def generate_layer_data_type_map(graph,
source_quantizer_list,
is_inference,
keras_quantizer=None,
keras_accumulator=None,
for_reference=False,
debug=False):
"""main funciton to generate datatype for each layer.
For each type of layer, this function calculates the sizes and minimum
number of bits required to represent the parameters and variables (e.g.,
weights, bias, multiplier and accumulator - MAC, etc.) embedded in
these layers.
Args:
graph: input graph that traverses the model
source_quantizer_list: a list of quantizers for model inputs
is_inference: whether model is pre-trained with weights available
keras_quantizer: default quantizer used to quantize un-quantized layers
keras_accumulator: default MAC quantizer to quantize un-quantized layers
for_reference: whether to generate a map for a baseline model
debug: whether to print debug messages
Returns:
a result containing the following fields:
source_quantizer_list similar as input
output_layers: names of the layers that are output layers
input_layers: names of the layers that are input_layers,
layer_data_type_map: data type map of each layer
"""
quantizer_factory = quantizer_factory_module.QuantizerFactory()
layer_data_type_map = collections.OrderedDict()
# get the output layers
output_layers = []
input_layers = []
predecessors = list(graph.predecessors(qgraph.SINK))
successors = list(graph.successors(qgraph.SOURCE))
for u in predecessors:
if u == qgraph.SOURCE or u == qgraph.SINK:
continue
output_layers.append(graph.nodes[u]["layer"][0])
for u in successors:
if u == qgraph.SOURCE or u == qgraph.SINK:
continue
input_layers.append(graph.nodes[u]["layer"][0])
for node_id in nx.topological_sort(graph):
node = graph.nodes[node_id]
node_type = node["type"][-1]
layer = node["layer"][0]
is_input_layer = layer in input_layers
w_shapes = None
b_shapes = None
output_shapes = None
if hasattr(layer, "output_shape"):
output_shapes = layer.output_shape
if hasattr(layer, "get_weights"):
weights = layer.get_weights()
if len(weights) != 0:
w_shapes = layer.get_weights()[0].shape
b_shapes = weights[0].shape[-1]
if debug:
print("########")
if layer is not None:
print(layer.name)
else:
print("None")
# Deals with keras layer or lack of input quantizer in qkeras layer.
input_qe_list = qtools_util.get_input_quantizers_advanced(
graph, node_id, is_input_layer, quantizer_factory, cfg)
if input_qe_list and node_id != qgraph.SINK:
input_quantizer_list = []
for node in input_qe_list:
input_quantizer_list.append(node[0])
# Calculates number of operations (multiplication/accumulation).
(_, edge_0) = input_qe_list[0]
input_shape = edge_0["shape"]
# for merge layers, all input_shape are identical
operation_count = qtools_util.get_operation_count(
layer, input_shape)
# Merges layers with multiple inputs.
if qtools_util.is_merge_layers(layer):
# output_shapes = layer.output_shape
merge_factory = quantized_operators.MergeFactory()
merge_quantizer = merge_factory.make_quantizer(
input_qe_list, layer.__class__.__name__)
if hasattr(layer, "get_quantizers"):
# QMerge layer from future.
qkeras_quantizer = layer.get_quantizers()[0]
merge_quantizer.output = quantizer_factory.make_quantizer(
qkeras_quantizer)
else:
# Merge layer is a Keras layer.
if for_reference:
merge_quantizer.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_accumulator:
# gate_factor and gate_bits remain the same as previously
# calculated; only change output quantizer as the keras_accumulator
merge_quantizer.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, merge_quantizer.output,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, merge_quantizer, None, None, None, None,
None, output_quantizer, output_shapes, operation_count)
# MaxPooling/reshape/flatten/UpSampling1D/2D/3D
elif (qtools_util.is_shape_alternation_layers(layer)
or "UpSampling" in layer.__class__.__name__):
input_quantizer = input_quantizer_list[0]
# Output quantizer
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, input_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, None, None, None, None, None, None,
output_quantizer, output_shapes, operation_count)
# AveragePooling and GlobalAveragePooling
elif layer.__class__.__name__ in [
"AveragePooling2D", "AvgPool2D", "GlobalAvgPool2D",
"GlobalAveragePooling2D", "QAveragePooling2D",
"QGlobalAveragePooling2D"
]:
(input_quantizer, _) = input_qe_list[0]
# This is a hack. We don't want to implement a new accumulator class
# just for averagpooling. So we re-use accumulator type in conv/dense
# layers which need multiplier and kernel as input parameters.
# In order to do so, we fake a multiplier which treat the pool_size as
# the kernel. since kernel needs 4 dimension, k_h, k_w, C_in, C_out,
# we set the last two dimension as [1, 1]
if layer.__class__.__name__ in [
"AveragePooling2D", "AvgPool2D", "QAveragePooling2D"
]:
pool_size = tuple(list(layer.pool_size) + [1, 1])
else:
pool_size = tuple(list(input_shape)[1:-1] + [1, 1])
multiplier_factory = quantized_operators.MultiplierFactory()
fake_multiplier = multiplier_factory.make_multiplier(
input_quantizer, input_quantizer)
fake_multiplier.output = input_quantizer
accumulator_factory = quantized_operators.AccumulatorFactory()
accumulator = accumulator_factory.make_accumulator(pool_size,
fake_multiplier,
use_bias=False)
if layer.__class__.__name__ in [
"QAveragePooling2D", "QGlobalAveragePooling2D"
]:
# For the quantized layer, there is an average_quantizer used for
# the inverse of division operation.
qkeras_average_quantizer = layer.get_quantizers()[0]
qtools_average_quantizer = quantizer_factory.make_quantizer(
qkeras_average_quantizer)
multiplier = multiplier_factory.make_multiplier(
accumulator.output, qtools_average_quantizer)
else:
multiplier = None
if debug:
print("accumulator:", accumulator.output.bits)
if for_reference:
if multiplier:
multiplier.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
accumulator.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_accumulator:
if multiplier:
multiplier.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
accumulator.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
if layer.__class__.__name__ in [
"QAveragePooling2D", "QGlobalAveragePooling2D"
]:
# If is quantized layer, last operation is multiply (averaging).
layer_quantizer = multiplier.output
else:
layer_quantizer = accumulator.output
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, multiplier, accumulator, None, None,
None, None, output_quantizer, output_shapes, operation_count)
# If it's a Quantized Activation layer.
elif node_type in ["QActivation", "QAdaptiveActivation", "Activation"]:
if for_reference or not hasattr(layer, "quantizer"):
# Keras activation layer -> use default_interm_quantizer
layer_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_quantizer:
layer_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
else:
layer_quantizer = layer.quantizer
if not quantizer_factory.is_quantizer_supported(
layer_quantizer):
raise TagMissingError(
"Unsupported activation quantizer {} on this layer: {}"
.format(layer_quantizer, layer))
if not layer_quantizer:
layer_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, None, None, None, w_shapes, None,
b_shapes, output_quantizer, output_shapes, operation_count)
elif node_type in ["QBatchNormalization", "BatchNormalization"]:
(input_quantizer, _) = input_qe_list[0]
# QKeras layers might be mixed with keras layers.
if for_reference or not hasattr(layer, "get_quantizers"):
# Keras BatchNorm layer mixed with quantized model
# -> no reference mode
gamma_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
beta_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
mean_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
variance_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_quantizer:
gamma_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
beta_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
mean_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
variance_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
else:
(qkeras_gamma_quantizer, qkeras_beta_quantizer,
qkeras_mean_quantizer,
qkeras_variance_quantizer) = layer.get_quantizers()
if not qkeras_beta_quantizer:
beta_quantizer = quantizer_factory.clone_quantizer(
input_quantizer)
else:
beta_quantizer = quantizer_factory.make_quantizer(
qkeras_beta_quantizer)
if not qkeras_mean_quantizer:
mean_quantizer = quantizer_factory.clone_quantizer(
input_quantizer)
else:
mean_quantizer = quantizer_factory.make_quantizer(
qkeras_mean_quantizer)
if not qkeras_variance_quantizer:
variance_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
else:
# If gamma is float, convert to input_quantizer.
variance_quantizer = quantizer_factory.make_quantizer(
qkeras_variance_quantizer)
if not qkeras_gamma_quantizer:
gamma_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
else:
gamma_quantizer = quantizer_factory.make_quantizer(
qkeras_gamma_quantizer)
# During inference, gamma, beta and variance are constants
# if they are po2 quantizers, we need to modify their bits
# with actual values and also update graph with the
# corresponding output_quantizer on the edge.
if is_inference:
weights = qtools_util.get_weights(layer)
# If no scale(gamma), num_weights --
# If no center(beta_quantizer) num_weights --
num_weights = 4
if not layer.scale:
num_weights -= 1
if not layer.center:
num_weights -= 1
if layer.scale and gamma_quantizer.is_po2:
gamma_quantizer.update_inference_values(weights[0])
if variance_quantizer.is_po2:
variance_quantizer.update_inference_values(
weights[num_weights - 1])
qbn = quantized_operators.QBNFactory()
qbn.make_quantizer(input_quantizer, gamma_quantizer,
beta_quantizer, mean_quantizer,
variance_quantizer, layer.scale, layer.center)
def set_output(op, output):
if op:
op.output = output
if for_reference or not hasattr(layer, "get_quantizers"):
set_output(
qbn.internal_divide_quantizer,
quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer))
set_output(
qbn.internal_multiplier,
quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer))
set_output(
qbn.internal_accumulator,
quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer))
set_output(
qbn.internal_output,
quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer))
if keras_accumulator:
set_output(
qbn.internal_divide_quantizer,
quantizer_factory.make_default_quantizer(
mode=keras_accumulator))
set_output(
qbn.internal_multiplier,
quantizer_factory.make_default_quantizer(
mode=keras_accumulator))
set_output(
qbn.internal_accumulator,
quantizer_factory.make_default_quantizer(
mode=keras_accumulator))
set_output(
qbn.internal_output.output,
quantizer_factory.make_default_quantizer(
mode=keras_accumulator))
layer_quantizer = qbn.internal_output.output
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
gamma_range = None
if hasattr(layer, "gamma_range"):
gamma_range = layer.gamma_range
beta_range = None
if hasattr(layer, "beta_range"):
beta_range = layer.beta_range
if not layer.center:
qbn.beta_quantizer = None
if not layer.scale:
qbn.gamma_quantizer = None
layer_data_type_map[layer] = {
"input_quantizer_list": input_quantizer_list,
"gamma_quantizer": gamma_quantizer,
"beta_quantizer": beta_quantizer,
"mean_quantizer": mean_quantizer,
"variance_quantizer": variance_quantizer,
"gamma_range": gamma_range,
"beta_range": beta_range,
"internal_divide_quantizer": qbn.internal_divide_quantizer,
"internal_multiplier": qbn.internal_multiplier,
"internal_accumulator": qbn.internal_accumulator,
"output_quantizer": output_quantizer,
"output_shapes": input_shape,
"operation_count": operation_count
}
# If qdense, qconv, qpool, qoctave
elif node_type in QKERAS_LAYERS or node_type in KERAS_LAYERS:
(input_quantizer, _) = input_qe_list[0]
if for_reference or not hasattr(layer, "get_quantizers"):
# for_reference: force all quantizers to keras_quantizer
weight_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
bias_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_quantizer:
weight_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
bias_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
else:
# qkeras layer
qkeras_weight_quantizer = layer.get_quantizers()[0]
qkeras_bias_quantizer = layer.get_quantizers()[1]
if not quantizer_factory.is_quantizer_supported(
qkeras_weight_quantizer):
raise TagMissingError(
"Unsupported weight quantizer {} on this layer: {}".
format(qkeras_weight_quantizer, layer))
if not quantizer_factory.is_quantizer_supported(
qkeras_bias_quantizer):
raise TagMissingError(
"Unsupported bias quantizer {} on this layer: {}".
format(qkeras_bias_quantizer, layer))
weight_quantizer = quantizer_factory.make_quantizer(
qkeras_weight_quantizer)
bias_quantizer = quantizer_factory.make_quantizer(
qkeras_bias_quantizer)
# TODO(lishanok): During inference, if weight and bias is po2,
# need to update corresponding quantizer type with min and max
# of the constant values.
if is_inference:
weights = qtools_util.get_weights(layer)
if weight_quantizer.is_po2:
weight_quantizer.update_inference_values(weights[0])
if bias_quantizer.is_po2:
bias_quantizer.update_inference_values(weights[1])
multiplier_factory = quantized_operators.MultiplierFactory()
multiplier = multiplier_factory.make_multiplier(
weight_quantizer, input_quantizer)
weights = layer.get_weights()
kernel = weights[0]
accumulator_factory = quantized_operators.AccumulatorFactory()
accumulator = accumulator_factory.make_accumulator(
kernel.shape, multiplier)
if not layer.use_bias:
bias_quantizer = None
if debug:
print(layer.name or "None")
print("weight_quantizer:", weight_quantizer.bits)
print("input_quantizer:", input_quantizer.bits)
print("multiplier_quantizer:", multiplier.output.bits)
print("multiplier_gate_bits:", multiplier.gate_bits)
print("accumulator:", accumulator.output.bits)
if for_reference or not hasattr(layer, "get_quantizers"):
accumulator.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
multiplier.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_accumulator:
accumulator.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
multiplier.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
layer_quantizer = accumulator.output
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, multiplier, accumulator,
weight_quantizer, w_shapes, bias_quantizer, b_shapes,
output_quantizer, output_shapes, operation_count)
# Folded conv/dense/depthwiseconv layer
elif node_type in ["QConv2DBatchnorm", "QDepthwiseConv2DBatchnorm"]:
(input_quantizer, _) = input_qe_list[0]
if for_reference or not hasattr(layer, "get_quantizers"):
# For_reference: force all quantizers to keras_quantizer.
weight_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
bias_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_quantizer:
weight_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
bias_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
else:
# QKeras layer
qkeras_weight_quantizer = layer.get_quantizers()[0]
qkeras_bias_quantizer = layer.get_quantizers()[1]
if not quantizer_factory.is_quantizer_supported(
qkeras_weight_quantizer):
raise TagMissingError(
"Unsupported weight quantizer {} on this layer: {}".
format(qkeras_weight_quantizer, layer))
if not quantizer_factory.is_quantizer_supported(
qkeras_bias_quantizer):
raise TagMissingError(
"Unsupported bias quantizer {} on this layer: {}".
format(qkeras_bias_quantizer, layer))
weight_quantizer = quantizer_factory.make_quantizer(
qkeras_weight_quantizer)
if qkeras_bias_quantizer:
bias_quantizer = quantizer_factory.make_quantizer(
qkeras_bias_quantizer)
else:
bias_quantizer = None
# TODO(lishanok): During inference, if weight and bias is po2,
# need to update corresponding quantizer type with min and max
# of the constant values
if is_inference:
weights = qtools_util.get_weights(layer)
if weight_quantizer.is_po2:
weight_quantizer.update_inference_values(weights[0])
if bias_quantizer and bias_quantizer.is_po2:
bias_quantizer.update_inference_values(weights[1])
multiplier_factory = quantized_operators.MultiplierFactory()
multiplier = multiplier_factory.make_multiplier(
weight_quantizer, input_quantizer)
weights = layer.get_weights()
kernel = weights[0]
accumulator_factory = quantized_operators.AccumulatorFactory()
accumulator = accumulator_factory.make_accumulator(
kernel.shape,
multiplier,
use_bias=True if bias_quantizer else False)
if not bias_quantizer:
# Sets bias the same as accumulator type.
bias_quantizer = copy.deepcopy(accumulator.output)
if not accumulator.output.is_floating_point:
# For fixed point accumulator, needs to add 1 to its bits to avoid
# possible satuation.
accumulator.output.bits += 1
accumulator.output.int_bits += 1
if for_reference or not hasattr(layer, "get_quantizers"):
accumulator.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
multiplier.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_accumulator:
accumulator.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
multiplier.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
layer_quantizer = accumulator.output
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, multiplier, accumulator,
weight_quantizer, w_shapes, bias_quantizer, b_shapes,
output_quantizer, output_shapes, operation_count)
elif node_type:
# Any other unsupported layer types -> pass the input quantizer
# type to output in qraph
(input_quantizer, _) = input_qe_list[0]
if for_reference and keras_quantizer:
input_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, input_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, None, None, None, None, None, None,
output_quantizer, output_shapes, operation_count)
result = {
"source_quantizer_list": source_quantizer_list,
"output_layers": output_layers,
"input_layers": input_layers,
"layer_data_type_map": layer_data_type_map
}
return result
def generate_layer_data_type_map(graph,
source_quantizer_list,
is_inference,
keras_quantizer=None,
keras_accumulator=None,
for_reference=False,
debug=False):
"""main funciton to generate datatype for each layer.
For each type of layer, this function calculates the sizes and minimum
number of bits required to represent the parameters and variables (e.g.,
weights, bias, multiplier and accumulator - MAC, etc.) embedded in
these layers.
Args:
graph: input graph that traverses the model
source_quantizer_list: a list of quantizers for model inputs
is_inference: whether model is pre-trained with weights available
keras_quantizer: default quantizer used to quantize un-quantized layers
keras_accumulator: default MAC quantizer to quantize un-quantized layers
for_reference: whether to generate a map for a baseline model
debug: whether to print debug messages
Returns:
a result containing the following fields:
source_quantizer_list similar as input
output_layers: names of the layers that are output layers
input_layers: names of the layers that are input_layers,
layer_data_type_map: data type map of each layer
"""
quantizer_factory = quantizer_factory_module.QuantizerFactory()
layer_data_type_map = collections.OrderedDict()
# get the output layers
output_layers = []
input_layers = []
predecessors = list(graph.predecessors(qgraph.SINK))
successors = list(graph.successors(qgraph.SOURCE))
for u in predecessors:
if u == qgraph.SOURCE or u == qgraph.SINK:
continue
output_layers.append(graph.nodes[u]["layer"][0])
for u in successors:
if u == qgraph.SOURCE or u == qgraph.SINK:
continue
input_layers.append(graph.nodes[u]["layer"][0])
for node_id in nx.topological_sort(graph):
node = graph.nodes[node_id]
node_type = node["type"][-1]
layer = node["layer"][0]
is_input_layer = layer in input_layers
w_shapes = None
b_shapes = None
output_shapes = None
if hasattr(layer, "output_shape"):
output_shapes = layer.output_shape
if hasattr(layer, "get_weights"):
weights = layer.get_weights()
if len(weights) != 0:
w_shapes = layer.get_weights()[0].shape
b_shapes = weights[0].shape[-1]
if debug:
print("########")
if layer is not None:
print(layer.name)
else:
print("None")
# deal with keras layer or lack of input quantizer in qkeras layer
input_qe_list = qtools_util.get_input_quantizers_advanced(
graph, node_id, is_input_layer, quantizer_factory, cfg)
if input_qe_list and node_id != qgraph.SINK:
input_quantizer_list = []
for node in input_qe_list:
input_quantizer_list.append(node[0])
# calculate number of operations (multiplication/accumulation)
(_, edge_0) = input_qe_list[0]
input_shape = edge_0["shape"]
# for merge layers, all input_shape are identical
operation_count = qtools_util.get_operation_count(
layer, input_shape)
# Merge Layers with multiple inputs
if qtools_util.is_merge_layers(layer):
# output_shapes = layer.output_shape
merge_factory = quantized_operators.MergeFactory()
merge_quantizer = merge_factory.make_quantizer(
input_qe_list, layer.__class__.__name__)
if hasattr(layer, "get_quantizers"):
# QMerge layer from future
qkeras_quantizer = layer.get_quantizers()[0]
merge_quantizer.output = quantizer_factory.make_quantizer(
qkeras_quantizer)
else:
# merge layer is a Keras layer
if for_reference:
merge_quantizer.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_accumulator:
# gate_factor and gate_bits remain the same as previously
# calculated; only change output quantizer as the keras_accumulator
merge_quantizer.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, merge_quantizer.output,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, merge_quantizer, None, None, None, None,
None, output_quantizer, output_shapes, operation_count)
# pooling/reshape/flatten
if qtools_util.is_shape_alternation_layers(layer):
input_quantizer = input_quantizer_list[0]
# output quantizer
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, input_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, None, None, None, None, None, None,
output_quantizer, output_shapes, operation_count)
# if Quantized Activation layer
elif node_type in ["QActivation", "Activation"]:
if for_reference or not hasattr(layer, "quantizer"):
# Keras activation layer -> use default_interm_quantizer
layer_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_quantizer:
layer_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
else:
layer_quantizer = layer.quantizer
if not quantizer_factory.is_quantizer_supported(
layer_quantizer):
raise TagMissingError(
"Unsupported activation quantizer {} on this layer: {}"
.format(layer_quantizer, layer))
if not layer_quantizer:
layer_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, None, None, None, w_shapes, None,
b_shapes, output_quantizer, output_shapes, operation_count)
elif node_type in ["QBatchNormalization", "BatchNormalization"]:
(input_quantizer, _) = input_qe_list[0]
# qkeras layers might be mixed with keras layers
if for_reference or not hasattr(layer, "get_quantizers"):
# Keras BatchNorm layer mixed with quantized model
# -> no reference mode
gamma_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
beta_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
mean_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
variance_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_quantizer:
gamma_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
beta_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
mean_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
variance_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
else:
(qkeras_gamma_quantizer, qkeras_beta_quantizer,
qkeras_mean_quantizer,
qkeras_variance_quantizer) = layer.get_quantizers()
if not qkeras_beta_quantizer:
beta_quantizer = quantizer_factory.clone_quantizer(
input_quantizer)
else:
beta_quantizer = quantizer_factory.make_quantizer(
qkeras_beta_quantizer)
if not qkeras_mean_quantizer:
mean_quantizer = quantizer_factory.clone_quantizer(
input_quantizer)
else:
mean_quantizer = quantizer_factory.make_quantizer(
qkeras_mean_quantizer)
if not qkeras_variance_quantizer:
variance_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
else:
# if gamma is float, convert to input_quantizer
variance_quantizer = quantizer_factory.make_quantizer(
qkeras_variance_quantizer)
if not qkeras_gamma_quantizer:
gamma_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
else:
gamma_quantizer = quantizer_factory.make_quantizer(
qkeras_gamma_quantizer)
# during inference, gamma, beta and variance are constants
# if they are po2 quantizers, we need to modify their bits
# with actual values and also update graph with the
# corresponding output_quantizer on the edge
if is_inference:
weights = qtools_util.get_weights(layer)
# if no scale(gamma), num_weights --
# if no center(beta_quantizer) num_weights --
num_weights = 4
if not layer.scale:
num_weights -= 1
if not layer.center:
num_weights -= 1
if layer.scale and gamma_quantizer.is_po2:
gamma_quantizer.update_inference_values(weights[0])
if variance_quantizer.is_po2:
variance_quantizer.update_inference_values(
weights[num_weights - 1])
qbn = quantized_operators.QBNFactory()
qbn.make_quantizer(input_quantizer, gamma_quantizer,
beta_quantizer, mean_quantizer,
variance_quantizer, layer.scale, layer.center)
def set_output(op, output):
if op:
op.output = output
if for_reference or not hasattr(layer, "get_quantizers"):
set_output(
qbn.internal_divide_quantizer,
quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer))
set_output(
qbn.internal_multiplier,
quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer))
set_output(
qbn.internal_accumulator,
quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer))
set_output(
qbn.internal_output,
quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer))
if keras_accumulator:
set_output(
qbn.internal_divide_quantizer,
quantizer_factory.make_default_quantizer(
mode=keras_accumulator))
set_output(
qbn.internal_multiplier,
quantizer_factory.make_default_quantizer(
mode=keras_accumulator))
set_output(
qbn.internal_accumulator,
quantizer_factory.make_default_quantizer(
mode=keras_accumulator))
set_output(
qbn.internal_output.output,
quantizer_factory.make_default_quantizer(
mode=keras_accumulator))
layer_quantizer = qbn.internal_output.output
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
gamma_range = None
if hasattr(layer, "gamma_range"):
gamma_range = layer.gamma_range
beta_range = None
if hasattr(layer, "beta_range"):
beta_range = layer.beta_range
if not layer.center:
qbn.beta_quantizer = None
if not layer.scale:
qbn.gamma_quantizer = None
layer_data_type_map[layer] = {
"input_quantizer_list": input_quantizer_list,
"gamma_quantizer": gamma_quantizer,
"beta_quantizer": beta_quantizer,
"mean_quantizer": mean_quantizer,
"variance_quantizer": variance_quantizer,
"gamma_range": gamma_range,
"beta_range": beta_range,
"internal_divide_quantizer": qbn.internal_divide_quantizer,
"internal_multiplier": qbn.internal_multiplier,
"internal_accumulator": qbn.internal_accumulator,
"output_quantizer": output_quantizer,
"output_shapes": input_shape,
"operation_count": operation_count
}
# if qdense, qconv, qpool, qoctave
elif node_type in QKERAS_LAYERS or node_type in KERAS_LAYERS:
(input_quantizer, _) = input_qe_list[0]
if for_reference or not hasattr(layer, "get_quantizers"):
# for_reference: force all quantizers to keras_quantizer
weight_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
bias_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_quantizer:
weight_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
bias_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
else:
# qkeras layer
qkeras_weight_quantizer = layer.get_quantizers()[0]
qkeras_bias_quantizer = layer.get_quantizers()[1]
if not quantizer_factory.is_quantizer_supported(
qkeras_weight_quantizer):
raise TagMissingError(
"Unsupported weight quantizer {} on this layer: {}".
format(qkeras_weight_quantizer, layer))
if not quantizer_factory.is_quantizer_supported(
qkeras_bias_quantizer):
raise TagMissingError(
"Unsupported bias quantizer {} on this layer: {}".
format(qkeras_bias_quantizer, layer))
weight_quantizer = quantizer_factory.make_quantizer(
qkeras_weight_quantizer)
bias_quantizer = quantizer_factory.make_quantizer(
qkeras_bias_quantizer)
# TODO(lishanok): during inference, if weight and bias is po2,
# need to update corresponding quantizer type with min and max
# of the constant values
if is_inference:
weights = qtools_util.get_weights(layer)
if weight_quantizer.is_po2:
weight_quantizer.update_inference_values(weights[0])
if bias_quantizer.is_po2:
bias_quantizer.update_inference_values(weights[1])
multiplier_factory = quantized_operators.MultiplierFactory()
multiplier = multiplier_factory.make_multiplier(
weight_quantizer, input_quantizer)
weights = layer.get_weights()
kernel = weights[0]
# if layer.use_bias:
# bias = weights[1]
accumulator_factory = quantized_operators.AccumulatorFactory()
accumulator = accumulator_factory.make_accumulator(
kernel.shape, multiplier)
if not layer.use_bias:
bias_quantizer = None
if debug:
print(layer.name or "None")
print("weight_quantizer:", weight_quantizer.bits)
print("input_quantizer:", input_quantizer.bits)
print("multiplier_quantizer:", multiplier.output.bits)
print("multiplier_gate_bits:", multiplier.gate_bits)
print("accumulator:", accumulator.output.bits)
if for_reference or not hasattr(layer, "get_quantizers"):
accumulator.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
multiplier.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_accumulator:
accumulator.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
multiplier.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
layer_quantizer = accumulator.output
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, multiplier, accumulator,
weight_quantizer, w_shapes, bias_quantizer, b_shapes,
output_quantizer, output_shapes, operation_count)
elif node_type:
# any other unsupported layer types -> pass the input quantizer
# type to output in qraph
(input_quantizer, _) = input_qe_list[0]
if for_reference and keras_quantizer:
input_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
update_output_quantizer_in_graph(graph, node_id, quantizer_factory,
input_quantizer, for_reference)
result = {
"source_quantizer_list": source_quantizer_list,
"output_layers": output_layers,
"input_layers": input_layers,
"layer_data_type_map": layer_data_type_map
}
return result
def generate_layer_data_type_map(graph,
source_quantizer_list,
is_inference,
keras_quantizer=None,
keras_accumulator=None,
for_reference=False,
debug=False,
model_weights_already_quantized=True,
hw_weight_dict=None):
"""main funciton to generate datatype for each layer.
For each type of layer, this function calculates the sizes and minimum
number of bits required to represent the parameters and variables (e.g.,
weights, bias, multiplier and accumulator - MAC, etc.) embedded in
these layers.
Args:
graph: input graph that traverses the model
source_quantizer_list: a list of quantizers for model inputs
is_inference: whether model is pre-trained with weights available
keras_quantizer: default quantizer used to quantize weights and bias
keras_accumulator: default MAC quantizer to quantize multiplier,
accumulator and output
for_reference: whether to generate a map for a baseline model
debug: whether to print debug messages
model_weights_already_quantized: bool. If model weights are already
quantized, no need to apply quantizer to weights here in this function.
hw_weight_dict: weight dictonary for hardware inference. For example, fused
bn op inference in hardware will need additional fused weights, which
can be extracted from this dictionary. This dictionary is the output from
utils.py/model_save_quantized_weights function.
Returns:
a result containing the following fields:
source_quantizer_list similar as input
output_layers: names of the layers that are output layers
input_layers: names of the layers that are input_layers,
layer_data_type_map: data type map of each layer
"""
quantizer_factory = quantizer_factory_module.QuantizerFactory()
layer_data_type_map = collections.OrderedDict()
# get the output layers
output_layers = []
input_layers = []
predecessors = list(graph.predecessors(qgraph.SINK))
successors = list(graph.successors(qgraph.SOURCE))
for u in predecessors:
if u == qgraph.SOURCE or u == qgraph.SINK:
continue
output_layers.append(graph.nodes[u]["layer"][0])
for u in successors:
if u == qgraph.SOURCE or u == qgraph.SINK:
continue
input_layers.append(graph.nodes[u]["layer"][0])
for node_id in nx.topological_sort(graph):
node = graph.nodes[node_id]
node_type = node["type"][-1]
layer = node["layer"][0]
is_input_layer = layer in input_layers
w_shapes = None
b_shapes = None
output_shapes = None
qkeras_weight_quantizer = None
if hasattr(layer, "output_shape"):
output_shapes = layer.output_shape
if hasattr(layer, "get_weights"):
weights = layer.get_weights()
if len(weights) != 0:
w_shapes = layer.get_weights()[0].shape
b_shapes = weights[0].shape[-1]
if debug:
print("########")
if layer is not None:
print(layer.name)
else:
print("None")
# Deals with keras layer or lack of input quantizer in qkeras layer.
input_qe_list = qtools_util.get_input_quantizers_advanced(
graph, node_id, is_input_layer, quantizer_factory, cfg)
if input_qe_list and node_id != qgraph.SINK:
input_quantizer_list = []
for node in input_qe_list:
input_quantizer_list.append(node[0])
# Calculates number of operations (multiplication/accumulation).
# Previously Merge layers's inputs all have the same shape, however, in
# MobilenetV3 we found that there is shape broadcast in the keras
# Multiply layer. Therefore we use the shape with max size as the
# input shape
if len(input_qe_list) > 0:
maxsize = -1
max_id = 0
for (idx, item) in enumerate(input_qe_list):
shape = item[1]["shape"]
size = np.prod(shape[1:])
if size > maxsize:
maxsize = size
max_id = idx
input_shape = input_qe_list[max_id][1]["shape"]
else:
(_, edge_0) = input_qe_list[0]
input_shape = edge_0["shape"]
operation_count = qtools_util.get_operation_count(
layer, input_shape)
# Merges layers with multiple inputs.
if qtools_util.is_merge_layers(layer):
# merge_factory.make_quantizer automatically calculates the merge output
# quantizer bitwidth according to input quantizer type.
merge_factory = quantized_operators.MergeFactory()
merge_quantizer = merge_factory.make_quantizer(
input_qe_list, layer.__class__.__name__)
if for_reference:
# The for_reference option overwrites the auto-calculated merge output
# quantizer
if keras_accumulator:
# gate_factor and gate_bits remain the same as previously
# calculated; only change output quantizer as the keras_accumulator
merge_quantizer.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
else:
merge_quantizer.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, merge_quantizer.output,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, merge_quantizer, None, None, None, None,
None, output_quantizer, output_shapes, operation_count)
# MaxPooling/reshape/flatten/UpSampling1D/2D/3D
elif (qtools_util.is_shape_alternation_layers(layer)
or "UpSampling" in layer.__class__.__name__):
input_quantizer = input_quantizer_list[0]
# Output quantizer
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, input_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, None, None, None, None, None, None,
output_quantizer, output_shapes, operation_count)
# AveragePooling and GlobalAveragePooling
elif layer.__class__.__name__ in [
"AveragePooling2D", "AvgPool2D", "GlobalAvgPool2D",
"GlobalAveragePooling2D", "QAveragePooling2D",
"QGlobalAveragePooling2D"
]:
(input_quantizer, _) = input_qe_list[0]
qtools_average_quantizer = None
# This is a hack. We don't want to implement a new accumulator class
# just for averagpooling. So we re-use accumulator type in conv/dense
# layers which need multiplier and kernel as input parameters.
# In order to do so, we fake a multiplier which treat the pool_size as
# the kernel. since kernel needs 4 dimension, k_h, k_w, C_in, C_out,
# we set the last two dimension as [1, 1]
if layer.__class__.__name__ in [
"AveragePooling2D", "AvgPool2D", "QAveragePooling2D"
]:
pool_size = tuple(list(layer.pool_size) + [1, 1])
else:
pool_size = tuple(list(input_shape)[1:-1] + [1, 1])
# Automatically calculates the accumulator bitwidth according to input
# quantizer type for both quantized pooling and regular pooling layers
multiplier_factory = quantized_operators.MultiplierFactory()
fake_multiplier = multiplier_factory.make_multiplier(
input_quantizer, input_quantizer)
fake_multiplier.output = input_quantizer
accumulator_factory = quantized_operators.AccumulatorFactory()
accumulator = accumulator_factory.make_accumulator(pool_size,
fake_multiplier,
use_bias=False)
# For quantized pooling layers, we also need to consider the division
# precision, which is controlled by the average quantizer
if layer.__class__.__name__ in [
"QAveragePooling2D", "QGlobalAveragePooling2D"
]:
# For the quantized layer, there is an average_quantizer used for
# the inverse of division operation.
qkeras_average_quantizer = layer.get_quantizers()[0]
qtools_average_quantizer = quantizer_factory.make_quantizer(
qkeras_average_quantizer)
multiplier = multiplier_factory.make_multiplier(
accumulator.output, qtools_average_quantizer)
else:
multiplier = None
if debug:
print("accumulator:", accumulator.output.bits)
# Re-calcualte accumulator/multiplier type when it's using
# for_reference option
if for_reference:
if keras_accumulator:
# If keras_accumulator exists, use keras_accumulator as multiplier
# or accumulator type
if multiplier:
# Quantized layers need to define multiplier type
multiplier.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
accumulator.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
else:
# If user didn't provide keras_accumulator, use the default settings
# in cfg to define multiplier/accumulator type
if multiplier:
multiplier.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
accumulator.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
layer_quantizer = accumulator.output
# set the output quantizer
if layer.__class__.__name__ in [
"QAveragePooling2D", "QGlobalAveragePooling2D"
]:
# If is quantized layer, last operation is multiply (averaging).
layer_quantizer = multiplier.output
else:
layer_quantizer = accumulator.output
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
layer_data_type_map[layer] = {
"input_quantizer_list": input_quantizer_list,
"average_quantizer": qtools_average_quantizer,
"pool_sum_accumulator": accumulator,
"pool_avg_multiplier": multiplier,
"output_quantizer": output_quantizer,
"output_shapes": output_shapes,
"operation_count": operation_count
}
# If it's a Quantized Activation layer.
elif node_type in ["QActivation", "QAdaptiveActivation", "Activation"]:
if for_reference or not hasattr(layer, "quantizer"):
# Keras activation layer -> use default_interm_quantizer
layer_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_accumulator:
layer_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
else:
layer_quantizer = layer.quantizer
if not quantizer_factory.is_quantizer_supported(
layer_quantizer):
raise TagMissingError(
"Unsupported activation quantizer {} on this layer: {}"
.format(layer_quantizer, layer))
if not layer_quantizer:
layer_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, None, None, None, w_shapes, None,
b_shapes, output_quantizer, output_shapes, operation_count)
elif node_type in ["QBatchNormalization", "BatchNormalization"]:
# If this batchnorm layer needs to be fused with the previous layer,
# we pass the input quantizer type as the output type in qraph.
(input_quantizer, _) = input_qe_list[0]
if (hw_weight_dict is not None
and hw_weight_dict[layer.name]["enable_bn_fusing"]):
if for_reference and keras_accumulator and not is_input_layer:
input_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, input_quantizer,
for_reference)
layer_data_type_map[layer] = {
"input_quantizer_list": input_quantizer_list,
"output_quantizer": output_quantizer,
"output_shapes": input_shape,
"operation_count": operation_count
}
else:
(gamma_quantizer, beta_quantizer, mean_quantizer,
variance_quantizer,
_) = get_bn_quantizers(layer, quantizer_factory, cfg,
keras_quantizer, input_quantizer,
is_inference, for_reference,
model_weights_already_quantized)
qbn = quantized_operators.QBNFactory()
qbn.make_quantizer(input_quantizer, gamma_quantizer,
beta_quantizer, mean_quantizer,
variance_quantizer, layer.scale,
layer.center)
def set_output(op, output):
if op:
op.output = output
if for_reference or not hasattr(layer, "get_quantizers"):
set_output(
qbn.internal_divide_quantizer,
quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer))
set_output(
qbn.internal_multiplier,
quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer))
set_output(
qbn.internal_accumulator,
quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer))
set_output(
qbn.internal_output,
quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer))
if keras_accumulator:
set_output(
qbn.internal_divide_quantizer,
quantizer_factory.make_default_quantizer(
mode=keras_accumulator))
set_output(
qbn.internal_multiplier,
quantizer_factory.make_default_quantizer(
mode=keras_accumulator))
set_output(
qbn.internal_accumulator,
quantizer_factory.make_default_quantizer(
mode=keras_accumulator))
set_output(
qbn.internal_output.output,
quantizer_factory.make_default_quantizer(
mode=keras_accumulator))
gamma_range = None
if hasattr(layer, "gamma_range"):
gamma_range = layer.gamma_range
beta_range = None
if hasattr(layer, "beta_range"):
beta_range = layer.beta_range
if not layer.center:
qbn.beta_quantizer = None
if not layer.scale:
qbn.gamma_quantizer = None
layer_quantizer = qbn.internal_output.output
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
layer_data_type_map[layer] = {
"input_quantizer_list": input_quantizer_list,
"gamma_quantizer": gamma_quantizer,
"beta_quantizer": beta_quantizer,
"mean_quantizer": mean_quantizer,
"variance_quantizer": variance_quantizer,
"gamma_range": gamma_range,
"beta_range": beta_range,
"internal_divide_quantizer": qbn.internal_divide_quantizer,
"internal_multiplier": qbn.internal_multiplier,
"internal_accumulator": qbn.internal_accumulator,
"output_quantizer": output_quantizer,
"output_shapes": input_shape,
"operation_count": operation_count
}
# If qdense, qconv, qpool, qoctave
elif node_type in QKERAS_LAYERS or node_type in KERAS_LAYERS:
(input_quantizer, _) = input_qe_list[0]
if for_reference or not hasattr(layer, "get_quantizers"):
# for_reference: force all quantizers to keras_quantizer
weight_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
bias_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_quantizer:
weight_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
bias_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
else:
# qkeras layer
qkeras_weight_quantizer = layer.get_quantizers()[0]
qkeras_bias_quantizer = layer.get_quantizers()[1]
if not quantizer_factory.is_quantizer_supported(
qkeras_weight_quantizer):
raise TagMissingError(
"Unsupported weight quantizer {} on this layer: {}".
format(qkeras_weight_quantizer, layer))
if not quantizer_factory.is_quantizer_supported(
qkeras_bias_quantizer):
raise TagMissingError(
"Unsupported bias quantizer {} on this layer: {}".
format(qkeras_bias_quantizer, layer))
weight_quantizer = quantizer_factory.make_quantizer(
qkeras_weight_quantizer)
bias_quantizer = quantizer_factory.make_quantizer(
qkeras_bias_quantizer)
# TODO(lishanok): During inference, if weight and bias is po2,
# need to update corresponding quantizer type with min and max
# of the constant values.
if is_inference:
weights = qtools_util.get_weights(
layer, model_weights_already_quantized)
if weight_quantizer.is_po2:
weight_quantizer.update_inference_values(weights[0])
if bias_quantizer.is_po2:
bias_quantizer.update_inference_values(weights[1])
multiplier_factory = quantized_operators.MultiplierFactory()
multiplier = multiplier_factory.make_multiplier(
weight_quantizer, input_quantizer)
if qkeras_weight_quantizer:
qtools_util.adjust_multiplier_for_auto_po2(
multiplier, qkeras_weight_quantizer)
weights = layer.get_weights()
kernel = weights[0]
kernel_accumulator_factory = quantized_operators.AccumulatorFactory(
)
# Set use_bias=False so that the accumulator doesn't account for bias
# bitwdith
kernel_accumulator = kernel_accumulator_factory.make_accumulator(
kernel.shape, multiplier, use_bias=False)
if not layer.use_bias:
bias_quantizer = None
accumulator = kernel_accumulator
else:
# Add bias quantizer bitwidth to the overall accumulator
bias_accumulator_instance = adder_factory.IAdder()
accumulator = bias_accumulator_instance.make_quantizer(
kernel_accumulator.output, bias_quantizer)
if debug:
print(layer.name or "None")
print("weight_quantizer:", weight_quantizer.bits)
print("input_quantizer:", input_quantizer.bits)
print("multiplier_quantizer:", multiplier.output.bits)
print("multiplier_gate_bits:", multiplier.gate_bits)
print("accumulator:", accumulator.output.bits)
if for_reference or not hasattr(layer, "get_quantizers"):
accumulator.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
multiplier.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_accumulator:
accumulator.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
multiplier.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
if (hw_weight_dict is not None
and hw_weight_dict.get(layer.name, None) and
hw_weight_dict[layer.name].get("enable_bn_fusing", None)):
bn_layer_name = hw_weight_dict[
layer.name]["fused_bn_layer_name"]
successor_ids = list(graph.successors(node_id))
bn_layer = graph.nodes[successor_ids[0]]["layer"][0]
assert bn_layer.name == bn_layer_name, (
"Batchnorm layer in the graph has different name from hw_weight"
f"_dict: {layer.name} vs {bn_layer_name}. Check both places to "
"ensure they are matching.")
# Add additional datatype for bn fused weights
(gamma_quantizer, beta_quantizer, mean_quantizer,
variance_quantizer, inverse_quantizer) = get_bn_quantizers(
bn_layer, quantizer_factory, cfg, keras_quantizer,
input_quantizer, is_inference, for_reference,
model_weights_already_quantized)
fused_bn = FusedBNFactory()
fused_bn.make_quantizer(
prev_output_quantizer=kernel_accumulator.output,
prev_bias_quantizer=bias_quantizer,
beta_quantizer=beta_quantizer,
mean_quantizer=mean_quantizer,
inverse_quantizer=inverse_quantizer,
use_beta=bn_layer.center,
use_bias=layer.use_bias,
)
if for_reference or not hasattr(layer, "get_quantizers"):
fused_bn.internal_accumulator.output = (
quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer))
if keras_accumulator:
fused_bn.internal_accumulator.output = (
quantizer_factory.make_default_quantizer(
mode=keras_accumulator))
fused_bn.internal_output.output = fused_bn.internal_accumulator.output
layer_quantizer = fused_bn.internal_accumulator.output
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
layer_data_type_map[layer] = {
"input_quantizer_list": input_quantizer_list,
"multiplier": multiplier,
"accumulator": accumulator,
"weight_quantizer": weight_quantizer,
"w_shapes": w_shapes,
"bias_quantizer": bias_quantizer,
"b_shapes": b_shapes,
"bn_inverse_quantizer": inverse_quantizer,
"bn_mean_quantizer": mean_quantizer,
"bn_beta_quantizer": beta_quantizer,
"fused_accumulator": fused_bn.internal_accumulator,
"output_quantizer": output_quantizer,
"output_shapes": output_shapes,
"operation_count": operation_count
}
else:
layer_quantizer = accumulator.output
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, multiplier, accumulator,
weight_quantizer, w_shapes, bias_quantizer, b_shapes,
output_quantizer, output_shapes, operation_count)
elif node_type in ["QConv2DBatchnorm", "QDepthwiseConv2DBatchnorm"]:
# Datatype for Folded Conv/DepthwiseConv layer
# TODO(lishanok): Add additional support for Folded Dense layer
(input_quantizer, _) = input_qe_list[0]
if for_reference or not hasattr(layer, "get_quantizers"):
# For_reference: force all quantizers to keras_quantizer.
weight_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
bias_quantizer = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_quantizer:
weight_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
bias_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_quantizer)
else:
# QKeras layer
qkeras_weight_quantizer = layer.get_quantizers()[0]
qkeras_bias_quantizer = layer.get_quantizers()[1]
if not quantizer_factory.is_quantizer_supported(
qkeras_weight_quantizer):
raise TagMissingError(
"Unsupported weight quantizer {} on this layer: {}".
format(qkeras_weight_quantizer, layer))
if not quantizer_factory.is_quantizer_supported(
qkeras_bias_quantizer):
raise TagMissingError(
"Unsupported bias quantizer {} on this layer: {}".
format(qkeras_bias_quantizer, layer))
weight_quantizer = quantizer_factory.make_quantizer(
qkeras_weight_quantizer)
if qkeras_bias_quantizer:
bias_quantizer = quantizer_factory.make_quantizer(
qkeras_bias_quantizer)
else:
bias_quantizer = None
# TODO(lishanok): During inference, if weight and bias is po2,
# need to update corresponding quantizer type with min and max
# of the constant values
if is_inference:
weights = qtools_util.get_weights(
layer, model_weights_already_quantized)
if weight_quantizer.is_po2:
weight_quantizer.update_inference_values(weights[0])
if bias_quantizer and bias_quantizer.is_po2:
bias_quantizer.update_inference_values(weights[1])
multiplier_factory = quantized_operators.MultiplierFactory()
multiplier = multiplier_factory.make_multiplier(
weight_quantizer, input_quantizer)
if qkeras_weight_quantizer:
qtools_util.adjust_multiplier_for_auto_po2(
multiplier, qkeras_weight_quantizer)
weights = layer.get_weights()
kernel = weights[0]
accumulator_factory = quantized_operators.AccumulatorFactory()
accumulator = accumulator_factory.make_accumulator(
kernel.shape,
multiplier,
use_bias=True if bias_quantizer else False)
if not bias_quantizer:
# Set bias the same as accumulator type.
bias_quantizer = copy.deepcopy(accumulator.output)
if not accumulator.output.is_floating_point:
# For fixed point accumulator, needs to add 1 to its bits to avoid
# possible satuation.
accumulator.output.bits += 1
accumulator.output.int_bits += 1
if for_reference or not hasattr(layer, "get_quantizers"):
accumulator.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
multiplier.output = quantizer_factory.make_default_quantizer(
mode=cfg.default_interm_quantizer)
if keras_accumulator:
accumulator.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
multiplier.output = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
layer_quantizer = accumulator.output
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, layer_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, multiplier, accumulator,
weight_quantizer, w_shapes, bias_quantizer, b_shapes,
output_quantizer, output_shapes, operation_count)
elif node_type:
# Any other unsupported layer types -> pass the input quantizer
# type to output in qraph
(input_quantizer, _) = input_qe_list[0]
if for_reference and keras_accumulator and not is_input_layer:
input_quantizer = quantizer_factory.make_default_quantizer(
mode=keras_accumulator)
output_quantizer = update_output_quantizer_in_graph(
graph, node_id, quantizer_factory, input_quantizer,
for_reference)
layer_data_type_map[layer] = LayerDataType(
input_quantizer_list, None, None, None, None, None, None,
output_quantizer, output_shapes, operation_count)
result = {
"source_quantizer_list": source_quantizer_list,
"output_layers": output_layers,
"input_layers": input_layers,
"layer_data_type_map": layer_data_type_map
}
return result