def test_graph_imu_auto_quant_and_execute_quant():
G = create_graph("tests/graph/imu.tflite", opts={"load_tensors": True})
G.add_dimensions()
G.adjust_order()
get_pow2_match_group().match(G)
G.add_dimensions()
stats_collector = ActivationStatsCollector()
for input_file in ['tests/images/imu0.pgm']:
input_tensor = import_data(input_file,
offset=0,
divisor=256,
nptype='int16')
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=16)
qrecs = quantizer.quantize(G)
G.quantization = qrecs
executer = GraphExecuter(G, qrecs=qrecs)
for input_file in ['tests/images/imu0.pgm']:
input_tensor = import_data(input_file,
offset=0,
divisor=256,
nptype='int16')
output_ = executer.execute([input_tensor],
qmode=QuantizationMode.all())
def test_validate_mn1_quantized1(mn1q_graph, mn1f_graph):
tfi = TfliteImporter()
Gf = tfi.create_graph(mn1f_graph, {'load_tensors': True})
Gf.add_dimensions()
Gf.adjust_order()
matcher = get_pow2_match_group()
matcher.match(Gf)
Gf.add_dimensions()
tfi = TfliteImporter()
G = tfi.create_graph(mn1q_graph, {
'load_tensors': True,
'load_quantization': True
})
G.add_dimensions()
G.adjust_order()
matcher = get_pow2_match_group()
matcher.match(G)
G.add_dimensions()
fpnode = Gf.graph_state.steps[2]['node']
fpcnode = fpnode.contained_filters()[0]
qpnode = G.graph_state.steps[2]['node']
qpcnode = qpnode.contained_filters()[0]
nid = NodeId(qpnode, qpcnode)
qrec = G.quantization[nid]
dqbiases = qrec.biases_q.get_dequantized(qpcnode.biases)
assert np.max(np.abs(fpcnode.biases - dqbiases)) < 0.1
input_tensor = np.load('tests/mobv1_valid/COCO_val2014_000000362331_0.npy')
input_tensor = input_tensor.reshape((224, 224, 3)).transpose((2, 0, 1))
executer = GraphExecuter(Gf)
foutput_tensors = executer.execute([input_tensor])
foutput_tensor = np.load(
'tests/mobv1_valid/output_COCO_val2014_000000362331_0_float.npy')
assert np.max(np.abs(foutput_tensors[-1][0] - foutput_tensor[0])) < 0.0001
executer = GraphExecuter(G, qrecs=G.quantization)
qfroutput_tensors = executer.execute([input_tensor],
qmode=QuantizationMode.none())
assert np.max(np.abs(qfroutput_tensors[-1][0] - foutput_tensor[0])) < 0.2
executer = GraphExecuter(G, qrecs=G.quantization)
qroutput_tensors = executer.execute(
[input_tensor], qmode=QuantizationMode.all_dequantize())
output_tensor = np.load(
'tests/mobv1_valid/output_COCO_val2014_000000362331_0_quant.npy')
# assert np.max(np.abs(qroutput_tensors[-1][0] - output_tensor[0])) < 0.16
assert np.max(np.abs(qroutput_tensors[-1][0] - output_tensor[0])) < 0.28
def test_fake_values_concat(concat_test_graph):
G = create_graph(concat_test_graph, opts={"load_tensors": True})
G.add_dimensions()
G.adjust_order()
matcher = get_pow2_match_group()
matcher.match(G)
G.add_dimensions()
G.constant_store.fake = True
stats_collector = ActivationStatsCollector()
stats_collector.collect_stats(
G, [np.random.rand(*node.dims.shape) for node in G.input_nodes()])
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)
G.quantization = qrecs
with tempfile.TemporaryDirectory() as tempdir:
opts = {
'default_input_location': 'ARG_LOC_L2',
'default_output_location': 'ARG_LOC_L2',
'default_global_location': 'ARG_LOC_L3_HFLASH',
'default_local_location': 'AT_MEM_UNDEF',
'at_ver': 3,
'tensor_directory': tempdir
}
code_gen = CodeGenerator(G, DefaultNamingConvension(G), opts)
print(default_template(G, code_generator=code_gen))
code_gen.write_constants()
def test_graph_kws_auto_quant(kws_graph, kws_sounds):
G = create_graph(kws_graph, opts={"load_tensors": True})
G.add_dimensions()
G.adjust_order()
get_pow2_match_group().match(G)
G.add_dimensions()
stats_collector = ActivationStatsCollector()
for input_file in kws_sounds:
data = import_data(input_file, offset=0, divisor=256, nptype='int16')
stats_collector.collect_stats(G, [data])
astats = stats_collector.reduce_stats()
stats_collector = FilterStatsCollector()
fstats = stats_collector.collect_stats(G)
quantizer = SymmetricQuantizer(astats, fstats, force_width=16)
qrecs = quantizer.quantize(G)
G.quantization = qrecs
def do_fusions(self, args):
"""
Carry out the default set of fusions on the graph"""
self._check_graph()
if args.list:
self.ppaged("\n".join(
["%s - %s" % (name, desc) for name, desc in get_fusions()]))
return
if args.apply:
fusions = [get_fusion(name) for name in args.apply]
if not fusions:
self.perror('fusion %s not found' % args.apply)
return
elif args.pow2:
fusions = [get_pow2_match_group()]
elif args.scale8:
fusions = [get_scale8_match_group()]
else:
self.perror(
"No fusion set selected. Nothing to do. Select --pow2 or --scale8."
)
return
for fusion in fusions:
fusion.match(self.G)
self.G.add_dimensions()
if self.G.quantization and not self.G.quantization.verify_quantization(
self.G):
self.G.quantization = None
def test_fusions4(ssd_graph):
tfi = TfliteImporter()
G = tfi.create_graph(ssd_graph, {})
G.add_dimensions()
matcher = get_pow2_match_group()
matcher.match(G)
G.add_dimensions()
def test_validate_mn1_quantized2(mn1q_graph):
tfi = TfliteImporter()
G = tfi.create_graph(mn1q_graph, {
'load_tensors': True,
'load_quantization': True
})
G.add_dimensions()
G.adjust_order()
matcher = get_pow2_match_group()
matcher.match(G)
G.add_dimensions()
def test_validate_mn1_dequant_quantfloat(mn1q_graph):
# load dequantized graph same results as quant graph and float execution
tfi = TfliteImporter()
G = tfi.create_graph(mn1q_graph, {
'load_tensors': True,
'load_quantization': True
})
G.add_dimensions()
G.adjust_order()
matcher = get_pow2_match_group()
matcher.match(G)
G.add_dimensions()
Gdq = tfi.create_graph(mn1q_graph, {
'load_tensors': True,
'load_dequantized': True
})
Gdq.add_dimensions()
Gdq.adjust_order()
matcher = get_pow2_match_group()
matcher.match(Gdq)
Gdq.add_dimensions()
input_tensor = np.load('tests/mobv1_valid/COCO_val2014_000000362331_0.npy')
input_tensor = input_tensor.reshape((224, 224, 3)).transpose((2, 0, 1))
executer = GraphExecuter(G, qrecs=G.quantization)
qfoutput_tensors = executer.execute([input_tensor],
qmode=QuantizationMode.none())
executer = GraphExecuter(Gdq)
dfoutput_tensors = executer.execute([input_tensor])
diff_list = [
np.abs(df[0] - qf[0])
for df, qf in zip(dfoutput_tensors, qfoutput_tensors)
]
max_diff = [np.max(elem) for elem in diff_list]
assert max(max_diff) < 0.003
def save_state(temp_dir, width, fusions=False, adjust=False):
file_name = os.path.join(temp_dir, "state_file")
G = create_graph(MNIST_GRAPH, opts={"load_tensors": True})
G.add_dimensions()
if adjust:
G.adjust_order()
if fusions:
get_pow2_match_group().match(G)
G.add_dimensions()
stats_collector = ActivationStatsCollector()
for input_file in MNIST_IMAGES:
data = import_data(input_file, offset=0, divisor=255)
if not adjust:
data = data.reshape((28, 28, 1))
stats_collector.collect_stats(G, [data])
astats = stats_collector.reduce_stats()
stats_collector = FilterStatsCollector()
fstats = stats_collector.collect_stats(G)
quantizer = SymmetricQuantizer(astats, fstats, force_width=width)
qrecs = quantizer.quantize(G)
G.quantization = qrecs
dump_state(G, include_parameters=True, state_path=file_name)
return file_name
def test_validate_mn1_float(mn1f_graph):
tfi = TfliteImporter()
G = tfi.create_graph(mn1f_graph, {'load_tensors': True})
G.add_dimensions()
matcher = get_pow2_match_group()
matcher.match(G)
G.add_dimensions()
input_tensor = np.load('tests/mobv1_valid/COCO_val2014_000000362331_0.npy')
input_tensor = input_tensor.reshape((224, 224, 3))
executer = GraphExecuter(G, qrecs=G.quantization)
routput_tensors = executer.execute([input_tensor])
output_tensor = np.load(
'tests/mobv1_valid/output_COCO_val2014_000000362331_0_float.npy')
assert np.max(np.abs(routput_tensors[-1][0] - output_tensor[0])) < 0.0001
def test_adjust7(concat_test_graph):
tfi = TfliteImporter()
G = tfi.create_graph(concat_test_graph, {'load_tensors': True})
G.node('input_1').fixed_order = True
G.node('output_1').fixed_order = True
G.node('output_2').fixed_order = True
G.add_dimensions()
G.adjust_order()
matcher = get_pow2_match_group()
matcher.match(G)
G.add_dimensions()
report = GraphReporter().report(G, None)
renderer = TextTableRenderer(maxwidth=200)
print(report.render(renderer))
report = GraphReporter(split_dims=True).report(G, None)
def do_fusions(self, args):
"""
Carry out the default set of fusions on the graph"""
if args.list:
table = texttable.Texttable()
table.set_cols_align(['l', 'l'])
table.set_max_width(120)
table.add_rows([['Name', 'Description']] + get_fusions())
self.ppaged(table.draw())
return
self._check_graph()
state = ConstantInputParameters.save_compression_state(self.G)
try:
if args.apply:
fusions = [get_fusion(name) for name in args.apply]
invalid_names = [
args.apply[idx] for idx, fusion in enumerate(fusions)
if fusion is None
]
if invalid_names:
self.perror(
f'fusion{"s" if len(invalid_names) > 1 else ""} {", ".join(invalid_names)} not found'
)
return
elif args.pow2:
fusions = [get_pow2_match_group()]
elif args.scale8:
fusions = [get_scale8_match_group()]
else:
self.perror(
"No fusion set selected. Nothing to do. Select --pow2 or --scale8."
)
return
for fusion in fusions:
fusion.match(self.G)
self.G.add_dimensions()
if self.G.quantization and verify_quantization(self.G):
quantizer = NewQuantizer(self.G)
quantizer.quantize()
problems = verify_quantization(self.G)
if problems:
self.perror('quantization issue after fusions')
for problem in problems:
self.perror(problem)
finally:
ConstantInputParameters.restore_compression_state(self.G, state)
def do_fusions(self, args):
"""
Carry out the default set of fusions on the graph"""
if args.list:
table = texttable.Texttable()
table.set_cols_align(['l', 'l'])
table.set_max_width(120)
table.add_rows([['Name', 'Description']] + get_fusions())
self.ppaged(table.draw())
return
self._check_graph()
if args.apply:
fusions = [get_fusion(name) for name in args.apply]
invalid_names = [
args.apply[idx] for idx, fusion in enumerate(fusions)
if fusion is None
]
if invalid_names:
self.perror(
f'fusion{"s" if len(invalid_names) > 1 else ""} {", ".join(invalid_names)} not found'
)
return
elif args.pow2:
fusions = [get_pow2_match_group()]
elif args.scale8:
fusions = [get_scale8_match_group()]
else:
self.perror(
"No fusion set selected. Nothing to do. Select --pow2 or --scale8."
)
return
for fusion in fusions:
fusion.match(self.G)
self.G.add_dimensions()
if self.G.quantization and not self.G.quantization.verify_quantization(
self.G):
self.G.quantization = None