Merge pull request #750 from vloncar/ir_cleanup

Additional cleanup of the codebase

Author: jmitrevs

contrib/garnet.py

@@ -3,21 +3,18 @@
 """
 
 import tensorflow.keras as keras
+from qkeras import QActivation, QDense, ternary
 
 K = keras.backend
 
-try:
-    from qkeras import QActivation, QDense, ternary
 
-    class NamedQDense(QDense):
-        def add_weight(self, name=None, **kwargs):
-            return super().add_weight(name=f'{self.name}_{name}', **kwargs)
+class NamedQDense(QDense):
+    def add_weight(self, name=None, **kwargs):
+        return super().add_weight(name=f'{self.name}_{name}', **kwargs)
 
-    def ternary_1_05():
-        return ternary(alpha=1.0, threshold=0.5)
 
-except ImportError:
-    pass
+def ternary_1_05():
+    return ternary(alpha=1.0, threshold=0.5)
 
 
 # Hack keras Dense to propagate the layer name into saved weights

hls4ml/backends/quartus/passes/merge_templates.py

@@ -68,7 +68,7 @@ def __init__(self):
     def format(self, node):
         inp1 = node.get_input_variable(node.inputs[0])
         inp2 = node.get_input_variable(node.inputs[1])
-        params = node._default_config_params()
+        params = self._default_config_params(node)
         params['n_out'] = 1
         params['n_in'] = inp1.shape[0]
         params['product_type'] = get_backend('quartus').product_type(inp1.type.precision, inp2.type.precision)

hls4ml/backends/vivado/passes/merge_templates.py

@@ -65,7 +65,7 @@ def __init__(self):
     def format(self, node):
         inp1 = node.get_input_variable(node.inputs[0])
         inp2 = node.get_input_variable(node.inputs[1])
-        params = node._default_config_params()
+        params = self._default_config_params(node)
         params['n_out'] = 1
         params['n_in'] = inp1.shape[0]
         params['product_type'] = get_backend('vivado').product_type(inp1.type.precision, inp2.type.precision)

hls4ml/converters/__init__.py

@@ -64,7 +64,8 @@
                         elif model_type == 'onnx':
                             register_onnx_layer_handler(layer, func)
 
-        except ImportError:
+        except ImportError as err:
+            print(f'WARNING: Failed to import handlers from {module}: {err.msg}.')
             continue
 
 

hls4ml/converters/keras/core.py

@@ -1,7 +1,5 @@
-import numpy as np
-
 from hls4ml.converters.keras_to_hls import keras_handler, parse_default_keras_layer
-from hls4ml.model.types import IntegerPrecisionType, Quantizer
+from hls4ml.model.types import BinaryQuantizer, IntegerPrecisionType, TernaryQuantizer
 
 
 @keras_handler('InputLayer')
@@ -25,37 +23,6 @@ def parse_input_layer(keras_layer, input_names, input_shapes, data_reader):
     return layer, output_shape
 
 
-class BinaryQuantizer(Quantizer):
-    def __init__(self, bits=2):
-        if bits == 1:
-            hls_type = IntegerPrecisionType(width=1, signed=False)
-        elif bits == 2:
-            hls_type = IntegerPrecisionType(width=2)
-        else:
-            raise Exception(f'BinaryQuantizer suppots 1 or 2 bits, but called with bits={bits}')
-        super().__init__(bits, hls_type)
-
-    def __call__(self, data):
-        zeros = np.zeros_like(data)
-        ones = np.ones_like(data)
-        quant_data = data
-        if self.bits == 1:
-            quant_data = np.where(data > 0, ones, zeros).astype('int')
-        if self.bits == 2:
-            quant_data = np.where(data > 0, ones, -ones)
-        return quant_data
-
-
-class TernaryQuantizer(Quantizer):
-    def __init__(self):
-        super().__init__(2, IntegerPrecisionType(width=2))
-
-    def __call__(self, data):
-        zeros = np.zeros_like(data)
-        ones = np.ones_like(data)
-        return np.where(data > 0.5, ones, np.where(data <= -0.5, -ones, zeros))
-
-
 dense_layers = ['Dense', 'BinaryDense', 'TernaryDense']
 
 

hls4ml/converters/keras/qkeras.py

@@ -1,81 +1,9 @@
-import tensorflow as tf
 from qkeras.quantizers import get_quantizer
 
-from hls4ml.converters.keras.core import BinaryQuantizer
-from hls4ml.model.types import ExponentPrecisionType, FixedPrecisionType, IntegerPrecisionType, Quantizer, XnorPrecisionType
-
-
-class QKerasQuantizer(Quantizer):
-    def __init__(self, config):
-        self.quantizer_fn = get_quantizer(config)
-        self.alpha = config['config'].get('alpha', None)
-        if config['class_name'] == 'quantized_bits':
-            self.bits = config['config']['bits']
-            self.hls_type = get_type(config)
-        # ! includes stochastic_ternary
-        elif 'ternary' in config['class_name']:
-            self.bits = 2
-            self.hls_type = IntegerPrecisionType(width=2, signed=True)
-        # ! includes stochastic_binary
-        elif 'binary' in config['class_name']:
-            self.bits = 1
-            self.hls_type = XnorPrecisionType()
-        else:
-            print("Unsupported quantizer: " + config['class_name'])
-            self.bits = 16
-            self.hls_type = FixedPrecisionType(width=16, integer=6, signed=True)
-
-    def __call__(self, data):
-        tf_data = tf.convert_to_tensor(data)
-        return self.quantizer_fn(tf_data).numpy()
-        # return self.quantizer_fn(data)
-
-
-class QKerasBinaryQuantizer:
-    def __init__(self, config, xnor=False):
-        self.bits = 1 if xnor else 2
-        self.hls_type = XnorPrecisionType() if xnor else IntegerPrecisionType(width=2, signed=True)
-        self.alpha = config['config']['alpha']
-        # Use the QKeras quantizer to handle any stochastic / alpha stuff
-        self.quantizer_fn = get_quantizer(config)
-        # Then we use our BinaryQuantizer to convert to '0,1' format
-        self.binary_quantizer = BinaryQuantizer(1) if xnor else BinaryQuantizer(2)
-
-    def __call__(self, data):
-        x = tf.convert_to_tensor(data)
-        y = self.quantizer_fn(x).numpy()
-        return self.binary_quantizer(y)
-
-
-class QKerasPO2Quantizer:
-    def __init__(self, config):
-        self.bits = config['config']['bits']
-        self.quantizer_fn = get_quantizer(config)
-        self.hls_type = ExponentPrecisionType(width=self.bits, signed=True)
-
-    def __call__(self, data):
-        '''
-        Weights are quantized to nearest power of two
-        '''
-        x = tf.convert_to_tensor(data)
-        y = self.quantizer_fn(x)
-        if hasattr(y, 'numpy'):
-            y = y.numpy()
-        return y
-
-
-def get_type(quantizer_config):
-    width = quantizer_config['config']['bits']
-    integer = quantizer_config['config'].get('integer', 0)
-    if quantizer_config['class_name'] == 'quantized_po2':
-        return ExponentPrecisionType(width=width, signed=True)
-    if width == integer:
-        if width == 1:
-            return XnorPrecisionType()
-        else:
-            return IntegerPrecisionType(width=width, signed=True)
-    else:
-        return FixedPrecisionType(width=width, integer=integer + 1, signed=True)
+from hls4ml.converters.keras.convolution import parse_conv1d_layer, parse_conv2d_layer
+from hls4ml.converters.keras.core import parse_batchnorm_layer, parse_dense_layer
+from hls4ml.converters.keras_to_hls import keras_handler, parse_default_keras_layer
+from hls4ml.model.types import FixedPrecisionType, QKerasBinaryQuantizer, QKerasPO2Quantizer, QKerasQuantizer
 
 
 def get_quantizer_from_config(keras_layer, quantizer_var):
@@ -88,3 +16,130 @@ def get_quantizer_from_config(keras_layer, quantizer_var):
         return QKerasPO2Quantizer(quantizer_config)
     else:
         return QKerasQuantizer(quantizer_config)
+
+
+@keras_handler('QDense')
+def parse_qdense_layer(keras_layer, input_names, input_shapes, data_reader):
+
+    layer, output_shape = parse_dense_layer(keras_layer, input_names, input_shapes, data_reader)
+
+    layer['weight_quantizer'] = get_quantizer_from_config(keras_layer, 'kernel')
+    if keras_layer['config']['bias_quantizer'] is not None:
+        layer['bias_quantizer'] = get_quantizer_from_config(keras_layer, 'bias')
+    else:
+        layer['bias_quantizer'] = None
+
+    return layer, output_shape
+
+
+@keras_handler('QConv1D', 'QConv2D')
+def parse_qconv_layer(keras_layer, input_names, input_shapes, data_reader):
+    assert 'QConv' in keras_layer['class_name']
+
+    if '1D' in keras_layer['class_name']:
+        layer, output_shape = parse_conv1d_layer(keras_layer, input_names, input_shapes, data_reader)
+    elif '2D' in keras_layer['class_name']:
+        layer, output_shape = parse_conv2d_layer(keras_layer, input_names, input_shapes, data_reader)
+
+    layer['weight_quantizer'] = get_quantizer_from_config(keras_layer, 'kernel')
+    if keras_layer['config']['bias_quantizer'] is not None:
+        layer['bias_quantizer'] = get_quantizer_from_config(keras_layer, 'bias')
+    else:
+        layer['bias_quantizer'] = None
+
+    return layer, output_shape
+
+
+@keras_handler('QActivation')
+def parse_qactivation_layer(keras_layer, input_names, input_shapes, data_reader):
+    assert keras_layer['class_name'] == 'QActivation'
+    supported_activations = [
+        'quantized_relu',
+        'quantized_tanh',
+        'binary_tanh',
+        'ternary_tanh',
+        'quantized_sigmoid',
+        'quantized_bits',
+        'binary',
+        'ternary',
+    ]
+
+    layer = parse_default_keras_layer(keras_layer, input_names)
+
+    activation_config = keras_layer['config']['activation']
+    quantizer_obj = get_quantizer(activation_config)
+    activation_config = {}
+    # some activations are classes
+    if hasattr(quantizer_obj, 'get_config'):
+        activation_config['class_name'] = quantizer_obj.__class__.__name__
+        if activation_config['class_name'] == 'ternary' or activation_config['class_name'] == 'binary':
+            activation_config['class_name'] += '_tanh'
+        activation_config['config'] = quantizer_obj.get_config()
+    # some activation quantizers are just functions with no config
+    else:
+        activation_config['config'] = {}
+        if 'binary' in quantizer_obj.__name__:
+            activation_config['class_name'] = 'binary_tanh'
+            activation_config['config']['bits'] = 1
+            activation_config['config']['integer'] = 1
+        elif 'ternary' in quantizer_obj.__name__:
+            activation_config['class_name'] = 'ternary_tanh'
+            activation_config['config']['bits'] = 2
+            activation_config['config']['integer'] = 2
+        else:
+            activation_config['class_name'] = 'unknown'
+
+    if activation_config['class_name'] not in supported_activations:
+        raise Exception('Unsupported QKeras activation: {}'.format(activation_config['class_name']))
+
+    if activation_config['class_name'] == 'quantized_bits':
+        activation_config['class_name'] = 'linear'
+
+    if activation_config['class_name'] == 'ternary_tanh':
+        layer['class_name'] = 'TernaryTanh'
+        layer['threshold'] = activation_config.get('config', {}).get('threshold', 0.33)
+        if layer['threshold'] is None:
+            layer['threshold'] = 0.33  # the default ternary tanh threshold for QKeras
+        layer['activation'] = 'ternary_tanh'
+    elif (
+        activation_config['class_name'] == 'quantized_sigmoid'
+        and not activation_config['config'].get('use_real_sigmoid', False)
+    ) or (
+        activation_config['class_name'] == 'quantized_tanh' and not activation_config['config'].get('use_real_tanh', False)
+    ):
+        layer['class_name'] = 'HardActivation'
+        layer['slope'] = 0.5  # the default values in QKeras
+        layer['shift'] = 0.5
+        # Quartus seems to have trouble if the width is 1.
+        layer['slope_prec'] = FixedPrecisionType(width=2, integer=0, signed=False)
+        layer['shift_prec'] = FixedPrecisionType(width=2, integer=0, signed=False)
+        layer['activation'] = activation_config['class_name'].replace('quantized_', 'hard_')
+    else:
+        layer['class_name'] = 'Activation'
+        layer['activation'] = activation_config['class_name'].replace('quantized_', '')
+
+    layer['activation_quantizer'] = activation_config
+    return layer, [shape for shape in input_shapes[0]]
+
+
+@keras_handler('QBatchNormalization')
+def parse_qbatchnorm_layer(keras_layer, input_names, input_shapes, data_reader):
+
+    layer, output_shape = parse_batchnorm_layer(keras_layer, input_names, input_shapes, data_reader)
+
+    layer['mean_quantizer'] = get_quantizer_from_config(keras_layer, 'mean')
+    layer['variance_quantizer'] = get_quantizer_from_config(keras_layer, 'variance')
+    layer['beta_quantizer'] = get_quantizer_from_config(keras_layer, 'beta')
+    layer['gamma_quantizer'] = get_quantizer_from_config(keras_layer, 'gamma')
+
+    return layer, output_shape
+
+
+@keras_handler('QConv2DBatchnorm')
+def parse_qconv2dbatchnorm_layer(keras_layer, input_names, input_shapes, data_reader):
+    intermediate_shape = list()
+    conv_layer, shape_qconv = parse_qconv_layer(keras_layer, input_names, input_shapes, data_reader)
+    intermediate_shape.append(shape_qconv)
+    temp_shape = intermediate_shape
+    batch_layer, out_shape = parse_batchnorm_layer(keras_layer, input_names, temp_shape, data_reader)
+    return {**conv_layer, **batch_layer}, out_shape