Retrained mobilenet model conversion error

[idata]

Hi,

We've recently retrained a mobilenet model using the Tensorflow's Image Retraining script given at : https://github.com/tensorflow/tensorflow/blob/r1.5/tensorflow/examples/image_retraining/retrain.py. This gives a Pb file which when used for profiling gives the following error :

[Error 5] Toolkit Error: Stage Details Not Supported: Top Not Supported - Constants input_1/BottleneckInputPlaceholder

How to remove this error ?

[idata]

@karthik Not sure why you're receiving that error. Does this model use multiple inputs? The current NCSDK (2.05.xx) doesn't support multiple inputs at the moment. Can you provide your model files for testing? It would also be helpful if you could provide any instructions you used to produce issue. Thanks.

[idata]

Hi @Tome_at_Intel.

Here's the model frozen graph : https://drive.google.com/file/d/1BhxsEyBuNQlv-sOXAnjWyk_-FOf8NrHQ/view?usp=sharing

I got the error while profiling it with mvNCProfile :

mvNCProfile retrained_graph.pb -in input -on final_result -s 12

The model doesn't take multiple inputs

[idata]

retrain.py. ? Thats the same one that everyone uses. You know there is a version 1.7 and 1.8 as well. What version of tensorflow do you have installed.?

https://github.com/tensorflow/tensorflow/blob/r1.8/tensorflow/examples/image_retraining/retrain.py.

FYI the PB file I got did not compile

[idata]

@chicagobob123 I have tensorflow v 1.6 as when i had compiled ncsdk, it didn't support 1.7

[idata]

@Tome_at_Intel any help ?

[idata]

@karthik I'm still investigating your network. Thanks for your patience.

[idata]

@karthik Since retraining your model, did you remove the training specific code? I think this could be the cause. Try removing all dropout layers and follow the guide at https://movidius.github.io/ncsdk/tf_compile_guidance.html.

[idata]

@Tome_at_Intel This is the frozen graph is got directly by the tensorflow retrain script. How do I remove the training specific code ?

retrain file adds a placeholder BottleneckInputPlaceholder which is a PlaceholderWithDefault and as i see from the release info is supported by sdk since v 2.04

[idata]

I would love to see a how to use retrain.py with tensorflow to created your own graphs.

So far I think the most frustrating thing about the stick is its not clear, and seems to often fail,when you try to reuse pre-made networks such as inceptionV3 to create and test your own classifications.

[idata]

@karthik This link might be helpful. FYI turning on the debug flag in TensorFlowParser.py, the error is occurring at the following op:

Identity 557 MatMul final_training_ops/Wx_plus_b/MatMul IN: input_1/BottleneckInputPlaceholder:0 IN: final_training_ops/weights/final_weights/read:0 OUT: final_training_ops/Wx_plus_b/MatMul:0

[idata]

@Tome_at_Intel thanks for the link. I did try the 'removing training-only nodes' but the BottleneckInputPlaceholder is not removed by that operation