def do_qshow(self, args):
"""
Show current quantization settings."""
self._check_graph()
self._check_quantized()
tab = QuantizationReporter(step=args.step).report(
self.G, self.G.quantization)
output_table(tab, args)
def test_simple_quantization(mnist_graph, mnist_images):
G = create_graph(mnist_graph, opts={"load_tensors": True})
G.add_dimensions()
input_tensor = import_data(mnist_images[0],
height=28,
width=28,
offset=0,
divisor=255)
input_tensor = input_tensor.reshape((28, 28, 1))
stats_collector = ActivationStatsCollector()
stats_collector.collect_stats(G, [input_tensor])
astats = stats_collector.reduce_stats()
stats_collector = FilterStatsCollector()
fstats = stats_collector.collect_stats(G)
quantizer = SymmetricQuantizer(astats, fstats, force_width=8)
qrecs = quantizer.quantize(G)
assert len(qrecs) == 11 # One more for saved quantizer
report = QuantizationReporter().report(G, qrecs)
renderer = TextTableRenderer(maxwidth=200)
print(report.render(renderer))
def do_qshow(self, args):
"""
Show current quantization settings."""
self._check_graph()
self._check_quantized()
if args.step:
nodes, _ = self.get_node_step_or_name(args.step)
if not nodes:
return
else:
nodes = None
tab = QuantizationReporter().report(self.G, self.G.quantization, nodes)
output_table(tab, args)
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 = ActivationStatsCollector()
input_tensors = [
np.random.normal(0, 0.2, input.dims.shape)
for input in self.G.input_nodes()
]
stats_collector.collect_stats(self.G, input_tensors)
astats = stats_collector.reduce_stats()
stats_collector = FakeFilterStatsCollector()
fstats = stats_collector.collect_stats(self.G)
quantizer = SymmetricQuantizer(astats,
fstats,
force_width=args.force_width)
qrecs = quantizer.quantize(self.G)
self.G.quantization = qrecs
tab = QuantizationReporter().report(self.G, qrecs)
output_table(tab, args)
self.G.constant_store.fake = False