Error in v0.9 using python3.x about tensorflow-on-raspberry-pi

[samjabrahams]

Thanks for opening this up! Very weird- I'll check this out as soon as I get the chance.

[wtchang13]

Tensorflow 0.9.0 is also giving the wrong classification with python 2.7:

window screen (score = 0.05178)
web site, website, internet site, site (score = 0.02289)
honeycomb (score = 0.01543)
window shade (score = 0.01286)
nematode, nematode worm, roundworm (score = 0.00703)
Build graph time: 3.489147
Eval time: 9.210417

It seems to be an error with how the image is being processed since the classifications indicate that it may be reading in a blank white screen or something like that.

[wtchang13]

v0.9 also has the same classification error with python 3.4

/usr/local/lib/python3.4/dist-packages/tensorflow/python/ops/array_ops.py:1197: VisibleDeprecationWarning: converting an array with ndim > 0 to an index will result in an error in the future
result_shape.insert(dim, 1)
W tensorflow/core/framework/op_def_util.cc:332] Op BatchNormWithGlobalNormalization is deprecated. It will cease to work in GraphDef version 9. Use tf.nn.batch_normalization()
...

it continues similar to my post above

[zxzhijia]

@shrimpwonton Did you also trained it on Raspberry Pi?

[wtchang13]

@zxzhijia no, I used the example in this repo (in benchmarks/inceptionv3). It also gives the same error if I use the file in the tensorflow python package.

[zxzhijia]

Ok, then I'm not sure what's going on. Sorry about that.

On Tuesday, July 12, 2016, shrimpwonton [email protected] wrote:

@zxzhijia
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_zxzhijia&d=CwMCaQ&c=imBPVzF25OnBgGmVOlcsiEgHoG1i6YHLR0Sj_gZ4adc&r=X3YsXL1AZsKHgsh28HFY8A&m=PAn8MN4A2TBbDqXriZVtDejk4jyqUTqq_HudvbLVTr0&s=GZR4RLcMnJzQbepql6D3KYVw_FjMd0qMq8FO5JRymIs&e=
no, I used the example in this repo (in benchmarks/inceptionv3). It also
gives the same error if I use the file in the tensorflow python package

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_samjabrahams_tensorflow-2Don-2Draspberry-2Dpi_issues_29-23issuecomment-2D232224200&d=CwMCaQ&c=imBPVzF25OnBgGmVOlcsiEgHoG1i6YHLR0Sj_gZ4adc&r=X3YsXL1AZsKHgsh28HFY8A&m=PAn8MN4A2TBbDqXriZVtDejk4jyqUTqq_HudvbLVTr0&s=GzXpAJMIoWGNqn52hTtRfHlHofjH7v50yS5xPnD6nxc&e=,
or mute the thread
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_notifications_unsubscribe_ALPsQLUvcJ9ngDX2jE7BrB9tzYuzw3Rpks5qVDVpgaJpZM4JIhJb&d=CwMCaQ&c=imBPVzF25OnBgGmVOlcsiEgHoG1i6YHLR0Sj_gZ4adc&r=X3YsXL1AZsKHgsh28HFY8A&m=PAn8MN4A2TBbDqXriZVtDejk4jyqUTqq_HudvbLVTr0&s=LQVMjUQJNEgO5PdyG7PGcNRjBFTd3ok-FjDsU3nIIvU&e=
.

[samjabrahams]

I'll be looking into this. I had to change some of the compiler flags for floating point variables in 0.9 due to the use of Eigen::half inside of TensorFlow, but I'm not sure if that would cause straight up miscalculations by itself.

[samjabrahams]

Hey all- not sure if this is going to be a fix or not, but I found that some of the changes introduced in the official repository related to specifying the platform for compilation may have messed up some functionality. Recompiling the binaries tonight/tomorrow and will try to see if it helps out with any of these errors.

Copied from #31:

If you're interested in the details (at least what I'm pretty sure is going on): Essentially, these lines are causing the compilation to target the Raspberry Pi as a mobile device, which in turn causes no 64-bit floating points to be registered for any Ops in this file. The current hack-y fix is to delete this line. If that works, I'll try to get a fix introduced into the official repository that is a bit more robust.

[samjabrahams]

@eyewy It's possible that the version of NumPy installed on your Raspberry Pi is simply the wrong version due to another package you've installed. You can either try creating a Virtualenv and reinstalling TensorFlow (as described here, just install the RPi binaries instead of the official binaries), or you can try pip uninstall numpy or pip3 uninstall numpy before reinstalling the TensorFlow binaries.

@shrimpwonton I'm still testing these, but I've updated these binaries in response to the aforementioned changes. Let me know if the classification error is still messed up for you- it works on my end with these guys.

Python2.7: https://github.com/samjabrahams/tensorflow-on-raspberry-pi/raw/master/bin/tensorflow-0.9.0-cp27-none-linux_armv7l.whl

Python3: https://github.com/samjabrahams/tensorflow-on-raspberry-pi/raw/master/bin/tensorflow-0.9.0-py3-none-any.whl

[wtchang13]

@samjabrahams it seems to be working now, thanks!

[samjabrahams]

@eyewy any status updates on your end?

I get the following NumPy version error:
RuntimeError: module compiled against API version 0xa but this version of numpy is 0x9

[samjabrahams]

Is this after uninstalling numpy and then reinstalling TensorFlow (or using virtualenv)?

[samjabrahams]

Closing this issue for now- @eyewy let me know if you're still experiencing issues.