- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Several users of this forum have asked the question how to fix error 34: [Error 34] Setup Error: Values for input contain placeholder. Pass an absolute value.
- https://ncsforum.movidius.com/discussion/339/is-there-a-tutorial-or-documentation-for-running-custom-tensorflow-network-in-movidius/p1
- https://ncsforum.movidius.com/discussion/368/error-5-or-error-34-converting-custom-tensorflow-model/p1
- https://ncsforum.movidius.com/discussion/497/any-guide-for-yolo-v2-graph-conversion-by-using-darkflow-with-tensorflow/p1: @AshwinVijayakumar mentioned working on an tutorial article
- https://ncsforum.movidius.com/discussion/441/problems-compiling/p1: @Tome_at_Intel was looking into the issue on November 27th
- https://ncsforum.movidius.com/discussion/517/compile-self-defined-tensorflow-model-with-mvnccompile/p1: @Yang solved it theirself
As far as I can find there has been no solution to this specific issue anywhere. Also, the promised blogpost explaining how to get custom TF models working on the NCS has not been posted. I hope to center the insight into this error in this thread. Does somebody have insight into the actual cause of 34? Can someone who fixed it for themselves share what fixed it? Or even better, can we get an update from the Movidius team on this issue?
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Same question.
Was unable to solve this.
Let me Know if u have any updates.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Apparently there has been a new release on Github: https://github.com/movidius/ncsdk/releases/tag/v1.12.00.01
It also mentions a new page in the documentation that guides users wanting to use a custom model: https://movidius.github.io/ncsdk/tf_compile_guidance.html
Going to look at it myself now!
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
I have not been able to solve the problem using either the new release or the new guide to using custom models.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@rjbruin We are working on a solution to this issue. I will let you know as soon as we have the fix and I will leave a reply post on this thread and tag you. Thank you for your patience and understanding.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@rjbruin @ashwinnair14 From https://movidius.github.io/ncsdk/tf_compile_guidance.html, "The ncsdk compiler does not resolve unknown placeholders. Often extra placeholders are used for training specific variables so they are not necessary for inference. Placeholder variables that cannot be removed should be replaced by constants in the inference graph."
For example, if we have a placeholder variable like this: tf.placeholder(tf.float32, [None, 784], name="input")
It would need to be changed to this: tf.placeholder(tf.float32, [1, 784], name="input")
There is also more information @ https://movidius.github.io/ncsdk/tf_compile_guidance.html about how to modify your Tensorflow program so it can work with the NCS.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@Tome_at_Intel
Replacing the None's in the placeholder with specific sizes did the trick! I have to say I find the guide on custom models misleading on this front though, since the very first code snippet there does use None's in the placeholder.
Anyway, that solves the error 34. I do however get a new error:
[Error 5] Toolkit Error: Stage Details Not Supported: StridedSlice attributes not supported
I've made a new thread for that error, so this thread doesn't derail. I'll let the posters of the other threads know about this solution as well!
Thanks @Tome_at_Intel for your help! :)
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
I have to say I find the guide on custom models misleading on this front though, since the very first code snippet there does use None's in the placeholder.
Thanks for the feedback @rjbruin, we will update the doc.
Below are a couple examples that show how to train simple TF models and then export them for inference on NCS:
MNIST example: https://github.com/ashwinvijayakumar/ncappzoo/tree/mnist/tensorflow/mnist
Fashion MNIST example: https://github.com/ashwinvijayakumar/ncappzoo/tree/fashion-mnist/tensorflow/fashion-mnist
A super-simple custom network trained on fashion MNIST: https://github.com/ashwinvijayakumar/ncappzoo/tree/single-conv/tensorflow/single-conv
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@AshwinVijayakumar trying to look at at the super-simple custom network trained on fashion MNIST but you removed it " ashwinvijayakumar Temporarily removed fashion-mnist, mnist & single-conv projects."
Can you please share something that could help with TF custom models?
Thanks
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@antoniocappiello, you can find these examples in dedicated branches. I have updated the original post with the proper links: https://ncsforum.movidius.com/discussion/comment/1579/#Comment_1579
I will pull these into movidius/ncappzoo soon.
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page