I have a toy example of CNN mnist digit keras tensorflow model , I have quantized the 2 Conv2D layers and 2 dense layers in it to 4-bit , now I want to access the weights of Conv2D layer. But if I try that using get_weights() it returns a list of np.array which is float 32bit , this can be because by default numpy stores weight in float 32, To access the weights of a QAT model we need to convert it to tf.lite , which I cant do in my case because of the annotated layers (4-bit quantized). Any suggestion on how can I properly access those 4-bit quantized weights ? this is my quantized model :
QAT_model = tfmot.quantization.keras.quantize_annotate_model( keras.Sequential([
tfmot.quantization.keras.quantize_annotate_layer( tf.keras.layers.Conv2D(6, activation='relu',kernel_size=(5, 5), input_shape= input_shape), DefaultDenseQuantizeConfig() ),
tf.keras.layers.MaxPool2D(pool_size=(2, 2)),
tf.keras.layers.Dropout(0.25),
tfmot.quantization.keras.quantize_annotate_layer( tf.keras.layers.Conv2D(10, activation='relu',kernel_size=(5, 5)), DefaultDenseQuantizeConfig() ),
tf.keras.layers.MaxPool2D(pool_size=(2, 2)),
tf.keras.layers.Dropout(0.25),
tf.keras.layers.Flatten(),
tfmot.quantization.keras.quantize_annotate_layer( tf.keras.layers.Dense(256, activation='softmax'), DefaultDenseQuantizeConfig() ),
tfmot.quantization.keras.quantize_annotate_layer( tf.keras.layers.Dense(10, activation='softmax'), DefaultDenseQuantizeConfig() )
]) )
with tfmot.quantization.keras.quantize_scope(
{'DefaultDenseQuantizeConfig': DefaultDenseQuantizeConfig}):
# Use `quantize_apply` to actually make the model quantization aware.
quantized_model = tfmot.quantization.keras.quantize_apply(QAT_model)
quantized_model.summary()
quantized_model.compile(optimizer='adam', # Good default optimizer to start with
loss= tensorflow.keras.losses.categorical_crossentropy , # how will we calculate our "error." Neural network aims to minimize loss.
metrics=['accuracy']) # what to track
quantized_model.fit(input_train, target_train, epochs=10)
val_loss, val_acc = quantized_model.evaluate(input_test,target_test)
print('test accuracy:', val_acc)
for layer in quantized_model.layers:
print(layer.name, layer)
print(quantized_model.layers[1].weights) #conv2d layer
