def __init__(self,
norm_layer,
nf0,
conv_res,
nclasses,
input_res,
pool_res,
fc_n,
nresblocks=3):
super(MeshAutoEncoder, self).__init__()
self.k = [nf0] + conv_res
self.res = [input_res] + pool_res
self.fc_n = fc_n
norm_args = get_norm_args(norm_layer, self.k[1:])
for i, ki in enumerate(self.k[:-1]):
setattr(self, 'conv{}'.format(i),
MResConv(ki, self.k[i + 1], nresblocks))
setattr(self, 'norm{}'.format(i), norm_layer(**norm_args[i]))
setattr(self, 'pool{}'.format(i), MeshPool(self.res[i + 1]))
# self.gp = torch.nn.AvgPool1d(self.res[-1])
self.gp = torch.nn.MaxPool1d(self.res[-1])
self.decoder_fc1 = nn.Linear(self.k[-1], fc_n[0])
self.decoder_fc2 = nn.Linear(fc_n[0], fc_n[1])
self.decoder_fc3 = nn.Linear(fc_n[1], 1402 * 3)
'''
def __init__(self,
norm_layer,
nf0,
conv_res,
nclasses,
input_res,
pool_res,
fc_n,
opt,
nresblocks=3,
num_features=0):
super(MeshConvNet, self).__init__()
self.k = [nf0] + conv_res
self.res = [input_res] + pool_res
self.opt = opt
norm_args = get_norm_args(norm_layer, self.k[1:])
for i, ki in enumerate(self.k[:-1]):
setattr(self, 'conv{}'.format(i),
MResConv(ki, self.k[i + 1], nresblocks))
setattr(self, 'norm{}'.format(i), norm_layer(**norm_args[i]))
setattr(self, 'pool{}'.format(i), MeshPool(self.res[i + 1]))
self.gp = torch.nn.AvgPool1d(self.res[-1])
# self.gp = torch.nn.MaxPool1d(self.res[-1])
if self.opt.dropout:
self.d = nn.Dropout()
self.fc1 = nn.Linear(self.k[-1] + num_features, fc_n)
if self.opt.dataset_mode == 'binary_class':
self.fc2 = nn.Linear(fc_n, 1)
else:
self.fc2 = nn.Linear(fc_n, nclasses)
def __init__(self,
norm_layer,
nf0,
conv_res,
nclasses,
input_res,
pool_res,
fc_n,
nresblocks=3,
stn=None):
super(MeshConvNet, self).__init__()
self.k = [nf0] + conv_res
self.res = [input_res] + pool_res
norm_args = get_norm_args(norm_layer, self.k[1:])
self.trans_inp = None
for i, ki in enumerate(self.k[:-1]):
setattr(self, 'conv{}'.format(i),
MResConv(ki, self.k[i + 1], nresblocks))
setattr(self, 'norm{}'.format(i), norm_layer(**norm_args[i]))
setattr(self, 'pool{}'.format(i), MeshPool(self.res[i + 1]))
self.gp = torch.nn.AvgPool1d(self.res[-1])
# self.gp = torch.nn.MaxPool1d(self.res[-1])
self.fc1 = nn.Linear(self.k[-1], fc_n)
self.fc2 = nn.Linear(fc_n, nclasses)
self.stn = stn
def __init__(
self,
norm_layer,
nf0,
conv_res,
nclasses,
input_res,
pool_res,
fc_n,
attn_n_heads, # d_k, d_v,
nresblocks=3,
attn_max_dist=None,
attn_dropout=0.1,
prioritize_with_attention=False,
attn_use_values_as_is=False,
double_attention=False,
attn_use_positional_encoding=False,
attn_max_relative_position=8):
super(MeshAttentionNet, self).__init__()
if double_attention:
assert attn_use_values_as_is, (
"must have attn_use_values_as_is=True if double_attention=True, "
"since the attention layer works on its own outputs")
self.k = [nf0] + conv_res
self.res = [input_res] + pool_res
self.prioritize_with_attention = prioritize_with_attention
self.double_attention = double_attention
self.use_values_as_is = attn_use_values_as_is
norm_args = get_norm_args(norm_layer, self.k[1:])
for i, ki in enumerate(self.k[:-1]):
setattr(self, 'conv{}'.format(i),
MResConv(ki, self.k[i + 1], nresblocks))
setattr(self, 'norm{}'.format(i), norm_layer(**norm_args[i]))
setattr(
self, 'attention{}'.format(i),
MeshAttention(
n_head=attn_n_heads,
d_model=self.k[i + 1],
d_k=int(self.k[i + 1] / attn_n_heads),
d_v=int(self.k[i + 1] / attn_n_heads),
attn_max_dist=attn_max_dist,
dropout=attn_dropout,
use_values_as_is=attn_use_values_as_is,
use_positional_encoding=attn_use_positional_encoding,
max_relative_position=attn_max_relative_position))
setattr(self, 'pool{}'.format(i), MeshPool(self.res[i + 1]))
self.gp = torch.nn.AvgPool1d(self.res[-1])
# self.gp = torch.nn.MaxPool1d(self.res[-1])
self.fc1 = nn.Linear(self.k[-1], fc_n)
self.fc2 = nn.Linear(fc_n, nclasses)
def __init__(self, in_channels, out_channels, blocks=0, pool=0):
super(DownConv, self).__init__()
self.bn = []
self.pool = None
self.conv1 = MeshConv(in_channels, out_channels)
self.conv2 = []
for _ in range(blocks):
self.conv2.append(MeshConv(out_channels, out_channels))
self.conv2 = nn.ModuleList(self.conv2)
for _ in range(blocks + 1):
self.bn.append(nn.InstanceNorm2d(out_channels))
self.bn = nn.ModuleList(self.bn)
if pool:
self.pool = MeshPool(pool)
def __init__(self, in_channels, out_channels, blocks=0, pool=0, nl_block = 0):
super(DownConv, self).__init__()
self.bn = []
self.pool = None
self.nl_block = None
self.conv1 = MeshConv(in_channels, out_channels)
self.conv2 = []
for _ in range(blocks):
self.conv2.append(MeshConv(out_channels, out_channels))
self.conv2 = nn.ModuleList(self.conv2)
for _ in range(blocks + 1):
self.bn.append(nn.InstanceNorm2d(out_channels))
self.bn = nn.ModuleList(self.bn)
if pool:
self.pool = MeshPool(pool)
if nl_block:
# Add non-local block before last downpool
self.nl_block = NLBlock(out_channels, block_type = nl_block)
def __init__(self,
norm_layer,
nf0,
conv_res,
nclasses,
input_res,
pool_res,
fc_n,
nresblocks=3,
dropout_p=0):
super(MeshConvNet, self).__init__()
self.p = dropout_p
self.k = [nf0] + conv_res
self.res = [input_res] + pool_res
norm_args = get_norm_args(norm_layer, self.k[1:])
for i, ki in enumerate(self.k[:-1]):
setattr(self, 'conv{}'.format(i),
MResConv(ki, self.k[i + 1], nresblocks))
setattr(self, 'norm{}'.format(i), norm_layer(**norm_args[i]))
setattr(self, 'pool{}'.format(i), MeshPool(self.res[i + 1]))
self.gp = torch.nn.AvgPool1d(self.res[-1])
#self.gp = torch.nn.MaxPool1d(self.res[-1])
self.fcs = []
self.fcs.append(nn.Linear(self.k[-1], fc_n[0]))
for i in range(len(fc_n) - 1):
self.fcs.append(nn.Linear(fc_n[i], fc_n[i + 1]))
self.fcs = nn.ModuleList(self.fcs)
self.last_fc = nn.Linear(fc_n[-1], nclasses)
self.bns = []
for i in range(len(fc_n)):
self.bns.append(nn.BatchNorm1d(fc_n[i]))
self.bns = nn.ModuleList(self.bns)
if self.p != 0:
print('dropout layers are applied (p = %f)' % (self.p))