Intel® Distribution of OpenVINO™ Toolkit
Community assistance about the Intel® Distribution of OpenVINO™ toolkit, OpenCV, and all aspects of computer vision-related on Intel® platforms.

Convert Tensorflow CNN model to NCS graph

idata
Employee
1,031 Views

Hello,

 

I've implemented a CNN with tensorflow i would like to get pointers on how i can convert/compile this model to the compute stick.

0 Kudos
7 Replies
idata
Employee
597 Views
0 Kudos
idata
Employee
597 Views

@GoldenWings, in order to run your own model on the NCS, you have to follow these steps:

 

     

  1. Export your model's graphdef (you can use tf.train.write_graph to export the graphdef from within your tf session)
  2.  

  3. Freeze your graphdef using freeze_graph.py (you'll need the checkpoint files)
  4.  

  5. Pass the frozen model to mvNCCompile along with the input and output nodes.
  6.  

 

I have explained this process in my webinar (see timestamp 00:26:34) - https://software.seek.intel.com/Edge_Devices_Webinar_Reg

 

Here's an example of a custom CNN deployed on NCS - https://github.com/ashwinvijayakumar/ncappzoo/tree/single-conv/tensorflow/single-conv

0 Kudos
idata
Employee
597 Views

@AshwinVijayakumar Thanks so much i will try to run my own model through the steps you provided and will let you know how things going, to keep you updated here is my tf model along with training code i want to know do i have to change anything to make it adapt with NCS:-

 

sess = tf.InteractiveSession(config=tf.ConfigProto())

 

x = tf.placeholder(tf.float32, shape=[None, 240, 320, 3], name='x') y_ = tf.placeholder(tf.float32, shape=[None, 3], name='y_') phase = tf.placeholder(tf.bool, name='phase') conv1 = batch_norm_pool_conv_layer('layer1', x, [6, 6, 3, 24], phase) conv2 = batch_norm_conv_layer('layer2',conv1, [6, 6, 24, 24], phase) conv3 = batch_norm_pool_conv_layer('layer3',conv2, [6, 6, 24, 36], phase) conv4 = batch_norm_conv_layer('layer4',conv3, [6, 6, 36, 36], phase) conv5 = batch_norm_pool_conv_layer('layer5',conv4, [6, 6, 36, 48], phase) conv6 = batch_norm_conv_layer('layer6',conv5, [6, 6, 48, 64], phase) conv7 = batch_norm_pool_conv_layer('layer7',conv6, [6, 6, 64, 64], phase) h_pool7_flat = tf.reshape(conv7, [-1, 15 * 20 * 64]) h8 = batch_norm_fc_layer('layer8',h_pool7_flat, [15 * 20 * 64, 512], phase) h9 = batch_norm_fc_layer('layer9',h8, [512, 256], phase) h10 = batch_norm_fc_layer('layer10',h9, [256, 128], phase) h11 = batch_norm_fc_layer('layer11',h10, [128, 64], phase) W_final = weight_variable('layer12',[64, 3]) b_final = bias_variable('layer12',[3]) logits = tf.add(tf.matmul(h11, W_final), b_final, name='logits') cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=y_)) train_step = tf.train.AdamOptimizer(1e-5,name='train_step').minimize(cross_entropy) correct_prediction = tf.equal(tf.argmax(logits,1), tf.argmax(y_,1)) accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32),name='accuracy') update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) with tf.control_dependencies(update_ops): train_step = tf.train.AdamOptimizer(1e-5).minimize(cross_entropy)

 

Also i have another model that is written in keras with tensorflow backend, i wanted to know is it possible to run on NCS?

0 Kudos
idata
Employee
597 Views

@Tome_at_Intel May i have your opinion please?

0 Kudos
idata
Employee
597 Views

@GoldenWings I don't see anything that would cause it not to run on the NCS. If you have issues with the NCSDK complaining about the placeholder, you may have to change the None in your code. x = tf.placeholder(tf.float32, shape=[None, 240, 320, 3], name='x') y_ = tf.placeholder(tf.float32, shape=[None, 3], name='y_') will become x = tf.placeholder(tf.float32, shape=[1, 240, 320, 3], name='x') y_ = tf.placeholder(tf.float32, shape=[1, 3], name='y_')

 

Let me know if there are any problems you face when running your network.

0 Kudos
idata
Employee
597 Views

@Tome_at_Intel I will keep you updated once i run it . i have another model that is written in keras with tensorflow backend, would like to know if it is supported by ncs and if it is is there any example out there?

0 Kudos
idata
Employee
597 Views

@GoldenWings I have not tried using a model that has been written in Keras with a TensorFlow backend on the NCS myself, but please let us know how it goes for you.

0 Kudos
Reply