QKeras support for RNN layers

[laurilaatu]

Description

:memo: This pull request adds QKeras support for RNN layers

The quantizers for kernel, recurrent kernel and bias are retrieved from the QKeras model. State quantizer corresponds to the output precision of the hidden state which is set by altering the default precision (discussed in issue #825). Currently the quantized activations for the RNN layers are not supported.

Type of change

• [x] New feature (non-breaking change which adds functionality)

Tests

:memo: Please describe the tests that you ran to verify your changes.

• Tested with several different network configurations
• Pytests implemented
• All layers produce expected output except for the Vivado implementation of the SimpleRNN (excluded from pytests)

Test Configuration: Quartus 21.1.0.169.pro

Example usage:

qmodel = Sequential()
qmodel.add(QSimpleRNN(4, input_shape=(5,1), name='rnn',
                      kernel_quantizer=quantized_bits(bits=9, integer=0, symmetric=False, alpha=1.0),
                      recurrent_quantizer=quantized_bits(bits=9, integer=0, symmetric=False, alpha=1.0),
                      bias_quantizer=quantized_bits(bits=9, integer=0, symmetric=False, alpha=1.0),
                      state_quantizer=quantized_bits(bits=9, integer=0, symmetric=False, alpha=1.0),
                      activation='relu'))
qmodel.summary()
qmodel.compile(loss='mse', optimizer='adam')

config = hls4ml.utils.config_from_keras_model(qmodel,
                                           default_precision="ap_fixed<9,1>",
                                           granularity='name')

hls_model = hls4ml.converters.convert_from_keras_model(qmodel,
                                           hls_config=config,
                                           output_dir='model_qrnn/hls4ml_prj',
                                           part='AGFB014R24A2E2VR0',
                                           backend='Quartus')
hls_model.compile()

Checklist

• [x] I have read the guidelines for contributing.
• [x] I have commented my code, particularly in hard-to-understand areas.
• [ ] 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]

@laurilaatu , is it possible for you to fix the conflicts so we can merge this PR?

[laurilaatu]

@jmitrevs , the conflict is now resolved.

[vloncar]

The newly added test fails with:

FAILED test_qkeras.py::test_qsimplernn[Quartus] - AssertionError: 
Not equal to tolerance rtol=1e-07, atol=0.1
Mismatched elements: 1 / 4 (25%)
Max absolute difference: 0.953125
Max relative difference: 1.
 x: array([[0.      , 1.004456, 0.024292, 0.352661]], dtype=float32)
 y: array([[0.953125, 0.921875, 0.015625, 0.351562]])

@laurilaatu can you investigate?

[laurilaatu]

@vloncar it seems that the test fails occasionally due to the random initialization. The latest commit fixes the issue by setting the seed. The issue does not affect QLSTM or QGRU.

Please let me know if any other changes are required.