`GRU` object has no attribute `get_quantizers` Error when using model_save_quantized_weights

[jay1601]

Objective

I'm using QKeras to apply Quantization-Aware Training to my TensorFlow model in order to deploy it on an embedded device. I use the model_quantize function to quantize my pre-defined model.

Issue Encountered:

After training, I attempted to extract the quantized weights to verify that all layers were properly quantized. However, I encountered an error when calling the function model_save_quantized_weights: GRU object has no attribute get_quantizers.

I see that QKeras supported QGRU, but why it raise that errors? I have thought that because I wrap in the Bidirectional layer but as in QRNNTutorial, they also can wrap LSTM in a Bidirectional and do the quantization with the same config.

Code Snippets:

Model Definition:

def create_model(signal_shape):
    input_signal = tf.keras.Input(shape=signal_shape, dtype=tf.float64)
    ecg_offset = tf.keras.Input(shape=signal_shape, dtype=tf.float64)

    inputs = {
        'signal': input_signal,
        'ecg_offset': ecg_offset
    }

    signal = inputs['signal']
    signal = tf.transpose(signal, perm=(0,2,1))

    conv1 = tf.keras.layers.Conv1D(16,3,activation='relu', padding='same')(signal)
    conv2 = tf.keras.layers.Conv1D(32,3,activation='relu', padding='same')(tf.keras.layers.MaxPooling1D(pool_size=2)(conv1))
    bi_gru = tf.keras.layers.Bidirectional(tf.keras.layers.GRU(64, return_sequences=True), merge_mode='sum')(tf.keras.layers.MaxPooling1D(pool_size=2)(conv2))

    conv3 = tf.keras.layers.Conv1D(32,3,activation='relu', padding='same')(bi_gru)
    dense = tf.keras.layers.Dense(10)(conv3)

    outputs = {
        'prob:': tf.transpose(tf.keras.layers.Activation("softmax", name="softmax")(dense), perm=(0,2,1))
    }

    model = tf.keras.Model(inputs=inputs, outputs=outputs)
    return model

Quantizer Configuration (as suggested by the QRNNTutorial):

bits = 4
quantizer_config = {
  "bidirectional": {
      'activation' : f"quantized_tanh({bits})",
      'recurrent_activation' : f"quantized_relu(4,0,1)",
      'kernel_quantizer' : f"quantized_bits({bits}, alpha='auto')",
      'recurrent_quantizer' : f"quantized_bits({bits}, alpha='auto')",
      'bias_quantizer' : f"quantized_bits({bits}, alpha='auto')",
  },
  "dense": {
      'kernel_quantizer' : f"quantized_bits({bits}), alpha='auto'",
      'bias_quantizer' : f"quantized_bits({bits}), alpha='auto'"
  },
"conv1d": {
      'kernel_quantizer' : f"quantized_bits({bits}), alpha='auto'",
      'bias_quantizer' : f"quantized_bits({bits}), alpha='auto'"
  },

Bonus Question:

• Since I'm using tf.transpose (which has no trainable weights), do I need to explicitly declare it in custom_objects when calling model_quantize? I see that tf.transpose appears as a Lambda layer in the model summary, but I haven't added it to custom_objects so far.
• When I use activation='relu' directly in Conv1D, does model_quantize automatically replace it with quantized_relu ? Or do I need to refactor the code to apply the activation in a separate layer, or explicitly include in QConv1D in the config dictionary?
• If I write a custom Quantization layer for my tf custom layer (QCustomLayer for CustomLayer for example), how can I define and use it with Qkeras? Is it done using _add_supported_quantized_objects(custom_objects)?

Thank you so much