def round_shift_weights(model, shift_base=2, clone=False):
if (clone):
model = copy.deepcopy(model)
for name, module in reversed(model._modules.items()):
if len(list(module.children())) > 0:
# recurse
model._modules[name] = round_shift_weights(model=module,
shift_base=shift_base)
if type(module) == deepshift.modules.LinearShift or type(
module) == deepshift.modules.Conv2dShift:
module.shift.data = module.shift.round()
module.sign.data = module.sign.round().sign()
if (module.bias is not None):
module.bias.data = utils.round_to_fixed(module.bias,
fraction_bits=16,
integer_bits=16)
elif type(module) == deepshift.modules_q.LinearShiftQ or type(
module) == deepshift.modules_q.Conv2dShiftQ:
module.weight.data = utils.round_power_of_2(
module.weight, shift_base)
if (module.bias is not None):
module.bias.data = utils.round_to_fixed(module.bias,
fraction_bits=16,
integer_bits=16)
return model
def round_shift_weights(model, clone=False):
# for name, param in model.named_parameters():
# if param.is_leaf:
# print(name)
model.eval()
if(clone):
model = copy.deepcopy(model)
for name, module in reversed(model._modules.items()):
if len(list(module.children())) > 0:
# recurse
model._modules[name] = round_shift_weights(model=module)
if type(module) == deepshift.modules.LinearShift or type(module) == deepshift.modules.Conv2dShift:
module.shift.data = module.shift.round()
module.sign.data = module.sign.round().sign()
if (module.bias is not None):
module.bias.data = utils.round_to_fixed(module.bias, fraction=16, integer=16)
elif type(module) == deepshift.modules_q.LinearShiftQ or type(module) == deepshift.modules_q.Conv2dShiftQ:
module.weight.data = utils.round_power_of_2(module.weight)
if (module.bias is not None):
module.bias.data = utils.round_to_fixed(module.bias, fraction=16, integer=16)
return model
def round_shift_weights(model, clone=False, weight_bits=5, act_integer_bits=16, act_fraction_bits=16):
if(clone):
model = copy.deepcopy(model)
for name, module in reversed(model._modules.items()):
if len(list(module.children())) > 0:
# recurse
model._modules[name] = round_shift_weights(model=module, weight_bits=weight_bits, act_integer_bits=act_integer_bits, act_fraction_bits=act_fraction_bits)
if type(module) == deepshift.modules.LinearShift or type(module) == deepshift.modules.Conv2dShift:
module.shift.data = module.shift.round()
module.sign.data = module.sign.round().sign()
if (module.bias is not None):
module.bias.data = utils.round_to_fixed(module.bias, integer_bits=act_integer_bits, fraction_bits=act_fraction_bits)
elif type(module) == deepshift.modules_q.LinearShiftQ or type(module) == deepshift.modules_q.Conv2dShiftQ:
module.weight.data = utils.clampabs(module.weight.data, 2**module.shift_range[0], 2**module.shift_range[1])
module.weight.data = utils.round_power_of_2(module.weight)
if (module.bias is not None):
module.bias.data = utils.round_to_fixed(module.bias, integer_bits=act_integer_bits, fraction_bits=act_fraction_bits)
return model
def forward(ctx, input, z):
return utils.round_power_of_2(input, z[0], z[1])
def forward(ctx, input, stochastic=False):
return utils.round_power_of_2(input, stochastic)
def forward(ctx, input):
return utils.round_power_of_2(input)