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('BinaryQuantizer suppots 1 or 2 bits, but called with bits={}'.format(bits))
super(BinaryQuantizer, self).__init__(bits, hls_type)
def transform(self, model, node):
# Compute the required precision and update the variables
# Number of bits for output is log2 of number of input nodes
# Since this is the number of uint<1>'s which are summed
nbits = int(np.ceil(np.log2(node.attributes['n_in'])) + 2)
out_type = IntegerPrecisionType(width=nbits)
accum_t = NamedType('layer{}_accum_t'.format(node.index), out_type)
node.set_attr('accum_t', accum_t)
out_var = node.get_output_variable()
out_var.type.precision = out_type
quantized_data = None
quantized_precision = None
quantizer = node.get_attr('weight_quantizer')
if quantizer.__class__.__name__ == 'BinaryQuantizer':
quantized_precision = XnorPrecisionType()
elif quantizer.__class__.__name__ == 'TernaryQuantizer':
quantized_precision = IntegerPrecisionType(width=2)
else:
print('WARNING: Unknown quantizer - {}. Bailing out'.format(
quantizer.__class__.__name__))
return False
quantizer.bits = quantized_precision.width
quantizer.hls_type = quantized_precision
quantized_data = quantizer(node.weights['weight'].data)
weights = node.weights['weight']
weights.data = quantized_data
weights.type.name = 'weight{index}_t'.format(index=node.index)
weights.update_precision(quantized_precision)
bias = node.weights['bias']
bias.data = np.zeros(shape=(node.get_attr('n_out')))
bias.type.name = 'bias{index}_t'.format(index=node.index)
bias.nzeros = 0
bias.update_precision(quantized_precision)
# If followed by the BatchNormalizationBinaryTanh, update its input
# Also requantise the weights
bd_out_nodes = node.get_output_nodes()
for out_node in bd_out_nodes:
if isinstance(out_node, BatchNormalizationQuantizedTanh):
var_names = []
if quantizer.__class__.__name__ == 'BinaryQuantizer':
var_names.append('threshold')
elif quantizer.__class__.__name__ == 'TernaryQuantizer':
var_names.append('threshold_hi')
var_names.append('threshold_lo')
for var_name in var_names:
threshold_var = out_node.weights[var_name]
threshold_var.update_precision(out_type)
threshold_var.data = np.floor(threshold_var.data)
return False
def transform(self, model, node):
old_precision = node.get_output_variable().type.precision
if isinstance(old_precision, IntegerPrecisionType):
new_precision = IntegerPrecisionType(old_precision.width,
old_precision.signed)
elif isinstance(old_precision, FixedPrecisionType):
new_precision = FixedPrecisionType(old_precision.width,
old_precision.integer,
old_precision.signed,
self.rounding_mode,
self.saturation_mode,
self.saturation_bits)
else: # in case the precision is a string
new_precision = self.precision_string_modify(old_precision)
out_var = node.get_output_variable()
out_t = NamedType(out_var.type.name, new_precision)
out_var.type = out_t
node.attributes['result_t'] = out_t
if node.get_attr('accum_t') is not None:
accum_t = NamedType('layer{}_accum_t'.format(node.index),
new_precision)
node.set_attr('accum_t', new_precision)
return False
def init_gru(self, layer):
reuse_factor = layer.model.config.get_reuse_factor(layer)
layer.set_attr('recurrent_reuse_factor', reuse_factor)
recurrent_bias = np.zeros(layer.weights['recurrent_weight'].shape[1])
layer.add_weights_variable(name='recurrent_bias',
var_name='br{index}',
data=recurrent_bias)
index_t = IntegerPrecisionType(width=1, signed=False)
if 'table_t' not in layer.attributes:
layer.set_attr('table_t', FixedPrecisionType(width=18, integer=8))
if 'table_size' not in layer.attributes:
layer.set_attr('table_size', 1024)
if layer.model.config.is_resource_strategy(layer):
n_in, n_out, n_in_recr, n_out_recr = self.get_layer_mult_size(
layer)
self.set_closest_reuse_factor(layer, n_in, n_out)
self.set_closest_reuse_factor(layer,
n_in_recr,
n_out_recr,
attribute='recurrent_reuse_factor')
layer.weights['weight'].data = np.transpose(
layer.weights['weight'].data)
layer.weights['recurrent_weight'].data = np.transpose(
layer.weights['recurrent_weight'].data)
layer.set_attr('strategy', 'resource')
else:
layer.set_attr('strategy', 'latency')
layer.set_attr('index_t', index_t)
def _convert_ac_type(cls, precision):
'''
Convert a precision string (e.g. "ac_fixed<16,6>" to the internal FixedPrecisionTypes etc)
'''
bits = re.search('.+<(.+?)>', precision).group(1).split(',')
signed = True # default is signed
sat_mode = None
round_mode = None
if 'fixed' in precision:
W = int(bits[0])
I = int(bits[1])
fields = 2
if len(bits) > 2:
signed = bool(bits[2])
fields = 3
elif 'int' in precision:
W = int(bits[0])
I = W
fields = 1
if len(bits) > 1:
signed = bool(bits[1])
fields = 2
if len(bits) > fields:
round_mode = bits[fields]
if len(bits) > fields + 1:
sat_mode = bits[fields + 1]
if 'fixed' in precision:
return FixedPrecisionType(W, I, signed, round_mode, sat_mode)
elif 'int' in precision:
return IntegerPrecisionType(W, signed)
def _convert_ap_type(cls, precision):
'''
Convert a precision string (e.g. "ap_fixed<16,6>" to the internal FixedPrecisionTypes etc)
'''
bits = re.search('.+<(.+?)>', precision).group(1).split(',')
sat_mode = None
round_mode = None
sat_bits = None
if 'fixed' in precision:
W = int(bits[0])
I = int(bits[1])
fields = 2
signed = not ('u' in precision)
elif 'int' in precision:
W = int(bits[0])
I = W
fields = 1
signed = not ('u' in precision)
if len(bits) > fields:
round_mode = bits[fields]
if len(bits) > fields + 1:
sat_mode = bits[fields + 1]
if len(bits) > fields + 2:
sat_bits = int(bits[fields + 2])
if 'fixed' in precision:
return FixedPrecisionType(W, I, signed, round_mode, sat_mode,
sat_bits)
elif 'int' in precision:
return IntegerPrecisionType(W, signed)
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 initialize(self):
inp = self.get_input_variable()
shape = inp.shape
dims = inp.dim_names
if self.get_attr('quantize') == 2:
self.add_output_variable(shape, dims, precision=XnorPrecisionType())
elif self.get_attr('quantize') == 3:
self.add_output_variable(shape, dims, precision=IntegerPrecisionType(width=2))
else:
raise Exception('Unsupported quantize attribute for BatchNormalizationQuantizedTanh: {}'.format(self.get_attr('quantize')))
def add_bias(self, quantizer=None):
data = self.model.get_weights_data(self.name, 'bias')
precision = None
type_name = None
if data is None:
data = np.zeros(self.get_output_variable().shape[-1])
precision = IntegerPrecisionType(width=1, signed=False)
type_name = 'bias{index}_t'
quantizer = None # Don't quantize non-existant bias
self.add_weights_variable(name='bias', var_name='b{index}', type_name=type_name, precision=precision, data=data, quantizer=quantizer)
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)
def init_dense(self, layer):
index_t = IntegerPrecisionType(width=1, signed=False)
compression = layer.model.config.get_compression(layer)
if layer.model.config.is_resource_strategy(layer):
self.set_target_reuse_factor(layer)
self.set_closest_reuse_factor(layer)
if compression:
layer.set_attr('strategy', 'compressed')
index_t = layer.get_weights('weight').type.index_precision
else:
layer.set_attr('strategy', 'resource')
else:
layer.set_attr('strategy', 'latency')
layer.set_attr('index_t',
NamedType('layer{}_index'.format(layer.index), index_t))
def _add_variable(self, name, var_name, data, frac_width=10, quantize=False):
# Wrapper for add_weights_variable with precision determination from data
# automatically make the variable unsigned if data are all positive
signed = (np.amin(data) < 0.)
int_width = find_minimum_width(data, signed=signed)
if quantize:
precision = IntegerPrecisionType(width=int_width, signed=signed)
else:
width = int_width + frac_width
precision = FixedPrecisionType(width=width, integer=int_width, signed=signed, rounding_mode='AP_RND', saturation_mode='AP_SAT')
self.add_weights_variable(name=name, var_name=var_name, data=data, precision=precision)
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 parse_input_layer(keras_layer, input_names, input_shapes, data_reader, config):
assert(keras_layer['class_name'] == 'InputLayer')
layer = parse_default_keras_layer(keras_layer, input_names)
layer['input_shape'] = keras_layer['config']['batch_input_shape'][1:]
dtype = keras_layer['config']['dtype']
if dtype.startswith('int') or dtype.startswith('uint'):
layer['type_name'] = 'integer_input_t'
width = int(dtype[dtype.index('int') + 3:])
signed = (not dtype.startswith('u'))
layer['precision'] = IntegerPrecisionType(width=width, signed=signed)
# elif bool, q[u]int, ...
output_shape = keras_layer['config']['batch_input_shape']
return layer, output_shape
def set_thresholds(self, scale, bias, ternary_threshold=0.5):
inp = self.get_input_variable()
shape = inp.shape
dims = inp.dim_names
precision = self.model.config.backend.convert_precision_string(inp.type.precision)
W, I, F = precision.width, precision.integer, precision.fractional
threshold = - bias / scale
if self.get_attr('quantize') == 2:
self.add_output_variable(shape, dims, precision=XnorPrecisionType())
threshold = np.floor(threshold * 2**F) / 2**F
self.add_weights_variable(name='threshold', var_name='t{index}', data=threshold, type_name='threshold{index}_t', precision=inp.type.precision)
elif self.get_attr('quantize') == 3:
self.add_output_variable(shape, dims, precision=IntegerPrecisionType(width=2))
threshold_hi = ternary_threshold / scale + threshold
threshold_lo = -ternary_threshold / scale + threshold
threshold_hi = np.floor(threshold_hi * 2**F) / 2**F
threshold_lo = np.floor(threshold_lo * 2**F) / 2**F
self.add_weights_variable(name='threshold_hi', var_name='th{index}', data=threshold_hi, type_name='threshold_hi_{index}_t', precision=inp.type.precision)
self.add_weights_variable(name='threshold_lo', var_name='tl{index}', data=threshold_lo, type_name='threshold_lo_{index}_t', precision=inp.type.precision)
def init_dense(self, layer):
index_t = IntegerPrecisionType(width=1, signed=False)
layer.set_attr('rfpad', 0)
layer.set_attr('bfpad', 0)
if layer.model.config.get_compression(layer):
layer.set_attr('strategy', 'compressed')
else:
self.set_closest_reuse_factor(layer)
self.gen_quartus_weight_array(layer)
layer.set_attr('strategy', 'resource')
if layer.model.config.is_resource_strategy(layer):
if layer.model.config.get_compression(layer):
index_t = layer.get_weights('weight').type.index_precision
layer.set_attr('index_t',
NamedType('layer{}_index'.format(layer.index), index_t))
def initialize(self):
if self.get_attr('data_format') == 'channels_last':
shape = [self.attributes['out_height'], self.attributes['out_width'], self.attributes['n_filt']]
dims = ['OUT_HEIGHT_{}'.format(self.index), 'OUT_WIDTH_{}'.format(self.index), 'N_FILT_{}'.format(self.index)]
else:
shape = [self.attributes['n_filt'], self.attributes['out_height'], self.attributes['out_width']]
dims = ['N_FILT_{}'.format(self.index), 'OUT_HEIGHT_{}'.format(self.index), 'OUT_WIDTH_{}'.format(self.index)]
self.add_output_variable(shape, dims)
depthwise_data = self.model.get_weights_data(self.name, 'depthwise_kernel')
pointwise_data = self.model.get_weights_data(self.name, 'pointwise_kernel')
self.add_weights_variable(name='depthwise', var_name='d{index}', data=depthwise_data, quantizer=self.get_attr('depthwise_quantizer'))
self.add_weights_variable(name='pointwise', var_name='p{index}', data=pointwise_data, quantizer=self.get_attr('pointwise_quantizer'))
zero_bias_data = np.zeros((self.attributes['n_chan'],))
precision = IntegerPrecisionType(width=1, signed=False)
self.add_weights_variable(name='zero_bias', var_name='z{index}', data=zero_bias_data, precision=precision)
self.add_bias(quantizer=self.get_attr('bias_quantizer'))
def __init__(self):
super(TernaryQuantizer, self).__init__(2,
IntegerPrecisionType(width=2))
