Re- Help with model optimizer for converting existing model from TF to IR

[SDwiv1]

Hi,

I have already posted here before asking for help but since that thread has died without any resolution,

I would like to request assistance in converting my model from TF to IR,

A few details:
My model has 2 inputs which are a simple matrix (non-image) called input1 and input2, the input shapes are as following for input 1 & input 2: [1,30,128] & [1,30,28]. 

The following bit of code for converting our model:

class SpatialTemporalModel(object):
    def __init__(self, n_frame = 30):

        self.geo_inputs = tf.placeholder(dtype = tf.float32, shape=(1, n_frame, 182), name='input1')
        self.traj_inputs = tf.placeholder(dtype = tf.float32, shape = (1, n_frame, 28), name='input2'

    tf.train.Saver().save(model.sess, save_path='model/fd')
    tf.saved_model.simple_save(model.sess, export_dir='model/fd', inputs={"input1": model.geo_inputs, "input2":model.traj_inputs}, outputs={"output": model.output})

I'm using the following command to convert:

sudo python3 mo_tf.py —input_model ~/Desktop/12-3-19/fall_v1.1/frozen_model_1.pb —input input1,input2 —output output —input_shape [1,30,128],[1,30,28] —output_dir irmodels —log_level=DEBUG

I have also attached the debug log. I'm unable to understand what's causing the error while conversion. Any help will be appreciated.

Thank you,

Sweta Dwivedi

[Shubha_R_Intel]

Dearest Sweta, so sorry that you have been ignored so far. 

Which release of OpenVino are you using ? I see that the ERROR is directly related to LSTM:

extensions.middle.TF_lstm_cell_to_generic.TensorFlowLSTMtoGeneric

Can you kindly Message me with a zip file containing your frozen.pb ? Of if you want you can attach it to this ticket also.

Thanks for using OpenVino, and I will debug your problem then report my findings here.

Sincerely,

Shubha

[SDwiv1]

Hi Shubha,

How do i send you the zip file, i see that there is no way to message you directly?

I'm using ubuntu 18.04.1 LTS on oracle virtual box with OpenVino 2018 R5 or R5.0.1 (im not too sure)

If there's a way I can check the version, please let me know.

Thank you for your assistance

- Sweta

[Shubha_R_Intel]

Dear Sweta glad to see that you are on OpenVino R5. Do you not see a link something like "Send Author a Message" ? You should be able to send me a private message directly from this forum.  (use Send Author A Message link).

Thanks,

Shubha

[SDwiv1]

Hi Shubha,

I do not see any such link over here or on your profile.

Thank you,

Sweta

[Shubha_R_Intel]

Dear Sweta,

sorry that you're having difficulty finding the (use Send Author A Message link). Other forum posters have definitely sent me private messages so I know it works. Can you attach everything I need (your model which you're trying to convert to MO IR) to this ticket or is it too big ? 

Thanks,

Shubha

[nikos1]

Not sure where the "Send Author A Message" link is either.  Is it browser specific? Or could it be that Intel internal members can only see that? Also double-checked I am logged in but still no link.

Cheers,

Nikos

[SDwiv1]

Hi,

I have not only looked at this thread but also your profile for send the author a message link, i have not found it anywhere. I think its only available to intel user's. Please look into it

If possible please enable this feature or let me know how i can directly send you the model without making it available here publicly.

Thank you,

Sweta

[Shubha_R_Intel]

Dear Sweta and Nikos, I sent you each a message. did you each get it ? Sweta now you can (I hope) reply to it and attach your model.

Let me know if you can accomplish this and again, so super sorry for the trouble you've been having.

Shubha

[Shubha_R_Intel]

Sweta PM'd me her model and I reproduced her issue as well as filed a bug ticket. It definitely seems like a bug to me.  I will keep you posted Sweta !

Thanks for using OpenVino !

Shubha

[SDwiv1]

Hi Shubha,

Thank you very much. I will look forward to your update regarding this issue.

Meanwhile, I will try to run only 1 model on movidius and the other on CPU.

Please, let me know If there are any existing python code samples that do this. 

-Sweta