LayerNormalization is not show as LayerNormalization in onnx graph

[Azulita0317]

Ask a Question

Question

The LayerNormalization was translate to some operators and BatchNormalization, when I use opset=17。
I hope it could be a unique operator rather than a series of operator list like this:

any parameter should I add?
here is my code :

`

class Test_Model(tf.keras.Model):
  def __init__(self):
      super(Test_Model, self).__init__()
      self.input_layer = tf.keras.layers.Input((1,256,1))
      self.conv1 = layers.Conv2D(128, (1, 3), activation='relu',padding='same')
      self.norm1 = layers.LayerNormalization(axis=-1)
      self.conv2 = layers.Conv2D(64, (1, 4), activation='relu',padding='same')
      self.norm2 = layers.LayerNormalization(axis=3)
 def call(self, inputs):
      x = self.conv1(inputs)
      x = self.norm1(x)
      x = self.conv2(x)
      x = self.norm2(x)
      return x

input_shape = [1, 1, 256, 1]
inputs = tf.random.uniform(input_shape)
model = Test_Model()
model(inputs)
model.summary() 

onnx_model, _ = tf2onnx.convert.from_keras(model,opset=18)
new_model = onnxoptimizer.optimize(onnx_model)
inferred_model = onnx.shape_inference.infer_shapes(new_model)
onnx_name = r'./models/operators/layer_norm_tf' + '.onnx'
onnx.save_model(inferred_model, onnx_name) 

`

[fatcat-z]

So far, LayerNormalization will be transferred to some tf nodes in tf graph by TensorFlow, tf2onnx doesn't have a pattern to identify this so we can't convert it to one node in ONNX graph.

[pwuertz]

#2183 (feature request)
#2250 (implementation attempt, deleted though)