class ConvThenLinear(_Net):
def __init__(self, start_shape, stop_shape, depth=2, conv_part=Conv1dAC):
_Net.__init__(self, start_shape, stop_shape, depth)
self.conv_part = conv_part(start_shape=start_shape,
stop_shape=stop_shape,
depth=depth,
bias_prob=0.3)
self.lin_part = LinearAC(start_shape=None,
stop_shape=(10, ),
depth=depth,
bias_prob=0.3)
self.change_num = IntParam(name="")
def change_when(self, x):
self.change_num.unrandomize(val=x)
def change_random(self):
self.change_num.randomize(limits=(1, self.depth - 1))
def generate(self):
self.change_random()
t = self.change_num.value
self.lin_part.depth = self.depth - t
cp = self.conv_part.generate()[:t]
self.lin_part.start_shape = cp[-1].out_shape
lp = self.lin_part.generate()
return cp + lp
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"EmbeddingBag",
[
"num_embeddings",
"embedding_dim",
"max_norm",
"norm_type",
"scale_grad_by_freq",
"mode",
"sparse",
"_weight",
],
)
self.params = self.template_fn(
num_embeddings=IntParam(name="num_embeddings", default=1),
embedding_dim=IntParam(name="embedding_dim", default=1),
max_norm=Param(name="max_norm", default=None),
norm_type=IntParam(name="norm_type", default=2.0),
scale_grad_by_freq=BinaryParam(name="scale_grad_by_freq",
default=False,
true_prob=0.5),
mode=ChoiceParam(name="mode",
choices=("mean", ),
cprobs=(1, ),
default="mean"),
sparse=BinaryParam(name="sparse", default=False, true_prob=0.5),
_weight=Param(name="_weight", default=None),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"GRU",
[
"input_size",
"hidden_size",
"num_layers",
"bias",
"batch_first",
"dropout",
"bidirectional",
],
)
self.params = self.template_fn(
input_size=IntParam(name="input_size", default=1),
hidden_size=Param(name="hidden_size", default=None),
num_layers=IntParam(name="num_layers", default=1),
bias=BinaryParam(name="bias", default=True, true_prob=0.5),
batch_first=BinaryParam(name="batch_first",
default=False,
true_prob=0.5),
dropout=IntParam(name="dropout", default=0.0),
bidirectional=BinaryParam(name="bidirectional",
default=False,
true_prob=0.5),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple("Flatten", ["start_dim", "end_dim"])
self.params = self.template_fn(
start_dim=IntParam(name="start_dim", default=1),
end_dim=IntParam(name="end_dim", default=-1),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple("BasicBlock", ["inplanes", "planes"])
self.params = self.template_fn(
inplanes=IntParam(name="inplanes", default=1),
planes=IntParam(name="planes", default=1),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple("Linear",
["in_features", "out_features", "bias"])
self.params = self.template_fn(
in_features=IntParam(name="in_features", default=1),
out_features=IntParam(name="out_features", default=1),
bias=BinaryParam(name="bias", default=True, true_prob=0.5),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, start_shape, stop_shape, depth=2, conv_part=Conv1dAC):
_Net.__init__(self, start_shape, stop_shape, depth)
self.conv_part = conv_part(start_shape=start_shape,
stop_shape=stop_shape,
depth=depth,
bias_prob=0.3)
self.lin_part = LinearAC(start_shape=None,
stop_shape=(10, ),
depth=depth,
bias_prob=0.3)
self.change_num = IntParam(name="")
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"Unfold", ["kernel_size", "dilation", "padding", "stride"])
self.params = self.template_fn(
kernel_size=Param(name="kernel_size", default=None),
dilation=IntParam(name="dilation", default=1),
padding=IntParam(name="padding", default=0),
stride=IntParam(name="stride", default=1),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"ConvTranspose3d",
[
"in_channels",
"out_channels",
"kernel_size",
"stride",
"padding",
"output_padding",
"groups",
"bias",
"dilation",
"padding_mode",
],
)
self.params = self.template_fn(
in_channels=IntParam(name="in_channels", default=1),
out_channels=IntParam(name="out_channels", default=1),
kernel_size=TupleParam(
name="kernel_size",
size=3,
limits=((1, 1, 1), (1, 1, 1)),
default=(1, 1, 1),
),
stride=TupleParam(
name="stride", size=3, limits=((1, 1, 1), (1, 1, 1)), default=(1, 1, 1)
),
padding=TupleParam(
name="padding", size=3, limits=((0, 0, 0), (0, 0, 0)), default=(0, 0, 0)
),
output_padding=TupleParam(
name="output_padding",
size=3,
limits=((0, 0, 0), (0, 0, 0)),
default=(0, 0, 0),
),
groups=IntParam(name="groups", default=1),
bias=BinaryParam(name="bias", default=True, true_prob=0.5),
dilation=TupleParam(
name="dilation",
size=3,
limits=((1, 1, 1), (1, 1, 1)),
default=(1, 1, 1),
),
padding_mode=ChoiceParam(
name="padding_mode", choices=("zeros",), cprobs=(1,), default="zeros"
),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"Hardtanh",
["min_val", "max_val", "inplace", "min_value", "max_value"])
self.params = self.template_fn(
min_val=IntParam(name="min_val", default=-1.0),
max_val=IntParam(name="max_val", default=1.0),
inplace=BinaryParam(name="inplace", default=False, true_prob=0.5),
min_value=Param(name="min_value", default=None),
max_value=Param(name="max_value", default=None),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"AdaptiveLogSoftmaxWithLoss",
["in_features", "n_classes", "cutoffs", "div_value", "head_bias"],
)
self.params = self.template_fn(
in_features=IntParam(name="in_features", default=1),
n_classes=Param(name="n_classes", default=None),
cutoffs=Param(name="cutoffs", default=None),
div_value=IntParam(name="div_value", default=4.0),
head_bias=BinaryParam(name="head_bias", default=False, true_prob=0.5),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"MultiheadAttention",
[
"embed_dim",
"num_heads",
"dropout",
"bias",
"add_bias_kv",
"add_zero_attn",
"kdim",
"vdim",
],
)
self.params = self.template_fn(
embed_dim=Param(name="embed_dim", default=None),
num_heads=Param(name="num_heads", default=None),
dropout=IntParam(name="dropout", default=0.0),
bias=BinaryParam(name="bias", default=True, true_prob=0.5),
add_bias_kv=BinaryParam(name="add_bias_kv",
default=False,
true_prob=0.5),
add_zero_attn=BinaryParam(name="add_zero_attn",
default=False,
true_prob=0.5),
kdim=Param(name="kdim", default=None),
vdim=Param(name="vdim", default=None),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple("PixelShuffle", ["upscale_factor"])
self.params = self.template_fn(
upscale_factor=IntParam(name="upscale_factor", default=1)
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple("PReLU", ["num_parameters", "init"])
self.params = self.template_fn(
num_parameters=IntParam(name="num_parameters", default=1),
init=FloatParam(name="init", default=0.25),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple("Softplus", ["beta", "threshold"])
self.params = self.template_fn(
beta=FloatParam(name="beta", default=1),
threshold=IntParam(name="threshold", default=20),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple("CosineSimilarity", ["dim", "eps"])
self.params = self.template_fn(
dim=IntParam(name="dim", default=1),
eps=FloatParam(name="eps", default=1e-08),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple("PairwiseDistance", ["p", "eps", "keepdim"])
self.params = self.template_fn(
p=IntParam(name="p", default=2.0),
eps=FloatParam(name="eps", default=1e-06),
keepdim=BinaryParam(name="keepdim", default=False, true_prob=0.5),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple("UpsamplingBilinear2d",
["size", "scale_factor"])
self.params = self.template_fn(
size=Param(name="size", default=None),
scale_factor=IntParam(name="scale_factor", default=1),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"MultiMarginLoss",
["p", "margin", "weight", "size_average", "reduce", "reduction"],
)
self.params = self.template_fn(
p=IntParam(name="p", default=1),
margin=IntParam(name="margin", default=1.0),
weight=Param(name="weight", default=None),
size_average=Param(name="size_average", default=None),
reduce=Param(name="reduce", default=None),
reduction=ChoiceParam(name="reduction",
choices=("mean", ),
cprobs=(1, ),
default="mean"),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple("GRUCell",
["input_size", "hidden_size", "bias"])
self.params = self.template_fn(
input_size=IntParam(name="input_size", default=1),
hidden_size=Param(name="hidden_size", default=None),
bias=BinaryParam(name="bias", default=True, true_prob=0.5),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"RNNCellBase", ["input_size", "hidden_size", "bias", "num_chunks"])
self.params = self.template_fn(
input_size=IntParam(name="input_size", default=1),
hidden_size=Param(name="hidden_size", default=None),
bias=Param(name="bias", default=None),
num_chunks=Param(name="num_chunks", default=None),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple("CrossMapLRN2d",
["size", "alpha", "beta", "k"])
self.params = self.template_fn(
size=Param(name="size", default=None),
alpha=FloatParam(name="alpha", default=0.0001),
beta=FloatParam(name="beta", default=0.75),
k=IntParam(name="k", default=1),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple("Transformer", [
'd_model', 'nhead', 'num_encoder_layers', 'num_decoder_layers',
'dim_feedforward', 'dropout', 'activation', 'custom_encoder',
'custom_decoder'
])
self.params = self.template_fn(
d_model=IntParam(name="d_model", default=512),
nhead=IntParam(name="nhead", default=8),
num_encoder_layers=IntParam(name="num_encoder_layers", default=6),
num_decoder_layers=IntParam(name="num_decoder_layers", default=6),
dim_feedforward=IntParam(name="dim_feedforward", default=2048),
dropout=FloatParam(name="dropout", default=0.1),
activation=ChoiceParam(name="activation",
choices=("relu", ),
cprobs=(1, ),
default="relu"),
custom_encoder=Param(name="custom_encoder", default=None),
custom_decoder=Param(name="custom_decoder", default=None),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"TripletMarginLoss",
[
"margin", "p", "eps", "swap", "size_average", "reduce",
"reduction"
],
)
self.params = self.template_fn(
margin=IntParam(name="margin", default=1.0),
p=IntParam(name="p", default=2.0),
eps=FloatParam(name="eps", default=1e-06),
swap=BinaryParam(name="swap", default=False, true_prob=0.5),
size_average=Param(name="size_average", default=None),
reduce=Param(name="reduce", default=None),
reduction=ChoiceParam(name="reduction",
choices=("mean", ),
cprobs=(1, ),
default="mean"),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"Upsample", ["size", "scale_factor", "mode", "align_corners"])
self.params = self.template_fn(
size=Param(name="size", default=None),
scale_factor=IntParam(name="scale_factor", default=1),
mode=ChoiceParam(name="mode",
choices=("nearest", ),
cprobs=(1, ),
default="nearest"),
align_corners=Param(name="align_corners", default=None),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"RNNCell", ["input_size", "hidden_size", "bias", "nonlinearity"])
self.params = self.template_fn(
input_size=IntParam(name="input_size", default=1),
hidden_size=Param(name="hidden_size", default=None),
bias=BinaryParam(name="bias", default=True, true_prob=0.5),
nonlinearity=ChoiceParam(name="nonlinearity",
choices=("tanh", ),
cprobs=(1, ),
default="tanh"),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple("CTCLoss",
["blank", "reduction", "zero_infinity"])
self.params = self.template_fn(
blank=IntParam(name="blank", default=0),
reduction=ChoiceParam(name="reduction",
choices=("mean", ),
cprobs=(1, ),
default="mean"),
zero_infinity=BinaryParam(name="zero_infinity",
default=False,
true_prob=0.5),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"CosineEmbeddingLoss",
["margin", "size_average", "reduce", "reduction"])
self.params = self.template_fn(
margin=IntParam(name="margin", default=0.0),
size_average=Param(name="size_average", default=None),
reduce=Param(name="reduce", default=None),
reduction=ChoiceParam(name="reduction",
choices=("mean", ),
cprobs=(1, ),
default="mean"),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
def __init__(self, **kwargs):
_Factory.__init__(self)
self.template_fn = namedtuple(
"NLLLoss",
["weight", "size_average", "ignore_index", "reduce", "reduction"])
self.params = self.template_fn(
weight=Param(name="weight", default=None),
size_average=Param(name="size_average", default=None),
ignore_index=IntParam(name="ignore_index", default=-100),
reduce=Param(name="reduce", default=None),
reduction=ChoiceParam(name="reduction",
choices=("mean", ),
cprobs=(1, ),
default="mean"),
)
for k, v in kwargs.items():
getattr(self.params, k).val = v
class ResNetStyle(_Net):
def __init__(self, start_shape, stop_shape, depth=3):
_Net.__init__(self, start_shape, stop_shape, depth)
self.layers.append(Conv2dAC(start_shape, [10, 1, 1], depth))
self.layers.append(Conv2dThenLinear(start_shape, [10, 1, 1], depth))
self.change_num = IntParam("", default=1)
def generate(self):
z1 = Conv2dAC.ac.val
Conv2dAC.ac.val = Conv2dAC.ac.choices[2]
z2 = LinearAC.ac.val
LinearAC.ac.val = LinearAC.ac.choices[2]
self.change_num.randomize(limits=(1, self.depth - 2 * self.depth // 3))
t = self.change_num.value
cp = self.layers[0].generate(resnet_cover=20)[:t]
skip_gen = BasicBlock()
self.change_num.randomize(
limits=(1, max([1, self.depth - (t + self.depth // 3)]))
)
t2 = self.change_num.value
sp = []
for i in range(t2):
sp.append(skip_gen(_in_shape=cp[-1].out_shape))
self.change_num.randomize(limits=(2, max([2, self.depth - (t + t2)])))
t3 = self.change_num.value
self.layers[1].start_shape = sp[-1].out_shape
self.layers[1].conv_part.start_shape = sp[-1].out_shape
self.layers[1].depth = t3
lp = self.layers[1].generate()
Conv2dAC.ac.val = z1
LinearAC.ac.val = z2
return cp + sp + lp
