TensorFlow Implementation

[3ygun]

Goals

Use TensorFlow for model back-ends to enable better extensibility. This issue will server as a base of operations regarding the progress and/or discussion around the implementation.

Aspects

@tylermzeller and I believe the following would be some of the required implementation plan to enable the above goal:

• Rewrite Server to use TensorFlow as it's testing interface
  • A simple extension of the existing NodeJS server application wouldn't be possible due to TensorFlow's lack of support for JavaScript bindings.
  • In light of this a rewrite in Python would probably be most appropriate. It supports the largest part of TensorFlow's implemented models and documentation while being accessible and performant
    • Could use pyrebase
• Rewrite of the Andorid app to use the TensorFlow bindings and TrainingInterface explored and developed in TensorFlow on Android
  • Lots of set-up but we've been good about documenting it mostly
  • The Java bindings are currently not stable BUT have been made more available for public comment see api_docs
  • Single model Android TensorFlow (with wifi & refractoring changes) #32
  • Any model

Pre-Requisites

NOTE: most of these can be explored through desktop TensorFlow applications

• Validate that weights can be changed after a model is loaded
  • If this is not the case a work around must be found (such as a model reload)
  • EDIT: view the TF documentation on variables.

Structure Project as

Single app vs Library

[tylermzeller]

StackOverflow article on how to assign values to variables in TF

[3ygun]

Testing with:

import tensorflow as tf

with tf.Session() as sess:

    x = tf.placeholder(tf.float32, shape=[None, 784], name="x")
    y = tf.placeholder(tf.float32, [None, 10], name="y")
    w = tf.placeholder(tf.float32, [784, 10], name="weights_in")

    W = tf.Variable(tf.zeros([784, 10]), name="weights")
    b = tf.Variable(tf.zeros([10]))

    y_out = tf.add(tf.matmul(x, W), b, name="y_out")

    #cross_entropy = tf.reduce_mean(-tf.reduce_sum(y * tf.log(y_out), reduction_indices=[1]))
    cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=y_out))
    train_step = tf.train.AdamOptimizer(0.005).minimize(cross_entropy, name="train")

    correct_prediction = tf.equal(tf.argmax(y_out,1), tf.argmax(y,1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name="test")

    init = tf.variables_initializer(tf.global_variables(), name="init")

    tf.train.write_graph(sess.graph_def,
                         './',
                         'mnist_adam_0.005_mlp.pb', as_text=False)

[3ygun]

Performance

When running the PR #32 I get the following performance bottleneck with the Data file reading taking up a large majority of the time:

[tylermzeller]

I don't know that there are good alternatives to buffered IO. The only other way that could be possible is reading the data once and keeping all the samples in memory. Just to allocate enough memory for the training features, we need 784 4-byte floats to represent one feature vector. 60000 samples gives us 784x4x60,000 bytes which is ~188 Mb of memory. Thats a bit much.