def do_aquant(self, args: argparse.Namespace):
"""
Attempt to calculate quantization for graph using one or more sample input files."""
self._check_graph()
stats_collector = ActivationRangesCollector()
# if replaying state file then load the activation stats if they are present
if args.scheme == 'SQ8':
bits = 8
else:
bits = args.force_width
if self.replaying_history and self.history_stats:
astats = self.history_stats
else:
input_args = self._get_input_args(args)
processed_input = False
for file_per_input in glob_input_files(args.input_files,
self.G.num_inputs):
LOG.info("input file %s", file_per_input)
processed_input = True
data = [
import_data(input_file, **input_args)
for input_file in file_per_input
]
stats_collector.collect_stats(self.G, data)
if not processed_input:
self.perror("No input files found")
return
astats = stats_collector.stats
self._record_stats(astats)
quantizer = UnifiedQuantizer(args.scheme,
astats,
quantized_dimension=args.quant_dimension,
narrow_weights=not args.no_narrow_weights,
bits=bits)
qrecs = quantizer.quantize(self.G)
self.G.quantization = qrecs
# These should now be unnecessary
# if args.scheme == 'SQ8':
# concats_matcher = EqualizeSymmetricMultiplicativeQuantivedConcats()
# concats_matcher.match(self.G, set_identity=False)
# rnns_matcher = PropagateUpRNNInputQ()
# rnns_matcher.match(self.G, set_identity=False)
# softmax_qrec_matcher = PropagateSoftmaxSymQrec()
# softmax_qrec_matcher.match(self.G, set_identity=False)
# sig_swish_qrec_matcher = PropagateUpSigSwishInputQ()
# sig_swish_qrec_matcher.match(self.G, set_identity=False)
LOG.info("Quantization set. Use qshow command to see it.")
def do_fquant(self, args: argparse.Namespace):
"""
Attempt to calculate a fake quantization for graph using random tensors and parameters.
This is intended to allow code generation for performance testing even if no real
weights and input data are avalaible."""
self._check_graph()
self.G.constant_store.fake = True
stats_collector = ActivationRangesCollector()
for _ in range(args.num_inference):
if args.uniform:
input_tensors = [
np.random.uniform(-args.uniform, args.uniform,
inp.dims.shape)
for inp in self.G.input_nodes()
]
else:
input_tensors = [
np.random.normal(0, 0.2, inp.dims.shape)
for inp in self.G.input_nodes()
]
stats_collector.collect_stats(self.G, input_tensors)
if args.scheme == 'SQ8':
bits = 8
else:
bits = args.force_width
astats = stats_collector.stats
quantizer = UnifiedQuantizer(args.scheme,
astats,
quantized_dimension=args.quant_dimension,
narrow_weights=not args.no_narrow_weights,
bits=bits)
self._record_stats(astats)
qrecs = quantizer.quantize(self.G)
self.G.quantization = qrecs
if args.scheme == 'SQ8':
concats_matcher = EqualizeSymmetricMultiplicativeQuantivedConcats()
concats_matcher.match(self.G, set_identity=False)
softmax_qrec_matcher = PropagateSoftmaxSymQrec()
softmax_qrec_matcher.match(self.G, set_identity=False)
self.G.constant_store.fake = False
def do_aquant(self, args: argparse.Namespace):
"""
Attempt to calculate quantization for graph using one or more sample input files."""
self._check_graph()
stats_collector = ActivationRangesCollector()
# if replaying state file then load the activation stats if they are present
opts = get_options_from_args(args)
if self.replaying_history and self.history_stats:
astats = self.history_stats
else:
input_args = self._get_input_args(args)
processed_input = False
for file_per_input in glob_input_files(args.input_files,
self.G.num_inputs):
LOG.info("input file %s", file_per_input)
processed_input = True
data = [
import_data(input_file, **input_args)
for input_file in file_per_input
]
stats_collector.collect_stats(self.G, data)
if not processed_input:
self.perror("No input files found")
return
astats = stats_collector.stats
self._record_stats(astats)
if args.force_width:
opts['bits'] = args.force_width
quantizer = UnifiedQuantizer(args.scheme, astats, **opts)
# clear the existing quantization
self.G.quantization = None
qrecs = quantizer.quantize(self.G)
self.G.quantization = qrecs
RemoveUnnecessaryQuantizeOperators().match(self.G)
self.G.add_dimensions()
LOG.info("Quantization set. Use qshow command to see it.")
def do_fquant(self, args: argparse.Namespace):
"""
Attempt to calculate a fake quantization for graph using random tensors and parameters.
This is intended to allow code generation for performance testing even if no real
weights and input data are avalaible."""
self._check_graph()
opts = get_options_from_args(args)
if self.replaying_history and self.history_stats:
astats = self.history_stats
else:
self.G.constant_store.fake = True
stats_collector = ActivationRangesCollector()
for _ in range(args.num_inference):
if args.uniform:
input_tensors = [np.random.uniform(-args.uniform, args.uniform, inp.dims.shape)
for inp in self.G.input_nodes()]
else:
input_tensors = [np.random.normal(0, 0.2, inp.dims.shape)
for inp in self.G.input_nodes()]
stats_collector.collect_stats(self.G, input_tensors)
astats = stats_collector.stats
self._record_stats(astats)
self.G.constant_store.fake = False
if args.force_width:
opts['bits'] = args.force_width
quantizer = UnifiedQuantizer(args.scheme, astats,
**opts)
# clear the existing quantization
self.G.quantization = None
qrecs = quantizer.quantize(self.G)
self.G.quantization = qrecs
RemoveUnnecessaryQuantizeOperators().match(self.G)
self.G.add_dimensions()
LOG.info("Quantization set. Use qshow command to see it.")