi

proto=$1
snapshot=$2
out_nodes=$3
out=$4

$TF_HOME/bazel-bin/tensorflow/python/tools/freeze_graph --input_graph=$proto \
    --input_checkpoint=$snapshot \
    --output_graph=$out \
    --output_node_names=$out_nodes

Having done that, you can optimize it for inference, or use any other tool.

Dmytro Prylipko
  • 4,762
  • 2
  • 25
  • 44
0

If you simply need to access weights and biases of your neural net you can use the get_weights() method from tf.keras.layers.Layer

#params vector includes weights and biases
params = your_model.get_weights()
gigis95
  • 66
  • 5