def get_quantizer_output(self) -> QuantizerOutput:
"""
Quantization information can be stored both in q_output attribute
and in meta attributes
TODO: Merge approaches
"""
if not self.is_quantized():
raise ValueError("No quantization output found. Quantize this"
" XGraph object before retrieving the"
" quantization output")
if (self.quantizer_output is not None
and "is_quantized" in self.meta_attrs
and self.meta_attrs["is_quantized"]):
warnings.warn("Quantization info found both in XGraph meta"
" attributes and q_output attribute")
if self.quantizer_output is not None:
return self.quantizer_output
# Retrieve quantization output from meta attributes
q_output = QuantizerOutput(self.get_name())
if "quant_keys" not in self.meta_attrs:
raise ValueError("Expected `quant_keys` attribute in meta"
" attributes")
for q_key in self.meta_attrs["quant_keys"]:
q_output.add(q_key=q_key,
q_file=self.meta_attrs[q_key]['q_file'],
q_info=self.meta_attrs[q_key]['q_info'],
orig_pb=self.meta_attrs[q_key]['orig_pb'])
logger.debug("QOutput q_info: {}".format(
self.meta_attrs[q_key]['q_info']))
return q_output
def quantize(self, stop=None, subgraphs_only=True):
# type: (str, boolean) -> None
"""
Start quantization of the executable graph model
Arguments
---------
stop: str (optional, default = None)
the name of the operation at which to stop quantization
"""
self._quantize(stop, subgraphs_only)
# quant_files = {}
q_output = QuantizerOutput(self.xgraph.get_name())
for qkey in self._quant_layers.keys():
if qkey != 'None':
quant_file = os.path.join(self.work_dir, qkey + '_quant.json')
self._quant_param.save_to_dpu_v1_json(self._quant_layers[qkey],
quant_file)
q_output.add(qkey, quant_file, None, None)
self.xgraph.set_quantizer_output(q_output)
logger.info("QUANTIZATION DONE")
return self.xgraph
def __init__(self,
xgraph,
inputs_func,
work_dir=os.path.join(os.getcwd(), 'work')):
super(ExternalQuantizer, self).__init__(xgraph, inputs_func, work_dir)
self.gen = TfGenerator()
self.partition_graphs = {}
self.res = {}
self.q_output = QuantizerOutput(name=xgraph.get_name())
def __init__(self,
xgraph,
inputs_func,
work_dir=os.path.join(os.getcwd(), 'work'),
quant_iter=1,
**kwargs):
super(DECENTQuantizer, self).__init__(xgraph, inputs_func, work_dir)
self.quant_iter = quant_iter
self.gen = TfGenerator()
self.partition_graphs = {}
self.res = {}
self.kwargs = kwargs
self.q_output = QuantizerOutput(name=xgraph.get_name())
def __init__(self,
xgraph,
inputs_func,
bitwidth=8,
work_dir=os.path.join(os.getcwd(), 'work'),
quant_iter=1,
mse_opt_num=50):
super(XGraphMSEThresholdQuantizer, self).__init__(xgraph)
self.inputs_func = inputs_func
self.work_dir = work_dir
self.bitwidth = bitwidth
self.mse_opt_num = mse_opt_num
self.quant_xgraph = None
self.runtime = None
self._quant_param = QuantParamFactory()
self._quant_layers = {}
self.q_output = QuantizerOutput(name=xgraph.get_name())