Freezing the graph and then reloading for inference

[camer314]

Hello,

I am wanting to freeze your model into a .PB file and then later reload that when i want to pass my own sentences. I am familiar with image training where operations are named (e.g 'input' and 'final_result') and i feed the necessary tensors.

However in your model there are no named operations and the testing script does a lot of heavy lifting with restoration and loading of vector files.

Is it possible to just load a PB file and pass in a sentence? And if so then what do I pass that sentence in to and then what tensor do i need as a result?

Basically i am trying to get this running on TF mobile, so only using the Java/C# library, all my other models i have just frozen to PB, restored and fed the correct tensor in, got the results and i am wanting to do the same here but not sure what the absolute minimum requirement is.

[camer314]

I think i am on the right track here, at the very least it returns me an answer, but given my very small number of training iterations I wont know until fully trained if its the right answer

To start i gave the input_data tensor a name:

input_data = tf.placeholder(tf.int32, [batchSize, maxSeqLength], name="camtest_input_data")

Then when i printed the value of prediction = (tf.matmul(last, weight) + bias) to came back with a name of "add:0" which i am taking as being my output tensor

I can now write out the graph using "add" as my output name:

`
from tensorflow.python.framework import graph_util
from tensorflow.python.platform import gfile

output_graph_def = graph_util.convert_variables_to_constants(sess, sess.graph.as_graph_def(), ["add"])

with gfile.FastGFile("frozen_graph.pb", 'wb') as f:
f.write(output_graph_def.SerializeToString())

`

Now when i want to pass new sentences in later on, I can restore the frozen graph and feed the new information like this, which at least doesnt crash but not sure if its a complete replica of the testing script using checkpoint files? In particular am i right in what i think is the crucial input and output?

`inputText = FLAGS.sentence
inputMatrix = getSentenceMatrix(inputText)

tf.reset_default_graph()
graph = load_graph(os.path.join(local_path, "frozen_graph.pb"))

with tf.Session(graph=graph) as sess:
prediction = sess.graph.get_tensor_by_name('add:0')
predictedSentiment = sess.run(prediction, {'camtest_input_data:0': inputMatrix})[0]
`

[adeshpande3]

Hmm, I'm actually not too familiar with saving into .pb files, but the general approach you have of naming the variables, using f.write, then load_graph, and then sess.run seems right to me. Will let you know if I find any more info on this.

[camer314]

Thanks that would be great

[dreamibor]

@camer314 Is it possible to share the frozen model (.pb)?