Convolution operator do not work as expected about gorgonnx HOT 7 CLOSED

@chewxy helped a lot with this investigation.

He told me to add extra logging to extract some information of the VM of Gorgonia

// ...
   gorgonia.Let(g.ByName("Input3")[0], t)
    logger := log.New(os.Stderr, "", 0)
    machine := gorgonia.NewTapeMachine(g, gorgonia.WithLogger(logger), gorgonia.WithWatchlist())
    if err = machine.RunAll(); err != nil {
        log.Fatal(err)
    }
// ...

Thanks to the logs, we can notice that the execution is OK and that the operators are correctly applied; the error I mentioned in the first comment was actually just related to the format command that rounds the numbers.

We are now thinking that the Conv2D operator should be the origin of the problem.
@chewxy pointed out that the shape of the filter was "strange" ((8, 1, 5, 5)). He thinks that it should be ((1,8,5,5)).

The original implementation of the neural network is here.

I am now trying to isolate and test the Conv2D operator to see if the parameters I gave him are ok.

from gorgonnx.

The convolution operator is not working, probably a bug in the padding mechanism and probably a bug into Gorgonia:

go test -run=ConvOp
--- FAIL: TestConvOp (0.00s)
    conv_test.go:131:
                Error Trace:    conv_test.go:131
                Error:          Not equal:
                                expected: ⎡ 12   21   27   33   24⎤
                                ⎢ 33   54   63   72   51⎥
                                ⎢ 63   99  108  117   81⎥
                                ⎢ 93  144  153  162  111⎥
                                ⎣ 72  111  117  123   84⎦

                                actual  : ⎡  0    0    0    0    0⎤
                                ⎢  0    0    0    6   72⎥
                                ⎢129   51  108  165   39⎥
                                ⎢120  101   76  137   35⎥
                                ⎣126  129   53  118   69⎦

                                Diff:
                Test:           TestConvOp
                Messages:       Bad result for the convolution operator
    conv_test.go:132: ⎡  0    0    0    0    0⎤
        ⎢  0    0    0    6   72⎥
        ⎢129   51  108  165   39⎥
        ⎢120  101   76  137   35⎥
        ⎣126  129   53  118   69⎦

FAIL
exit status 1
FAIL    github.com/owulveryck/gorgonnx  0.017s

from gorgonnx.

The conv operator has been fixed in Gorgonia: gorgonia/gorgonia#233

The vendor directory has been updated in consequence

from gorgonnx.

Please keep this open. There is one fix for col2im (for backprop) that needs to go in first. Gimme about 2 hrs

from gorgonnx.

Actually I just finished a simple check. The col2im op looks ok (if a little aggressive)

from gorgonnx.

hm. still not working?

from gorgonnx.

The unit test of my implementation passes; I will close this issue.
I still don't have the good result at the very end of the computation of the MNIST graph, but now I can go to bug hunting again. Many thanks again for you help

from gorgonnx.