Can't Compile a Simple Network

idata · ‎06-13-2018

I am trying to compile a _very_ simple neural network written in TensorFlow and saved as a .meta graph. The network was taken directly from this tutorial:

http://stackabuse.com/tensorflow-save-and-restore-models/

which I followed in order to learn how to save and restore graphs. I have made minor changes to name inputs and outputs (I'm including all my Python notebooks and the saved .meta). When trying to compile using mcNCCompile I am stuck on the TensorFlow error:

tensorflow.python.framework.errors_impl.InvalidArgumentError: Input 1 (tensor_name) must be a string scalar; got a tensor of 2elements
     [[Node: save/RestoreV2 = RestoreV2[dtypes=[DT_FLOAT, DT_FLOAT], _device="/job:localhost/replica:0/task:0/device:CPU:0"](_arg_save/Const_0_0, save/RestoreV2/tensor_names, save/RestoreV2/shape_and_slices)]]

It is not obvious to me where it expects a tensor with a type "string scalar" and so I'm not clear where it is being fed a tensor of "2elements".

This was intended to be a simple test of mvNCCompile. This command is giving me errors on a larger, more complicated network and I wanted to run a simple experiment with a simple network to help me better understand the process. I am disappointed that I am still running into hiccups here. Any help would be much appreciated.

I am including my code below. I tried uploading the files but they aren't allowed (!?). I cannot upload the .meta file either (why can't I attach any files???). The .meta can easily be produced by running SaveTest.py as long as there is a folder named "saved" in the same directory.

After producing the .meta graph I can produce the error by running:

mvNCCompile model_final.meta -in=X -on=y

or even just

mvNCCompile model_final.meta

"""
SaveTest.py

This file builds, trains and saves the network.
"""

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt

tf.reset_default_graph()

X = tf.placeholder(tf.float32, name="X")  
Y = tf.placeholder(tf.float32, name="y")

h_est = tf.Variable(0.0, name='h_estimate')  
v_est = tf.Variable(0.0, name='v_estimate')

y_est = tf.square(X - h_est) + v_est

cost = (tf.pow(Y - y_est, 2))

trainop = tf.train.GradientDescentOptimizer(0.001).minimize(cost)  

h = 1  
v = -2

x_train = np.linspace(-2, 4, 201)  
noise = np.random.randn(*x_train.shape) * 0.4  
y_train = (x_train - h) ** 2 + v + noise

plt.rcParams['figure.figsize'] = (10, 6)  
plt.scatter(x_train, y_train)  
plt.xlabel('x_train')  
plt.ylabel('y_train') 

saver = tf.train.Saver()

init = tf.global_variables_initializer()

def train_graph():  
    with tf.Session() as sess:
        sess.run(init)
        for i in range(100):
            for (x, y) in zip(x_train, y_train):

                sess.run(trainop, feed_dict={X: x, Y: y})

            saver.save(sess, 'saved/model_iter', global_step=i)

        saver.save(sess, 'saved/model_final')
        print("Training complete!")
        h_ = sess.run(h_est)
        v_ = sess.run(v_est)
    return h_, v_

result = train_graph()  
print("h_est = %.2f, v_est = %.2f" % result)  



"""
RestoreTest.py

This file reloads the network via the saved .meta and prints
the learned parameters
"""

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt

tf.reset_default_graph()

imported_meta = tf.train.import_meta_graph("saved/model_final.meta")

with tf.Session() as sess:  
    imported_meta.restore(sess, tf.train.latest_checkpoint('saved'))
    h_est2 = sess.run('h_estimate:0')
    v_est2 = sess.run('v_estimate:0')
    print("h_est: %.2f, v_est: %.2f" % (h_est2, v_est2))