def parse_args(self, **kwargs):
self.kernel = num_as_tuple(cget(kwargs, "kernel", tuple, (1, 1)))
self.stride = num_as_tuple(cget(kwargs, "stride", tuple, (1, 1)))
self.pad = num_as_tuple(
cget(kwargs, "pad", tuple,
(int(self.kernel[0] / 2), int(self.kernel[1] / 2))))
self.pool_type = cget(kwargs, "pool_type", str, 'max')
self.global_pool = cget(kwargs, "global_pool", bool, False)
def get_weight_bias(self, **kwargs):
weight_initializer = cget(kwargs, 'weight_initializer', lambda x: x,
None)
bias_initializer = cget(kwargs, 'bias_initializer', lambda x: x, None)
self.no_bias = cget(kwargs, 'no_bias', bool, False)
self.weight = mx.sym.var(self.name + '_weight',
init=weight_initializer)
self.bias = mx.sym.var(
self.name +
'_bias', init=bias_initializer) if not self.no_bias else None
return self.weight, self.bias, self.no_bias
def __init__(self, **kwargs):
self.layout = cget(kwargs, "layout", str, "NCHW")
self.channel_id = self.layout.find("C")
self.height_id = self.layout.find("H")
self.width_id = self.layout.find("W")
self.time_id = self.layout.find("T")
self.batch_id = self.layout.find("N")
def parse_args(self, **kwargs):
self.kernel = num_as_tuple(cget(kwargs, "kernel", tuple, (1, 1)))
self.stride = num_as_tuple(cget(kwargs, "stride", tuple, (1, 1)))
self.dilate = num_as_tuple(cget(kwargs, "dilate", tuple, (1, 1)))
self.pad = num_as_tuple(
cget(kwargs, "pad", tuple,
(int(self.kernel[0] / 2), int(self.kernel[1] / 2))))
self.num_group = cget(kwargs, "num_group", int, 1)
self.num_filter = cget(kwargs, "num_filter", int, 1)
def __init__(self, **kwargs):
self.mode = cget(kwargs, 'mode', str, 'train')
def parse_args(self, **kwargs):
self.eps = cget(kwargs, "eps", float, 2e-5)
self.fix_gamma = cget(kwargs, "fix_gamma", bool, True)
self.momentum = cget(kwargs, "momentum", float, .9)
self.use_global_stats = cget(kwargs, "use_global_stats", bool, False)
def parse_args(self, **kwargs):
self.prob = cget(kwargs, "prob", float, 0.0)