def __init__(self,
units,
activation=None,
use_bias=True,
kernel_initializer='glorot_uniform',
bias_initializer='zeros',
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
bias_constraint=None,
**kwargs):
super().__init__(units,
activation=activation,
use_bias=use_bias,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer,
activity_regularizer=activity_regularizer,
kernel_constraint=kernel_constraint,
bias_constraint=bias_constraint,
**kwargs)
self.quant_mode, self.quantizer = maybe_get_quantizer()
if self.quant_mode > 0:
self._quant_vars = {}
self.valid_inputs = None
self.valid_output = None
self.quant_info = None
self.params_name = None
self.node = None
self.quant_vars_initialized = False
self.params_quantized = False
def __init__(self, *args, **kwards):
super(deephi_BatchNorm, self).__init__(*args, **kwards)
self.params_name = None
self.node = None
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.param_saved = False
self.param_quantized = False
def __init__(self, *args, **kwards):
super(deephi_MaxPool2d, self).__init__(*args, **kwards)
self.valid_inputs = None
self.valid_output = None
self.node = None
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.need_quant_output = True
def __init__(self, *args, **kwargs):
# only support the specified slope and inplace operation
super().__init__(*args, **kwargs)
if NndctOption.nndct_leaky_relu_approximate.value:
self.negative_slope = 0.1015625
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.node = None
def __init__(self, *args, **kwards): super(deephi_Interpolate, self).__init__(*args, **kwards) self.valid_inputs = None self.valid_output = None self.params_name = None self.node = None self.quant_mode, self.quantizer = maybe_get_quantizer()
def __init__(self): super(deephi_Sub, self).__init__() self.quant_mode, self.quantizer = maybe_get_quantizer() self.valid_inputs = None self.valid_output = None self.quant_info = None self.params_name = None self.node = None
def __init__(self, op_type, *args, **kwards):
super().__init__()
self.valid_inputs = None
self.valid_output = None
self.node = None
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.need_quant_output = True
self.op_type = op_type
def Mean(*args, **kwargs):
#quant_mode,_ = maybe_get_quantizer()
#if quant_mode==None:
# return
quant_mode, _ = maybe_get_quantizer()
if quant_mode is None or NndctOption.nndct_quant_off.value:
return torch.mean
return deephi_Mean(*args, **kwargs)
def __init__(self, *args, **kwargs): super(deephi_ReLU, self).__init__(*args, **kwargs) self.quant_mode, self.quantizer = maybe_get_quantizer() self.valid_inputs = None self.valid_output = None self.params_name = None self.node = None self.need_quant_output = False
def __init__(self, caller, *args, **kwards):
super().__init__()
self.valid_inputs = None
self.valid_output = None
self.node = None
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.need_quant_output = True
self.caller = caller
self._match_inputs = []
def __init__(self, *args, **kwargs):
super(deephi_Cat, self).__init__()
# self.dim = kwargs.get('dim', 0)
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.valid_inputs = None
self.valid_output = None
self.params_name = None
self.node = None
self.need_quant_output = True
def __init__(self):
super().__init__()
self.quant_mode, self.quantizer = maybe_get_quantizer()
if self.quant_mode > 0:
self._quant_vars = {}
self.valid_inputs = None
self.valid_output = None
self.quant_info = None
self.params_name = None
self.node = None
def __init__(self, *args, **kwards):
super(deephi_Linear, self).__init__(*args, **kwards)
self.valid_inputs = None
self.valid_output = None
self.params_name = None
self.bias_valid_inputs = None
self.bias_valid_output = None
self.node = None
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.param_quantized = False
self.need_quant_output = True
def __init__(self, **kwargs):
super().__init__(**kwargs)
self.quant_mode, self.quantizer = maybe_get_quantizer()
if self.quant_mode and self.quant_mode > 0:
self._quant_vars = {}
self.valid_inputs = None
self.valid_output = None
self.quant_info = None
self.params_name = None
self.node = None
self.quant_vars_initialized = False
self.params_quantized = False
def BatchNorm(*args, **kwargs):
quant_mode, _ = maybe_get_quantizer()
if quant_mode == None:
def _check_input_dim(self, input):
pass
import types
nn = torch.nn.modules.batchnorm._BatchNorm(*args, **kwargs)
nn._check_input_dim = types.MethodType(_check_input_dim, nn)
return nn
return deephi_BatchNorm(*args, **kwargs)
def __init__(self, *args, **kwards):
super(deephi_Conv2d, self).__init__(*args, **kwards)
self.params_name = None
self.node = None
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.param_saved = False
self.param_quantized = False
self.weight_f = None
self.bias_f = None
self.err = None
self.stop = False
self.efficency = 0.0
self.deviation = 0.0
def __init__(self, *args, **kwards):
super(deephi_Linear, self).__init__(*args, **kwards)
self.params_name = None
self.node = None
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.param_saved = False
self.param_quantized = False
# self.weight and self.bias are not quantized float parameters
self.weight_bak = None # backup of float bias for bias correction
self.bias_bak = None # backup of float bias for bias correction
self.stop = False
self.rate = NndctOption.nndct_param_corr_rate.value
self.efficency = 0.0
self.deviation = 0.0
def MaxPool2d(*args, **kwargs):
quant_mode, _ = maybe_get_quantizer()
if quant_mode == None:
return torch.nn.MaxPool2d(*args, **kwargs)
return deephi_MaxPool2d(*args, **kwargs)
def __init__(self, *args, **kwards): super(deephi_MaxPool2d, self).__init__(*args, **kwards) self.node = None self.quant_mode, self.quantizer = maybe_get_quantizer()
def __init__(self, *args, **kwargs):
super(deephi_Cat, self).__init__()
# self.dim = kwargs.get('dim', 0)
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.node = None
def ConvTranspose2d(*args, **kwargs):
quant_mode, _ = maybe_get_quantizer()
if quant_mode == None:
return torch.nn.ConvTranspose2d(*args, **kwargs)
return deephi_ConvTranspose2d(*args, **kwargs)
def Sigmoid(*args, **kwargs):
quant_mode, _ = maybe_get_quantizer()
if quant_mode == None:
return torch.nn.Sigmoid(*args, **kwargs)
return deephi_Sigmoid(*args, **kwargs)
def __init__(self):
super(deephi_Tanh, self).__init__()
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.node = None
def LeakyReLU(*args, **kwargs):
quant_mode, _ = maybe_get_quantizer()
if quant_mode is None or NndctOption.nndct_quant_off.value:
return torch.nn.LeakyReLU(*args, **kwargs)
return deephi_LeakyReLU(*args, **kwargs)
def __init__(self, *args, **kwargs): # only support the specified slope and inplace operation super().__init__(*args, **kwargs) self.negative_slope = 0.1015625 self.quant_mode, self.quantizer = maybe_get_quantizer() self.node = None
def __init__(self, caller, *args, **kwards):
super().__init__()
self.node = None
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.caller = caller
def Linear(*args, **kwargs):
quant_mode, _ = maybe_get_quantizer()
if quant_mode == None:
return torch.nn.Linear(*args, **kwargs)
return deephi_Linear(*args, **kwargs)
def __init__(self, inplace=False, *args, **kwards):
super(deephi_Hardswish, self).__init__()
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.node = None
self.inplace = inplace
def Tanh(*args, **kwargs):
quant_mode, _ = maybe_get_quantizer()
if quant_mode == None:
return torch.nn.Tanh(*args, **kwargs)
return deephi_Tanh(*args, **kwargs)
def __init__(self, dim=None):
super(deephi_Softmax, self).__init__()
self.dim = dim
self.quant_mode, self.quantizer = maybe_get_quantizer()
self.node = None
