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)
state = ConstantInputParameters.save_compression_state(self.G)
try:
if self.replaying_history and self.history_stats:
astats = self.history_stats
else:
if args.seed:
np.random.seed(args.seed)
ConstantInputParameters.fake(self.G, 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, args.normal, 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)
ConstantInputParameters.fake(self.G, False)
if args.force_width:
opts['bits'] = args.force_width
quantizer = NewQuantizer(self.G, reset_all=True)
quantizer.schemes.append(args.scheme)
quantizer.set_stats(current_stats=astats, current_options=opts)
quantizer.quantize()
LOG.info("Quantization set. Use qshow command to see it.")
finally:
ConstantInputParameters.restore_compression_state(self.G, state)
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)
state = ConstantInputParameters.save_compression_state(self.G)
try:
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 = NewQuantizer(self.G, reset_all=True)
quantizer.schemes.append(args.scheme)
quantizer.set_stats(astats, opts)
quantizer.quantize()
self.G.add_dimensions()
LOG.info("Quantization set. Use qshow command to see it.")
finally:
ConstantInputParameters.restore_compression_state(self.G, state)
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.")
def qtune_set_options(self, nodes, node_descr, args):
self._check_quantized()
options = get_options_from_args(args)
tune_options(self.G, nodes, options)
self.pfeedback(f'set options on {node_descr}')