Add LayerNorm support for Vivado

hls4ml/backends/vivado/passes/core_templates.py

@@ -1,6 +1,15 @@
 from hls4ml.backends.backend import get_backend
 from hls4ml.backends.template import FunctionCallTemplate, LayerConfigTemplate
-from hls4ml.model.layers import Activation, BatchNormalization, Dense, HardActivation, ParametrizedActivation, PReLU, Softmax
+from hls4ml.model.layers import (
+    Activation,
+    BatchNormalization,
+    Dense,
+    HardActivation,
+    LayerNormalization,
+    ParametrizedActivation,
+    PReLU,
+    Softmax,
+)
 from hls4ml.model.optimizer.passes.hgq_proxy_model import UnaryLUT
 
 # Dense templates
@@ -119,6 +128,59 @@ def format(self, node):
         return self.template.format(**params)
 
 
+# LayerNormalization templates
+
+layernorm_config_template = """struct config{index} : nnet::layernorm_config {{
+    static const unsigned n_in = {n_in};
+    static const unsigned seq_len = {seq_len};
+    static const unsigned table_size = {table_size};
+    static constexpr double table_range = {table_range};
+    static const unsigned io_type = nnet::{iotype};
+    static const unsigned reuse_factor = {reuse};
+    static const bool store_weights_in_bram = false;
+    static constexpr double epsilon = {epsilon};
+    typedef {bias_t.name} bias_t;
+    typedef {scale_t.name} scale_t;
+    typedef {mean_t.name} mean_t;
+    typedef {table_t.name} table_t;
+    template<class x_T, class y_T>
+    using product = nnet::product::{product_type}<x_T, y_T>;
+}};\n"""
+
+layernorm_function_template = 'nnet::layernormalize<{input_t}, {output_t}, {config}>({input}, {output}, {scale}, {bias});'
+
+layernorm_include_list = ['nnet_utils/nnet_layernorm.h']
+
+
+class LayerNormalizationConfigTemplate(LayerConfigTemplate):
+    def __init__(self):
+        super().__init__(LayerNormalization)
+        self.template = layernorm_config_template
+
+    def format(self, node):
+        params = self._default_config_params(node)
+        params['n_in'] = node.get_input_variable().size_cpp()
+        params['seq_len'] = node.get_attr('seq_len')
+        params['product_type'] = get_backend('vivado').product_type(
+            node.get_input_variable().type.precision, node.get_weights('scale').type.precision
+        )
+
+        return self.template.format(**params)
+
+
+class LayerNormalizationFunctionTemplate(FunctionCallTemplate):
+    def __init__(self):
+        super().__init__(LayerNormalization, include_header=layernorm_include_list)
+        self.template = layernorm_function_template
+
+    def format(self, node):
+        params = self._default_function_params(node)
+        params['scale'] = node.get_weights('scale').name
+        params['bias'] = node.get_weights('bias').name
+
+        return self.template.format(**params)
+
+
 # Activation templates
 
 activ_config_template = """struct {type}_config{index} : nnet::activ_config {{

hls4ml/backends/vivado/vivado_backend.py

@@ -21,6 +21,7 @@
     GarNet,
     GarNetStack,
     Layer,
+    LayerNormalization,
     Pooling1D,
     Pooling2D,
     SeparableConv1D,
@@ -558,6 +559,21 @@ def init_softmax(self, layer):
                 len(layer.get_input_variable().shape) == 1
             ), 'Softmax with io_parallel strategy cannot be used on multidimensional tensors.'
 
+    @layer_optimizer(LayerNormalization)
+    def init_layernormalization(self, layer):
+        if 'table_t' not in layer.attributes:
+            layer.set_attr(
+                'table_t', NamedType(name=layer.name + '_table_t', precision=FixedPrecisionType(width=16, integer=6))
+            )
+        if 'table_size' not in layer.attributes:
+            layer.set_attr('table_size', 4096)  # table size
+        if 'table_range' not in layer.attributes:
+            layer.set_attr('table_range', 1.0)  # table range
+        if 'mean_t' not in layer.attributes:
+            layer.set_attr(
+                'mean_t', NamedType(name=layer.name + '_mean_t', precision=FixedPrecisionType(width=19, integer=6))
+            )
+
     @layer_optimizer(Embedding)
     def init_embed(self, layer):
         if layer.attributes['n_in'] is None:

hls4ml/converters/keras/core.py

@@ -129,6 +129,34 @@ def parse_batchnorm_layer(keras_layer, input_names, input_shapes, data_reader):
     return layer, [shape for shape in input_shapes[0]]
 
 
+@keras_handler('LayerNormalization')
+def parse_layernorm_layer(keras_layer, input_names, input_shapes, data_reader):
+    assert 'LayerNormalization' in keras_layer['class_name']
+
+    layer = parse_default_keras_layer(keras_layer, input_names)
+
+    in_size = 1
+    for dim in input_shapes[0][1:]:
+        in_size *= dim
+    layer['n_in'] = layer['n_out'] = in_size
+
+    if not ((len(input_shapes[0])) == 3):
+        raise Exception('input size is not currently supported by hls4ml, only dim3 is supported')
+    layer['seq_len'] = input_shapes[0][-2]
+
+    if not (keras_layer['config']['axis'][0] == 2):
+        raise Exception('assigning the axis is not currently supported by hls4ml, only axis 2 is supported')
+
+    layer['gamma_data'] = get_weights_data(data_reader, layer['name'], 'gamma')
+    layer['beta_data'] = get_weights_data(data_reader, layer['name'], 'beta')
+
+    layer['epsilon'] = keras_layer['config']['epsilon']
+    if layer['epsilon'] <= 0:
+        raise Exception('epsilon must be positive')
+
+    return layer, [shape for shape in input_shapes[0]]
+
+
 @keras_handler('Embedding')
 def parse_embedding_layer(keras_layer, input_names, input_shapes, data_reader):
     assert 'Embedding' in keras_layer['class_name']

hls4ml/converters/pytorch/core.py

@@ -138,3 +138,32 @@ def parse_batchnorm_layer(operation, layer_name, input_names, input_shapes, node
         layer['n_filt'] = input_shapes[0][1]  # Always channel first for Pytorch
 
     return layer, [shape for shape in input_shapes[0]]
+
+
+@pytorch_handler('LayerNorm')
+def parse_layernorm_layer(operation, layer_name, input_names, input_shapes, node, class_object, data_reader, config):
+    assert 'LayerNorm' in operation
+
+    layer = {}
+
+    layer['class_name'] = 'LayerNormalization'
+    layer['name'] = layer_name
+    layer['inputs'] = input_names
+
+    in_size = 1
+    for dim in input_shapes[0][1:]:
+        in_size *= dim
+    layer['n_in'] = layer['n_out'] = in_size
+
+    if not ((len(input_shapes[0])) == 3):
+        raise Exception('input size is not currently supported by hls4ml, only dim3 is supported')
+    layer['seq_len'] = input_shapes[0][-2]
+
+    layer['gamma_data'] = class_object.weight.data.numpy()
+    layer['beta_data'] = class_object.bias.data.numpy()
+
+    layer['epsilon'] = class_object.eps
+    if layer['epsilon'] <= 0:
+        raise Exception('epsilon must be positive')
+
+    return layer, [shape for shape in input_shapes[0]]

hls4ml/model/layers.py

@@ -1058,6 +1058,30 @@ def add_bias(self, bias, quantizer=None, precision=None):
         self.add_weights_variable(name='bias', var_name='b{index}', data=bias, quantizer=quantizer, precision=precision)
 
 
+class LayerNormalization(Layer):
+    _expected_attributes = [
+        Attribute('n_in'),
+        Attribute('seq_len'),
+        Attribute('epsilon', value_type=float, default=1e-3),
+        WeightAttribute('scale'),
+        WeightAttribute('bias'),
+        TypeAttribute('scale'),
+        TypeAttribute('bias'),
+    ]
+
+    def initialize(self):
+        inp = self.get_input_variable()
+        shape = inp.shape
+        dims = inp.dim_names
+        self.add_output_variable(shape, dims)
+
+        scale = self.get_attr('gamma_data')
+        bias = self.get_attr('beta_data')
+
+        self.add_weights_variable(name='scale', var_name='s{index}', data=scale)
+        self.add_weights_variable(name='bias', var_name='b{index}', data=bias)
+
+
 class Merge(Layer):
     def initialize(self):
         assert len(self.inputs) == 2
@@ -1682,6 +1706,7 @@ def initialize(self):
     'BatchNormOnnx': BatchNormOnnx,
     'LayerGroup': LayerGroup,
     'SymbolicExpression': SymbolicExpression,
+    'LayerNormalization': LayerNormalization,
     # TensorFlow-specific layers:
     'BiasAdd': BiasAdd,
 }

hls4ml/model/optimizer/passes/convert_to_channels_last.py

@@ -2,7 +2,7 @@
 # Based on https://github.com/fastmachinelearning/qonnx/blob/
 # 12c96a3ded06beacab08e0f554e4ed014476c0aa/src/qonnx/transformation/channels_last.py
 
-from hls4ml.model.layers import Concatenate, Dense, Input, Reshape, Transpose
+from hls4ml.model.layers import Concatenate, Dense, Input, LayerNormalization, Reshape, Transpose
 from hls4ml.model.optimizer import OptimizerPass
 from hls4ml.model.types import WeightVariable
 
@@ -45,6 +45,24 @@ def transform(self, model, node):
                 node.get_output_variable().shape = input_shape
                 dim_names = [f'N_INPUT_{i}_{node.index}' for i in range(1, len(input_shape) + 1)]
                 node.get_output_variable().dim_names = dim_names
+        elif isinstance(node, LayerNormalization):
+            # LayerNorm only works on the last dimension in PyTorch
+            perm = [1, 0]
+            pre_transpose = model.make_node(
+                'Transpose', f'pre_transpose_for_{node.get_attr("name")}', {'perm': perm}, [node.get_input_node().name]
+            )
+            pre_transpose.channels_last_converted = True
+            model.insert_node(pre_transpose)
+
+            # If not the output layer, transpose again
+            if not (
+                node.get_attr('name') in model.outputs and model.config.config['HLSConfig']['Model']['TransposeOutputs']
+            ):
+                post_transpose = model.make_node(
+                    'Transpose', f'post_transpose_for_{node.get_attr("name")}', {'perm': perm}, [node.name]
+                )
+                post_transpose.channels_last_converted = True
+                model.insert_node(post_transpose)
         else:
             # Transpose weight tensors
             tensors = ['weight', 'depthwise', 'pointwise', 'zero_bias', 'scale', 'recurrent_weight']

hls4ml/model/optimizer/passes/infer_precision.py

@@ -51,7 +51,7 @@ def _infer_precision(self, node, types_to_infer):
         if node_class in ['Dense']:
             return self._infer_dense_precision(node, types_to_infer)
 
-        if node_class in ['BatchNormalization', 'ApplyAlpha']:
+        if node_class in ['BatchNormalization', 'ApplyAlpha', 'LayerNormalization']:
             return self._infer_bn_precision(node, types_to_infer)
 
         if node_class in ['Conv1D', 'Conv2D', 'PointwiseConv1D', 'PointwiseConv2D', 'Conv2DBatchnorm']:

hls4ml/model/profiling.py

@@ -273,6 +273,18 @@ def _keras_layer(layer):
     return layer.get_weights(), ['w', 'b']
 
 
+def _keras_layernorm(layer):
+    weights = layer.get_weights()
+
+    gamma = weights[0]
+    beta = weights[1]
+
+    scale = gamma
+    bias = beta
+
+    return [scale, bias], ['s', 'b']
+
+
 def _keras_lstm(layer):
     return layer.get_weights(), ['w', 'u', 'b']
 
@@ -282,6 +294,7 @@ def _keras_lstm(layer):
     {
         'BatchNormalization': _keras_batchnorm,
         'QBatchNormalization': _keras_batchnorm,
+        'LayerNormalization': _keras_layernorm,
         'LSTM': _keras_lstm,
         'QLSTM': _keras_lstm,
     },