def __init__(self,
arch,
normalization='batch',
activation='leakyrelu-0.2',
padding='same',
pooling='avg',
input_size=(128, 128),
layout_dim=0,
args=None):
super(PatchDiscriminator, self).__init__()
print(
"i2g2i.combine_sg2im_neural_motifs.discriminator.PatchDiscriminator"
)
input_dim = 3 + layout_dim
arch = 'I%d,%s' % (input_dim, arch)
cnn_kwargs = {
'arch': arch,
'normalization': normalization,
'activation': activation,
'pooling': pooling,
'padding': padding,
}
self.down_to_1channel = args.down_to_1channel
self.cnn, output_dim = build_cnn(**cnn_kwargs)
self.classifier = nn.Conv2d(output_dim, 1, kernel_size=1, stride=1)
def __init__(self,
vocab,
arch,
normalization='none',
activation='relu',
padding='same',
pooling='avg',
args=None):
super(AcDiscriminator, self).__init__()
print(
"i2g2i.combine_sg2im_neural_motifs.discriminator.AcDiscriminator")
self.vocab = vocab
cnn_kwargs = {
'arch': arch,
'normalization': normalization,
'activation': activation,
'pooling': pooling,
'padding': padding,
}
cnn, D = build_cnn(**cnn_kwargs)
self.reconstruct_frature = args.d_obj_rec_feat_weight > 0
if self.reconstruct_frature:
self.cnn = cnn
self.feature = nn.Sequential(GlobalAvgPool(), nn.Linear(D, 1024))
self.rec_feature = nn.Sequential(GlobalAvgPool(),
nn.Linear(D, 4096))
else:
self.cnn = nn.Sequential(cnn, GlobalAvgPool(), nn.Linear(D, 1024))
num_objects = len(vocab['object_idx_to_name']) - 1
self.real_classifier = nn.Linear(1024, 1)
self.obj_classifier = nn.Linear(1024, num_objects)
def __init__(self,
arch,
normalization='batch',
activation='leakyrelu-0.2',
padding='same',
pooling='avg',
args=None):
super(ImageEncoder, self).__init__()
print("i2g2i.combine_sg2im_neural_motifs.encoder.ImageEncoder")
input_dim = 3
arch = 'I%d,%s' % (input_dim, arch)
cnn_kwargs = {
'arch': arch,
'normalization': normalization,
'activation': activation,
'pooling': pooling,
'padding': padding,
}
self.layout_noise_dim = args.layout_noise_dim
self.object_noise_dim = args.object_noise_dim
assert not (
self.layout_noise_dim > 0 and self.object_noise_dim > 0
), "layout noise, object noise should not be applied at the same time"
self.cnn, D = build_cnn(**cnn_kwargs)
for i in range(len(self.cnn)):
if isinstance(self.cnn[i], nn.Conv2d):
origin_cnn = self.cnn[i]
self.cnn[i] = nn.Conv2d(origin_cnn.in_channels,
origin_cnn.out_channels,
kernel_size=origin_cnn.kernel_size,
stride=origin_cnn.stride,
padding=1)
def __init__(self, arch, normalization='batch', activation='leakyrelu-0.2',
padding='same', pooling='avg', input_size=(128,128),
layout_dim=0):
super(PatchDiscriminator, self).__init__()
input_dim = 3 + layout_dim
arch = 'I%d,%s' % (input_dim, arch)
cnn_kwargs = {
'arch': arch,
'normalization': normalization,
'activation': activation,
'pooling': pooling,
'padding': padding,
}
self.cnn, output_dim = build_cnn(**cnn_kwargs)
self.classifier = nn.Conv2d(output_dim, 1, kernel_size=1, stride=1)
def __init__(self, vocab, arch, normalization='none', activation='relu',
padding='same', pooling='avg'):
super(AcDiscriminator, self).__init__()
self.vocab = vocab
cnn_kwargs = {
'arch': arch,
'normalization': normalization,
'activation': activation,
'pooling': pooling,
'padding': padding,
}
cnn, D = build_cnn(**cnn_kwargs)
self.cnn = nn.Sequential(cnn, GlobalAvgPool(), nn.Linear(D, 1024))
num_objects = len(vocab['object_idx_to_name'])
self.real_classifier = nn.Linear(1024, 1)
self.obj_classifier = nn.Linear(1024, num_objects)
def __init__(self,
vocab,
arch,
normalization='none',
activation='relu',
padding='same',
vecs_size=1024,
pooling='avg'):
super(AppearanceEncoder, self).__init__()
self.vocab = vocab
cnn_kwargs = {
'arch': arch,
'normalization': normalization,
'activation': activation,
'pooling': pooling,
'padding': padding,
}
cnn, channels = build_cnn(**cnn_kwargs)
self.cnn = nn.Sequential(cnn, GlobalAvgPool(),
nn.Linear(channels, vecs_size))
def __init__(self,
arch,
normalization='none',
activation='relu',
padding='same',
pooling='avg',
args=None):
super(Encoder, self).__init__()
print("i2g2i.combine_sg2im_neural_motifs.encoder.Encoder")
self.object_noise_dim = args.object_noise_dim
cnn_kwargs = {
'arch': arch,
'normalization': normalization,
'activation': activation,
'pooling': pooling,
'padding': padding,
}
cnn, D = build_cnn(**cnn_kwargs)
self.cnn = nn.Sequential(cnn, GlobalAvgPool(),
nn.Linear(D, self.object_noise_dim * 2))
def __init__(self,
arch,
normalization='batch',
activation='leakyrelu-0.2',
padding='same',
pooling='avg',
input_size=(128, 128),
layout_dim=0,
sg_context_dim=0):
super(CondGANDiscriminator, self).__init__()
input_dim = 3 + layout_dim + sg_context_dim
arch = 'I%d,%s' % (input_dim, arch)
cnn_kwargs = {
'arch': arch,
'normalization': normalization,
'activation': activation,
'pooling': pooling,
'padding': padding,
}
cnn, output_dim = build_cnn(**cnn_kwargs)
self.cnn = nn.Sequential(cnn, GlobalAvgPool(),
nn.Linear(output_dim, 1024))
self.classifier = nn.Linear(1024, 1)