Error with nb 00_load_data

SwiftAI
1

I’m just trying to run the NBs to get my head around things, and have hit an error pretty quickly.

As of about 20 minutes ago (~2019-04-22 1140 PDT) I’ve

  • Reinstalled the swift-tensorflow-DEVELOPMENT package
  • git pulled and reinstalled the jupyter kernel
  • git pulled the fastai_docs repo

Running the notebook 00_load_data is fine until cell 11 when I get this:

error: <Cell 12>:7:40: error: cannot convert value of type 'Int' to expected element type 'Int32'
    let shape: TensorShape = labels ? [batch] : (flat ? [batch, 784] : [batch, 28, 28])
                                       ^~~~~
                                       Int32( )

This is bringing back the utter bafflement of new error types in a new language :slight_smile: I’m not sure when it’s trying to convert an Int to an Int32. Is this PEBCAK or a bug?

Edit: Changing line 6 from:

    let batch = training ? 60000 : 10000

to

    let batch = training ? Int32(60000) : Int32(10000)

Has let me proceed. Perhaps this a “teachable moment” waiting to be explained in the lesson?

2

I think you don’t have properly registered your new build with jupyter. Just updated and 00 is running without any issue here.

3

Thanks! OK well that’s probably good. I used

python register.py --sys-prefix --swift-python-use-conda --use-conda-shared-libs --swift-toolchain ~/swift

In my swift-jupyter directory to register the kernel. Something changed because before that, I was getting a different error (complaining about the %install magic lines), and those went away after I did that. Is there anything else I should check?

4

Converting from Int32 to Int in general happened a few days ago: https://github.com/tensorflow/swift-apis/commit/23c16ae33a3826399b01caeb1b0b736531d00bde

Then, you’re just seeing side-effects from this. Use a nightly for now, or wait for a new official release and after that things should be fine!

5

Thanks Brett, is there a difference between a nightly and the swift-tensorflow-DEVELOPMENT-* bundle? Because I thought I was using that.

Oh from looking at the commit I’ll try installing a TensorFlow nightly as well, if that’s what you mean. I’ll give it a shot and report back. (Edit: no that’s the wrong tree to be barking up).

OK I’m happy to consider this a me problem. Hopefully v0.3 will be released any minute now and this’ll be fine. :slight_smile:

6

I think that installing the nightly build of S4TF should suffice. Perhaps jupyter is using a different version for some reason?

7

:man_facepalming: so it helps if you fg before you ctrl+c to restart jupyter…

…Nothing to see here, move along :face_with_head_bandage:

8

I think your version of swift is fine then? Registering the kernel here and using a nightly, runs beginning to end for me. For what it’s worth, wrapping everything in Int32() casts is the correct way of dealing with this error, you’re not doing anything wrong there.

9

That is good to know that the casting is right, thanks. But it was just user error, I hadn’t restarted jupyter after all. Cheers for the help!