return self.quantize(input_, self.scale, self.num_bits)
if __name__ == '__main__':
per_tensor_scale_shape = (1, )
for input_name, input_size, gpu_runs in TEST_PARAMS_STRUCT:
weight_per_channel_scale_shape = get_per_channel_scale_shape(
input_size, is_weights=True)
act_per_channel_scale_shape = get_per_channel_scale_shape(
input_size, is_weights=False)
print("CUDA " + input_name)
print("------------------------------------------------")
print("Pytorch Symmetric (cuda 0) impl:")
print("input size: {0}".format(input_size))
run_profile(
ReferenceQuantize(NBITS).cuda(), input_size, 'cuda', gpu_runs)
print()
print("Custom Symmetric (cuda 0 ) impl:")
print("input size: {0}".format(input_size))
run_profile(
SymmetricQuantizer(
DefaultedPTQuantizerSpec(scale_shape=per_tensor_scale_shape,
num_bits=NBITS)).cuda(), input_size,
'cuda', gpu_runs)
print()
print("Pytorch Symmetric Per Weight Channel (cuda 0) impl:")
print("input size: {0}".format(input_size))
run_profile(
ReferenceQuantize(NBITS,
self.quantize = ReferenceQuantizeSymmetric.apply
def get_scale(self):
return self.scale
def forward(self, input_):
return self.quantize(input_, self.scale, self.num_bits)
if __name__ == '__main__':
for input_name, input_size, gpu_runs in TEST_PARAMS_STRUCT:
print("CUDA " + input_name)
print("------------------------------------------------")
print("Pytorch Symmetric (cuda 0) impl:")
print("input size: {0}".format(input_size))
run_profile(
ReferenceQuantize(NBITS).cuda(), input_size, 'cuda', gpu_runs)
print()
print("Custom Symmetric (cuda 0 ) impl:")
print("input size: {0}".format(input_size))
run_profile(
SymmetricQuantizer(QuantizerConfig(
QuantizationParams(bits=NBITS))).cuda(), input_size, 'cuda',
gpu_runs)
print()
print("Pytorch Symmetric Per Weight Channel (cuda 0) impl:")
print("input size: {0}".format(input_size))
run_profile(
ReferenceQuantize(NBITS,
input_shape=input_size,
def forward(self, input_):
return ReferenceActivationBinarize.apply(input_, self.scale,
self.threshold)
if __name__ == '__main__':
for input_name, input_size, gpu_runs in TEST_PARAMS_STRUCT:
print()
print("CUDA " + input_name)
print("------------------------------------------------")
print("Pytorch XNOR weight binarization (cuda 0) impl:")
print("input size: {0}".format(input_size))
run_profile(ReferenceWeightBinarizationModule('xnor').cuda(),
input_size,
'cuda',
gpu_runs,
forward_only=True)
print()
print("Custom XNOR weight binarization (cuda 0) impl:")
print("input size: {0}".format(input_size))
run_profile(XNORBinarize(enabled=True).cuda(),
input_size,
'cuda',
gpu_runs,
forward_only=True)
print()
print("Pytorch DoReFa weight binarization (cuda 0) impl:")
print("input size: {0}".format(input_size))