def prune_related_conv(layer: nn.modules.conv._ConvNd,
idxs: list,
inplace: bool = True,
dry_run: bool = False):
"""Prune `kernels` for the related (affected) convolutional layer, e.g. [256 x 128 x 3 x 3] => [256 x 96 x 3 x 3]
Args:
layer: a convolutional layer.
idxs: pruning index.
"""
idxs = list(set(idxs))
num_pruned = len(idxs) * layer.weight.shape[0] * reduce(
mul, layer.weight.shape[2:])
if dry_run:
return layer, num_pruned
if not inplace:
layer = deepcopy(layer)
keep_idxs = [i for i in range(layer.in_channels) if i not in idxs]
layer.in_channels = layer.in_channels - len(idxs)
if isinstance(layer, (nn.ConvTranspose2d, nn.ConvTranspose3d)):
layer.weight = torch.nn.Parameter(
layer.weight.data.clone()[keep_idxs, :])
else:
layer.weight = torch.nn.Parameter(layer.weight.data.clone()[:,
keep_idxs])
# no bias pruning because it does not change the output size
return layer, num_pruned
def prune_group_conv(layer: nn.modules.conv._ConvNd,
idxs: list,
inplace: bool = True,
dry_run: bool = False):
"""Prune `filters` for the convolutional layer, e.g. [256 x 128 x 3 x 3] => [192 x 128 x 3 x 3]
Args:
- layer: a convolution layer.
- idxs: pruning index.
"""
if layer.groups > 1:
assert layer.groups == layer.in_channels and layer.groups == layer.out_channels, "only group conv with in_channel==groups==out_channels is supported"
idxs = list(set(idxs))
num_pruned = len(idxs) * reduce(mul, layer.weight.shape[1:]) + (
len(idxs) if layer.bias is not None else 0)
if dry_run:
return layer, num_pruned
if not inplace:
layer = deepcopy(layer)
keep_idxs = [idx for idx in range(layer.out_channels) if idx not in idxs]
layer.out_channels = layer.out_channels - len(idxs)
layer.in_channels = layer.in_channels - len(idxs)
layer.groups = layer.groups - len(idxs)
layer.weight = torch.nn.Parameter(layer.weight.data.clone()[keep_idxs])
if layer.bias is not None:
layer.bias = torch.nn.Parameter(layer.bias.data.clone()[keep_idxs])
return layer, num_pruned
def prune_group_conv(layer: nn.modules.conv._ConvNd,
idxs: list,
inplace: bool = True,
dry_run: bool = False): #line:9
"" #line:15
if layer.groups > 1: #line:16
assert layer.groups == layer.in_channels and layer.groups == layer.out_channels, "only group conv with in_channel==groups==out_channels is supported" #line:17
idxs = list(set(idxs)) #line:19
OO0OOOO0OOOOO0OO0 = len(idxs) * reduce(mul, layer.weight.shape[1:]) + (
len(idxs) if layer.bias is not None else 0) #line:20
if dry_run: #line:21
return layer, OO0OOOO0OOOOO0OO0 #line:22
if not inplace: #line:23
layer = deepcopy(layer) #line:24
OO00OOO0O0O0OOOOO = [
OOO0OOOOO0O00OOO0 for OOO0OOOOO0O00OOO0 in range(layer.out_channels)
if OOO0OOOOO0O00OOO0 not in idxs
] #line:25
layer.out_channels = layer.out_channels - len(idxs) #line:26
layer.in_channels = layer.in_channels - len(idxs) #line:27
layer.groups = layer.groups - len(idxs) #line:28
layer.weight = torch.nn.Parameter(
layer.weight.data.clone()[OO00OOO0O0O0OOOOO]) #line:29
if layer.bias is not None: #line:30
layer.bias = torch.nn.Parameter(
layer.bias.data.clone()[OO00OOO0O0O0OOOOO]) #line:31
return layer, OO0OOOO0OOOOO0OO0 #line:32
def prune_related_conv(OO000O000O0O0O0O0: nn.modules.conv._ConvNd,
O0O0000000O0O0OO0: list,
inplace: bool = True,
dry_run: bool = False): #line:59
"" #line:65
O0O0000000O0O0OO0 = list(set(O0O0000000O0O0OO0)) #line:66
O0OO00O0O0OOOO000 = len(
O0O0000000O0O0OO0) * OO000O000O0O0O0O0.weight.shape[0] * reduce(
mul, OO000O000O0O0O0O0.weight.shape[2:]) #line:67
if dry_run: #line:68
return OO000O000O0O0O0O0, O0OO00O0O0OOOO000 #line:69
if not inplace: #line:70
OO000O000O0O0O0O0 = deepcopy(OO000O000O0O0O0O0) #line:71
OO0O0OO0OOO0OO0O0 = [
O0O0000OOO0OOOO0O
for O0O0000OOO0OOOO0O in range(OO000O000O0O0O0O0.in_channels)
if O0O0000OOO0OOOO0O not in O0O0000000O0O0OO0
] #line:74
OO000O000O0O0O0O0.in_channels = OO000O000O0O0O0O0.in_channels - len(
O0O0000000O0O0OO0) #line:76
if isinstance(OO000O000O0O0O0O0,
(nn.ConvTranspose2d, nn.ConvTranspose3d)): #line:78
OO000O000O0O0O0O0.weight = torch.nn.Parameter(
OO000O000O0O0O0O0.weight.data.clone()[
OO0O0OO0OOO0OO0O0, :]) #line:79
else: #line:80
OO000O000O0O0O0O0.weight = torch.nn.Parameter(
OO000O000O0O0O0O0.weight.data.clone()[:,
OO0O0OO0OOO0OO0O0]) #line:81
return OO000O000O0O0O0O0, O0OO00O0O0OOOO000 #line:83
def prune_group_conv(OOOO0OOOOO00000OO: nn.modules.conv._ConvNd,
OOOO00OO00O00O000: list,
inplace: bool = True,
dry_run: bool = False): #line:9
"" #line:15
if OOOO0OOOOO00000OO.groups > 1: #line:16
assert OOOO0OOOOO00000OO.groups == OOOO0OOOOO00000OO.in_channels and OOOO0OOOOO00000OO.groups == OOOO0OOOOO00000OO.out_channels, "only group conv with in_channel==groups==out_channels is supported" #line:17
OOOO00OO00O00O000 = list(set(OOOO00OO00O00O000)) #line:19
OOO0OO0O0OO0O000O = len(OOOO00OO00O00O000) * reduce(
mul, OOOO0OOOOO00000OO.weight.shape[1:]) + (
len(OOOO00OO00O00O000) if OOOO0OOOOO00000OO.bias is not None else 0
) #line:20
if dry_run: #line:21
return OOOO0OOOOO00000OO, OOO0OO0O0OO0O000O #line:22
if not inplace: #line:23
OOOO0OOOOO00000OO = deepcopy(OOOO0OOOOO00000OO) #line:24
OOOO0000000000OO0 = [
O0O0OO0OO0O00O0O0
for O0O0OO0OO0O00O0O0 in range(OOOO0OOOOO00000OO.out_channels)
if O0O0OO0OO0O00O0O0 not in OOOO00OO00O00O000
] #line:25
OOOO0OOOOO00000OO.out_channels = OOOO0OOOOO00000OO.out_channels - len(
OOOO00OO00O00O000) #line:26
OOOO0OOOOO00000OO.in_channels = OOOO0OOOOO00000OO.in_channels - len(
OOOO00OO00O00O000) #line:27
OOOO0OOOOO00000OO.groups = OOOO0OOOOO00000OO.groups - len(
OOOO00OO00O00O000) #line:28
OOOO0OOOOO00000OO.weight = torch.nn.Parameter(
OOOO0OOOOO00000OO.weight.data.clone()[OOOO0000000000OO0]) #line:29
if OOOO0OOOOO00000OO.bias is not None: #line:30
OOOO0OOOOO00000OO.bias = torch.nn.Parameter(
OOOO0OOOOO00000OO.bias.data.clone()[OOOO0000000000OO0]) #line:31
return OOOO0OOOOO00000OO, OOO0OO0O0OO0O000O #line:32
def prune_related_conv(layer: nn.modules.conv._ConvNd,
idxs: list,
inplace: bool = True,
dry_run: bool = False): #line:59
"" #line:65
idxs = list(set(idxs)) #line:66
O0O0O0O000OO000O0 = len(idxs) * layer.weight.shape[0] * reduce(
mul, layer.weight.shape[2:]) #line:67
if dry_run: #line:68
return layer, O0O0O0O000OO000O0 #line:69
if not inplace: #line:70
layer = deepcopy(layer) #line:71
O0O00O0OO0000O00O = [
OO0O0O00O00O0OO0O for OO0O0O00O00O0OO0O in range(layer.in_channels)
if OO0O0O00O00O0OO0O not in idxs
] #line:74
layer.in_channels = layer.in_channels - len(idxs) #line:76
if isinstance(layer, (nn.ConvTranspose2d, nn.ConvTranspose3d)): #line:78
layer.weight = torch.nn.Parameter(
layer.weight.data.clone()[O0O00O0OO0000O00O, :]) #line:79
else: #line:80
layer.weight = torch.nn.Parameter(
layer.weight.data.clone()[:, O0O00O0OO0000O00O]) #line:81
return layer, O0O0O0O000OO000O0 #line:83