discuss-gnuradio
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Discuss-gnuradio] gr-inspector example `amc_cnn` fails


From: Ben Hilburn
Subject: Re: [Discuss-gnuradio] gr-inspector example `amc_cnn` fails
Date: Mon, 9 Jan 2017 17:53:48 -0500

Hey Chris -

What is the `TF AMC Model` block expecting, exactly, in the "Graph File" field? Based on the earlier discussion, my understanding is that the block needs not just the graph file, but also the weights.

A different way to ask the question, perhaps, is how are you writing your trained models to disk so that you can then use them from within this block?

Cheers,
Ben

On Sat, Jan 7, 2017 at 5:34 AM, Christopher Richardson <address@hidden> wrote:
Hi Ben,

Yeah with respect to the shifts, that was just to do with the Python flowgraph where I generated the training data for the CNN model in my
cnn_generate.py script.  I notice the generation code for RadioML is now available https://github.com/radioML/dataset :)

I'm currently trying to find out why Keras & TensorFlow Serving, no longer seem to be saving the model in the same way as before, as I get an error message when
loading the model via the amc_cnn graph:

"""
handler caught exception: Attempting to use uninitialized value convolution2d_1_W_1
     [[Node: convolution2d_1_W_1/read = Identity[T=DT_FLOAT, _class=["loc:@convolution2d_1_W_1"], _device="/job:localhost/replica:0/task:0/cpu:0"](convolution2d_1_W_1)]]

Caused by op u'convolution2d_1_W_1/read', defined at:
"""

Yeah I think that's a good idea adding notes in the flow graph.  I'll just try to solve the loading issue, then I'll do that.

Cheers

Chris




On Sat, Jan 7, 2017 at 2:14 AM, Ben Hilburn <address@hidden> wrote:
Hey Chris -

Thanks for the guidance! With Tim O'Shea's help, I was able to create & train the TensorFlow model in the RadioML repository. I noticed that you made an update to the README steering folks away from the CNN model, for the moment:

Please use the FAM model at the moment, I'm just tweaking the CNN model, as I need to implement 64 sample shifts.

I'm assuming that refers to the CNN model in your repository (https://github.com/chrisruk/models), and not CNN models, generally? I'm going to give the `amc_cnn` flowgraph another shot with my newly trained CNN as soon as I get the chance.

Once I get everything up and running and have a better understanding of how it works, I'd like to create some more documentation about how to get this up and running. Your (and Sebastian's) work in `gr-inspector` is not only really impressive but also timely & very useful, and I want to make it accessible to as many people as possible. Even some small notes in the flowgraph about needing a trained model would be helpful, as right now this example relies on the user just knowing that's a necessary prerequisite.

I'll keep you posted on my progress =)

Cheers,
Ben



On Fri, Jan 6, 2017 at 5:24 AM, Christopher Richardson <address@hidden> wrote:
Hi Ben,

Which I'd not seen before, I'll update my CNN model generation to make sure it creates the model in the same way as that. 

(Sorry if you get this mail twice Ben, I forgot to hit reply to all first time)

Cheers
Chris

On Fri, Jan 6, 2017 at 12:21 AM, Christopher Richardson <address@hidden> wrote:
Hi Ben

I think it looks like you haven't generated the model, if I'm understanding correctly, for the GRC file.

At the moment the code to generate the models is at:
https://github.com/chrisruk/models

Tomorrow I'll check that the model generation works with the latest version of TensorFlow for you.

If anyone has experience with Keras, I'd love to have someone have a look over of the correctness of the models,
as I seem to have lower accuracy with the Keras models, than when they where implemented with TFLearn if I recall
correctly, which is very odd.

Cheers
Chris


On Thu, Jan 5, 2017 at 9:45 PM, Ben Hilburn <address@hidden> wrote:
Hey all -

I'm trying to get some of the tensorflow gr-inspector examples up and running, and running into a few hiccups (note that the non-tensorflow flowgraps work fine). I'm using the non-GPU version of tensorflow (`$ sudo pip install tensorflow`) as I don't have an nVidia GPU.

With the `amc_cnn` example, I was first getting the following error:

RuntimeError: Expected meta graph file missing /tmp/cnn/00000001/export.meta

Not really being sure what that was, I created the `/tmp/cnn/00000001/` directory, touched the `export.meta` file, and moved on (this may have been my first mistake).

Anyway, it gets a bit further, but dies with this:

Executing: /usr/bin/python2 -u /home/bhilburn/code/gnuradio/gr-inspector.git/examples/top_block.py

Traceback (most recent call last):
  File "/home/bhilburn/code/gnuradio/gr-inspector.git/examples/top_block.py", line 139, in <module>
    main()
  File "/home/bhilburn/code/gnuradio/gr-inspector.git/examples/top_block.py", line 127, in main
    tb = top_block_cls()
  File "/home/bhilburn/code/gnuradio/gr-inspector.git/examples/top_block.py", line 75, in __init__
    self.inspector_tfmodel_vcf_0 = inspector.tfmodel_vcf("complex64",128,'/tmp/cnn/00000001',(),0)
  File "/home/bhilburn/usr/lib64/python2.7/site-packages/inspector/tfmodel_vcf.py", line 64, in __init__
    sess, inp, out,classes = self.load_graph(graphfile)
  File "/home/bhilburn/usr/lib64/python2.7/site-packages/inspector/tfmodel_vcf.py", line 88, in load_graph
    signatures_any[0].Unpack(signatures)
  File "/usr/lib/python2.7/site-packages/google/protobuf/internal/containers.py", line 204, in __getitem__
    return self._values[key]
IndexError: list index out of range

Any ideas / pointers?

Cheers,
Ben






reply via email to

[Prev in Thread] Current Thread [Next in Thread]