tensorflow
AttributeError: 'CustomLayerMaxPooling1D' object has no attribute 'kernel'
Hi, when I am trying Quantization Aware Training on my model, I get the following error in my 'CustomLayerMaxPooling1D' :
AttributeError Traceback (most recent call last)
Cell In [62], line 9
1 # quantize_apply requires mentioning MyLSTMQuantizeConfig with quantize_scope
2 # as well as the custom Keras layer.
3 with tfmot.quantization.keras.quantize_scope({'MyLSTMQuantizeConfig': MyLSTMQuantizeConfig, 'CustomLayerLSTM': CustomLayerLSTM,
4 'MyConv1DQuantizeConfig': MyConv1DQuantizeConfig,
5 'CustomLayerConv1D': CustomLayerConv1D,
6 'MyMaxPooling1DQuantizeConfig': MyMaxPooling1DQuantizeConfig,
7 'CustomLayerMaxPooling1D': CustomLayerMaxPooling1D}):
8 # Use quantize_apply to actually make the model quantization aware.
----> 9 quant_aware_model = tfmot.quantization.keras.quantize_apply(new_model)
11 quant_aware_model.summary()
File ~\AppData\Roaming\Python\Python39\site-packages\tensorflow_model_optimization\python\core\keras\metrics.py:74, in MonitorBoolGauge.call.
File ~\AppData\Roaming\Python\Python39\site-packages\tensorflow_model_optimization\python\core\keras\metrics.py:69, in MonitorBoolGauge.call.
File ~\AppData\Roaming\Python\Python39\site-packages\tensorflow_model_optimization\python\core\quantization\keras\quantize.py:496, in quantize_apply(model, scheme, quantized_layer_name_prefix)
490 quantize_registry = scheme.get_quantize_registry()
492 # 4. Actually quantize all the relevant layers in the model. This is done by
493 # wrapping the layers with QuantizeWrapper, and passing the associated
494 # QuantizeConfig.
--> 496 return keras.models.clone_model(
497 transformed_model, input_tensors=None, clone_function=_quantize)
File ~\anaconda3\lib\site-packages\keras\models\cloning.py:502, in clone_model(model, input_tensors, clone_function) 499 clone_function = _clone_layer 501 if isinstance(model, Sequential): --> 502 return _clone_sequential_model( 503 model, input_tensors=input_tensors, layer_fn=clone_function 504 ) 505 else: 506 return _clone_functional_model( 507 model, input_tensors=input_tensors, layer_fn=clone_function 508 )
File ~\anaconda3\lib\site-packages\keras\models\cloning.py:372, in _clone_sequential_model(model, input_tensors, layer_fn)
367 layers, ancillary_layers = _remove_ancillary_layers(
368 model, layer_map, layers
369 )
371 if input_tensors is None:
--> 372 cloned_model = Sequential(layers=layers, name=model.name)
373 elif len(generic_utils.to_list(input_tensors)) != 1:
374 raise ValueError(
375 "To clone a Sequential model, we expect at most one tensor as "
376 f"part of input_tensors. Received: input_tensors={input_tensors}"
377 )
File ~\anaconda3\lib\site-packages\tensorflow\python\trackable\base.py:205, in no_automatic_dependency_tracking.
File ~\anaconda3\lib\site-packages\keras\utils\traceback_utils.py:70, in filter_traceback.tf.debugging.disable_traceback_filtering()
---> 70 raise e.with_traceback(filtered_tb) from None
71 finally:
72 del filtered_tb
File ~\AppData\Roaming\Python\Python39\site-packages\tensorflow_model_optimization\python\core\quantization\keras\quantize_wrapper.py:251, in QuantizeWrapperV2.build(self, input_shape) 249 def build(self, input_shape): 250 self._trainable_weights.extend(self.layer.trainable_weights) --> 251 super(QuantizeWrapperV2, self).build(input_shape)
File ~\AppData\Roaming\Python\Python39\site-packages\tensorflow_model_optimization\python\core\quantization\keras\quantize_wrapper.py:110, in QuantizeWrapper.build(self, input_shape) 102 self.optimizer_step = self.add_weight( 103 'optimizer_step', 104 initializer=tf.keras.initializers.Constant(-1), 105 dtype=tf.dtypes.int32, 106 trainable=False) 108 self._weight_vars = [] 109 for weight, quantizer in ( --> 110 self.quantize_config.get_weights_and_quantizers(self.layer)): 111 quantizer_vars = quantizer.build(weight.shape, 112 self._weight_name(weight.name), self) 114 self._weight_vars.append((weight, quantizer, quantizer_vars))
Cell In [48], line 6, in MyMaxPooling1DQuantizeConfig.get_weights_and_quantizers(self, layer) 5 def get_weights_and_quantizers(self, layer): ----> 6 return [(layer.kernel, LastValueQuantizer())]
AttributeError: 'CustomLayerMaxPooling1D' object has no attribute 'kernel'
The class MyMaxPooling1DQuantizeConfig is the following:
LastValueQuantizer = tfmot.quantization.keras.quantizers.LastValueQuantizer
MovingAverageQuantizer = tfmot.quantization.keras.quantizers.MovingAverageQuantizer
class MyMaxPooling1DQuantizeConfig(tfmot.quantization.keras.QuantizeConfig):
def get_weights_and_quantizers(self, layer):
return [(layer.kernel, LastValueQuantizer())]
def get_activations_and_quantizers(self, layer):
return [(layer.activation, MovingAverageQuantizer())]
def set_quantize_activations(self, layer, quantize_activations):
layer.activation = quantize_activations[0]
def set_quantize_weights(self, layer, quantize_weights):
layer.kernel = quantize_weights[0]
def get_config(self):
return {}
def get_output_quantizers(self, layer):
return []
The model along with the CustomLayerMaxPooling1D class is the following:
class CustomLayerMaxPooling1D(tf.keras.layers.MaxPooling1D):
pass
new_model = tfmot.quantization.keras.quantize_annotate_model(tf.keras.Sequential([tfmot.quantization.keras.quantize_annotate_layer(CustomLayerConv1D(64, kernel_size=3, activation = 'relu', input_shape=(120,51)), MyConv1DQuantizeConfig()),
tfmot.quantization.keras.quantize_annotate_layer(CustomLayerMaxPooling1D(pool_size=2), MyMaxPooling1DQuantizeConfig()),
tfmot.quantization.keras.quantize_annotate_layer(CustomLayerConv1D(64, kernel_size=3, activation = 'relu'), MyConv1DQuantizeConfig()),
tfmot.quantization.keras.quantize_annotate_layer(CustomLayerMaxPooling1D(pool_size=2), MyMaxPooling1DQuantizeConfig()),
tfmot.quantization.keras.quantize_annotate_layer(CustomLayerLSTM(32, return_sequences=False, activation='relu'), MyLSTMQuantizeConfig()),
tf.keras.layers.Dropout(0.2),
tf.keras.layers.Dense(64, activation='relu'),
tf.keras.layers.Dropout(0.2),
tf.keras.layers.Dense(actions.shape[0], activation='softmax')
]))
And this is how I quantize the model:
with tfmot.quantization.keras.quantize_scope({'MyLSTMQuantizeConfig': MyLSTMQuantizeConfig, 'CustomLayerLSTM':
CustomLayerLSTM, 'MyConv1DQuantizeConfig': MyConv1DQuantizeConfig, 'CustomLayerConv1D': CustomLayerConv1D, 'MyMaxPooling1DQuantizeConfig': MyMaxPooling1DQuantizeConfig, 'CustomLayerMaxPooling1D': CustomLayerMaxPooling1D}):
# Use `quantize_apply` to actually make the model quantization aware.
quant_aware_model = tfmot.quantization.keras.quantize_apply(new_model)
Apparently sth is wrong with the function def get_weights_and_quantizers in the class MyMaxPooling1DQuantizeConfig, but I am not aware on how to fix the problem. I tried replacing layer.kernel with [ ] to the return value of the function def get_weights_and_quantizers(self, layer) as well as replaced pass in place of layer.kernel = quantize_weights[0] in function def set_quantize_weights(self, layer, quantize_weights), but it returns ValueError: not enough values to unpack (expected 2, got 0).
new_model.summary() info:
Can you add how CustomLayerMaxPooling1D is implemented?
If it doesn't have kernel, than you have to change the MyMaxPooling1DQuantizeConfig as:
class MyMaxPooling1DQuantizeConfig(tfmot.quantization.keras.QuantizeConfig):
def get_weights_and_quantizers(self, layer):
return [] # No kernel.
# return [(layer.kernel, LastValueQuantizer())]
def get_activations_and_quantizers(self, layer):
return [(layer.activation, MovingAverageQuantizer())]
def set_quantize_activations(self, layer, quantize_activations):
layer.activation = quantize_activations[0]
def set_quantize_weights(self, layer, quantize_weights):
layer.kernel = quantize_weights[0]
def get_config(self):
return {}
def get_output_quantizers(self, layer):
return []
