def __init__(self, n_grapheme=168, n_vowel=11, n_consonant=7, dropout=.0):
super().__init__()
self.n_grapheme = n_grapheme
self.n_vowel = n_vowel
self.n_consonant = n_consonant
self.dropout = dropout
# feature extraction
# input 128x128 -> 4x4; 160x160 -> 5x5; 168x168 -> 6x6
self.feature_extractor = se_resnext50_32x4d()
# dense feature
self.gfeats = nn.Sequential(GeM(), nn.Flatten())
# classifier
self.cls_g = nn.Sequential(
nn.Linear(2048, 256),
nn.BatchNorm1d(256),
Mish(),
nn.Dropout(p=dropout, inplace=True),
nn.Linear(256, self.n_grapheme),
)
self.cls_v = nn.Sequential(
nn.Linear(2048, 256),
nn.BatchNorm1d(256),
Mish(),
nn.Dropout(p=dropout, inplace=True),
nn.Linear(256, self.n_vowel),
)
self.cls_c = nn.Sequential(
nn.Linear(2048, 256),
nn.BatchNorm1d(256),
Mish(),
nn.Dropout(p=dropout, inplace=True),
nn.Linear(256, self.n_consonant),
)
def __init__(self, n_grapheme=168, n_vowel=11, n_consonant=7, dropout=.0):
super().__init__()
self.n_grapheme = n_grapheme
self.n_vowel = n_vowel
self.n_consonant = n_consonant
self.dropout = dropout
# feature extraction
# input 128x128 -> 4x4; 160x160 -> 5x5; 168x168 -> 6x6
self.feature_extractor = se_resnext50_32x4d()
# dense feature
self.cls_pool = nn.Sequential(
# 2048x2x2
nn.MaxPool2d(2, stride=2, padding=0),
nn.BatchNorm2d(2048),
Mish(),
# 8192
nn.Flatten(),
)
self.gfeats = nn.Sequential(
GeM(), nn.Flatten(),
nn.BatchNorm1d(n_grapheme + n_vowel + n_consonant), Mish())
# classifier
self.cls_g = nn.Linear(8192, n_grapheme)
self.cls_v = nn.Linear(8192, n_vowel)
self.cls_c = nn.Linear(8192, n_consonant)
def options2model_kwargs(parameters):
kwargs = options2dataset_kwargs(parameters)
if parameters.mish:
kwargs['activation'] = Mish()
else:
kwargs['activation'] = nn.ReLU()
return kwargs
def options2model_kwargs(parameters):
kargs = dict()
if parameters.mish:
kargs['activation'] = Mish()
else:
kargs['activation'] = nn.ReLU()
return kargs
def __init__(self,
n_grapheme=168,
n_vowel=11,
n_consonant=7,
dropout=.1,
output_features=False):
super().__init__()
self.n_grapheme = n_grapheme
self.n_vowel = n_vowel
self.n_consonant = n_consonant
self.dropout = dropout
self.feat_out = output_features
# feature extraction
# input 128x128 -> 4x4; 160x160 -> 5x5; 168x168 -> 6x6
self.feature_extractor = se_resnext50_32x4d()
# global pooling
self.gpoolmax = nn.AdaptiveMaxPool2d((1, 1))
self.gpoolavg = nn.AdaptiveAvgPool2d((1, 1))
# classifier
self.cls_g = nn.Sequential(
nn.Linear(2048, 512),
nn.BatchNorm1d(512),
Mish(),
nn.Dropout(p=dropout, inplace=True),
nn.Linear(512, self.n_grapheme),
)
self.cls_v = nn.Sequential(
nn.Linear(2048, 512),
nn.BatchNorm1d(512),
Mish(),
nn.Dropout(p=dropout, inplace=True),
nn.Linear(512, self.n_vowel),
)
self.cls_c = nn.Sequential(
nn.Linear(2048, 512),
nn.BatchNorm1d(512),
Mish(),
nn.Dropout(p=dropout, inplace=True),
nn.Linear(512, self.n_consonant),
)
def __init__(self, n_grapheme=168, n_vowel=11, n_consonant=7, dropout=.0):
super().__init__()
self.n_grapheme = n_grapheme
self.n_vowel = n_vowel
self.n_consonant = n_consonant
self.dropout = dropout
# feature extraction
# input 128x128 -> 4x4; 160x160 -> 5x5; 168x168 -> 6x6
self.feature_extractor = se_resnext50_32x4d()
# dense feature
self.gfeats = nn.Sequential(GeM_Spatial(), nn.Flatten(),
nn.Linear(8192, 512), nn.BatchNorm1d(512),
Mish())
self.vcfeats = nn.Sequential(
nn.BatchNorm1d(self.n_vowel + self.n_consonant), Mish())
# classifier
self.cls_g = nn.Linear(512 + self.n_vowel + self.n_consonant,
self.n_grapheme)
self.cls_v = nn.Linear(512, self.n_vowel)
self.cls_c = nn.Linear(512, self.n_consonant)