Make binary CNN match between Keras and hls4ml

[jmitrevs]

Description

The check on the pytest for binary CNNs had to be disabled because it failed. This enables it, and attempts to fix issues that it discovers

Type of change

• [x] Bug fix (non-breaking change that fixes an issue)

Tests

The test checking has been enabled. It still fails in most cases.

Checklist

• [x] I have read the guidelines for contributing.
• [x] I have commented my code, particularly in hard-to-understand areas.
• [x] I have made corresponding changes to the documentation.
• [x] My changes generate no new warnings.
• [x] I have installed and run pre-commit on the files I edited or added.
• [x] I have added tests that prove my fix is effective or that my feature works.

[jmitrevs]

This now matches for io_parallel implementations with the Latency strategy. Resource still fails (potentially for unrelated reasons also on Quartus) and for io_stream. @vloncar , I am curious if you think there's a better way to preserve the XnorPrecision designation when doing MaxPooling than I am doing here. It is a bit related to precision propagation, but I think in this case you really need to preserve the type for the network to make sense.

[jmitrevs]

With the latest push, this works for Vitis and Vivado (but not Quartus). Note that #817 was triggered with the test previously, which has been modified to not trigger the problem. However, a better fix for #817 would be preferred. It is somewhat beyond the scope of this PR, though.

[jmitrevs]

The way Winograd is written it doesn't support the xnor types, so I have forced the Quartus io_parallel tests to use im2col. If we want to update Winograd to do binary, I'll leave that as a separate development.

[jmitrevs]

Don't fully understand why the pytests fail here but succeeded on my laptop. I thought I was using the same random number, but maybe not.

[jmitrevs]

I think the issue was just accumulator precision in the test (big number - almost equal big number) and then quantization based on > or < 0. Let's see if increasing the bitwidth makes the tests succeed.

[bo3z]

The way Winograd is written it doesn't support the xnor types, so I have forced the Quartus io_parallel tests to use im2col. If we want to update Winograd to do binary, I'll leave that as a separate development.

Enabling binary support for Winograd needs to change (i) the explicit multiplication for binary weights, to use the product function from nnet_mult and (ii) for binary inputs, see how the addition of input elements behaves. But Winograd's algorithm is really meant to reduce the number of multiplications/DSP utilisation and with binary CNNs, that's not really a concern, so I don't think it is necessarily important in this case.

[jmitrevs]

I see there are some errors. Will make it draft till I fix them.

[jmitrevs]

The failures seem to be due to precision of the quantization, so I am removing the draft designation

[bo3z]

Looks good to me.

[jmitrevs]

The test passed, so I think we can maybe merge this.