Weights export

[Sorry, the first time in GitHub] Hi, I used Netron to see and export the binarized parameters and I see that the parameters aren’t. In the attached file you can see what I mean. So, the parameters shouldn’t be only +1 and -1? Why is this happening?

The layer:

kwargs = dict(
    kernel_quantizer=lq.quantizers.SteSign(clip_value=1.0),
    kernel_constraint="weight_clip",
    padding="same"
)

activation = 'ste_sign'

x = tf.keras.Input(shape=(1, Time_series_length, 1))
ConvL1 = lq.layers.QuantConv2D(128, (1, 8), use_bias=False, **kwargs)(x)
ConvL1 = tf.keras.layers.BatchNormalization(scale=False)(ConvL1)
ConvL1 = tf.keras.layers.Activation(activation)(ConvL1)

In the image, you can see the parameters of the layer.

TensorFlow version: 2.0.0 Larq version: 0.8.2

[AlexRux]

For now, until our TFLite inference implementation is ready, you could apply this quantization rule yourself as preprocessing when you’re uploading the weights to the machine.

Oh, perfect. A Matlab script and I’m ready to go.

1. Use an optimizer like Bop which doesn’t rely on latent weights
2. Install larq from GitHub master and use the experimental quantization_scope API. Note that the details might change before the 0.9 release, but I’ll update this issue if the API changes.

Thank you both. I’ll try.

[lgeiger]

Great! Feel free to reopen if you have more questions.