def get_model_case10_fc_z2(model_type='vae'):
if model_type == 'vae':
last_layer = NV_.VAEChain
loss_chain = NV_.VAELoss
else:
last_layer = NA_.LAEChain
loss_chain = NA_.AELoss
model = NA_.CAEList(
NA_.LAEChain(None, 196, activation=(F.sigmoid, None)),
NA_.LAEChain(None, 49),
NA_.LAEChain(None, 7),
last_layer(None, 2))
return loss_chain(model)
def get_model_case0():
''' 畳み込みオートエンコーダ '''
# 入出力チャンネル数を指定
model = NA_.CAEList(
NA_.CAEChain(2, 10, activation=(F.relu, None)), # in: 512
NA_.CAEChain(10, 20), # in: 256
NA_.CAEChain(20, 20), # in: 128
NA_.CAEChain(20, 20), # in: 64
NA_.CAEChain(20, 20), # in: 32
NA_.CAEChain(20, 20), # in: 16
NA_.CAEChain(20, 20), # in: 8
NA_.CAEChain(20, 20), # in: 3
NA_.LAEChain(None, 20), # in: 10*3*3
NA_.LAEChain(20, 10, activation=(None, F.relu)))
return model
def get_model_case7_1():
''' AE
'''
# 入出力チャンネル数を指定
model = NA_.CAEList(
NA_.CAEChain(3, 8, activation=(F.sigmoid, None)),
NA_.CAEChain(8, 16, activation=F.sigmoid),
NA_.CAEChain(16, 32, activation=F.sigmoid),
NA_.CAEChain(32, 64, activation=F.sigmoid),
NA_.CAEChain(64, 64, activation=F.sigmoid),
NA_.CAEChain(64, 128, activation=F.sigmoid),
NA_.CAEChain(128, 256, activation=F.sigmoid),
NA_.LAEChain(None, 64, activation=(None, F.sigmoid)))
loss = NA_.AELoss(model)
return loss
def get_model_case6_4():
''' VAE
'''
# 入出力チャンネル数を指定
model = NA_.CAEList(
NA_.CAEChain(3, 8, activation=(F.relu, None)),
NA_.CAEChain(8, 16, activation=F.relu),
NA_.CAEChain(16, 32, activation=F.relu),
NA_.CAEChain(32, 64, activation=F.relu),
NA_.CAEChain(64, 64, activation=F.relu),
NA_.CAEChain(64, 128, activation=F.relu),
NA_.CAEChain(128, 256, activation=F.relu),
NV_.VAEChain(None, 64, activation=F.relu))
loss = NV_.VAELoss(model, beta=1.0, k=1)
return loss
def get_model_case1_z30():
''' VAE '''
# 入出力チャンネル数を指定
model = NA_.CAEList(
NA_.CAEChain(2, 10, activation=(F.relu, None)), # in: 512, 1024
NA_.CAEChain(10, 20), # in: 256
NA_.CAEChain(20, 20), # in: 128
NA_.CAEChain(20, 20), # in: 64
NA_.CAEChain(20, 20), # in: 32
NA_.CAEChain(20, 20), # in: 16
NA_.CAEChain(20, 20), # in: 8
NA_.CAEChain(20, 20), # in: 4
# NA_.LAEChain(None, 40), # in: 10*3*3
NV_.VAEChain(None, 30)) # in: 10*3*3
loss = NV_.VAELoss(model, beta=1.0, k=1)
return loss
def get_model_case10_2(model_type='vae'):
if model_type == 'vae':
last_layer = NV_.VAEChain
loss_chain = NV_.VAELoss
else:
last_layer = NA_.LAEChain
loss_chain = NA_.AELoss
model = NA_.CAEList(
NA_.CAEChain(None, 16, activation=(F.sigmoid, None)), # => 14
NA_.CAEChain(None, 32), # => 7
NA_.CAEChain(None, 64), # => 4
NA_.CAEChain(None, 128), # => 2 => 512
NA_.LAEChain(None, 128),
NA_.LAEChain(None, 32),
last_layer(None, 3))
return loss_chain(model)
def get_model_case9_1():
''' VAE
z: 32dim
'''
model = NA_.CAEList(
NA_.CAEChain(3, 8, activation=(F.sigmoid, None)),
NA_.CAEChain(None, 16),
NA_.CAEChain(None, 32),
NA_.CAEChain(None, 64),
NA_.CAEChain(None, 64),
NA_.CAEChain(None, 64),
NA_.CAEChain(None, 128), # => (128, 3, 3)
NA_.LAEChain(None, 288), # 1152 -> 288 (1 / 4)
NA_.LAEChain(None, 96), # 288 -> 96 (1 / 3)
NV_.VAEChain(None, 32)) # 96 -> 32 (1 / 3)
loss = NV_.VAELoss(model)
return loss
def get_model_case9_0():
''' VAE
z: 2dim
'''
model = NA_.CAEList(
NA_.CAEChain(3, 8, activation=(F.sigmoid, None)),
NA_.CAEChain(None, 16),
NA_.CAEChain(None, 32),
NA_.CAEChain(None, 64),
NA_.CAEChain(None, 64),
NA_.CAEChain(None, 64),
NA_.CAEChain(None, 128), # => (128, 3, 3)
NA_.LAEChain(None, 128), # 1152 -> 128 (1 / 9)
NA_.LAEChain(None, 16), # 128 -> 16 (1 / 8)
NV_.VAEChain(None, 2)) # 16 -> 2 (1 / 8)
loss = NV_.VAELoss(model)
return loss
def get_model_case8_0():
''' VAE
activation: F.relu => F.sigmoid
'''
# 入出力チャンネル数を指定
model = NA_.CAEList(
NA_.CAEChain(3, 8, activation=(F.sigmoid, None)),
NA_.CAEChain(8, 16, activation=F.sigmoid),
NA_.CAEChain(16, 32, activation=F.sigmoid),
NA_.CAEChain(32, 64, activation=F.sigmoid),
NA_.CAEChain(64, 64, activation=F.sigmoid),
NA_.CAEChain(64, 128, activation=F.sigmoid),
NA_.CAEChain(128, 256, activation=F.sigmoid), # => (256, 3, 3)
NV_.VAEChain(None, 64, activation=(None, F.sigmoid))) # 2304 -> 64 (1 / 36)
loss = NV_.VAELoss(model, beta=1.0, k=1)
return loss
def get_model_case6_3():
''' VAE
'''
# 入出力チャンネル数を指定
model = NA_.CAEList(
NA_.CAEChain(3, 8, activation=(F.relu, None), batch_norm='re',
padding=True),
NA_.CAEChain(8, 16, batch_norm='re', padding=True),
NA_.CAEChain(16, 32, batch_norm='re', padding=True),
NA_.CAEChain(32, 64, batch_norm='re', padding=True),
NA_.CAEChain(64, 64, batch_norm='re', padding=True),
NA_.CAEChain(64, 128, batch_norm='re', padding=True),
NA_.CAEChain(128, 256, batch_norm='re', padding=True),
NV_.VAEChain(None, 64))
loss = NV_.VAELoss(model, beta=1.0, k=1)
return loss
def get_model_case2n():
''' AE '''
# 入出力チャンネル数を指定
model = NA_.CAEList(
NA_.CAEChain(2, 10, activation=(F.relu, None)), # in: 512, 1024
NA_.CAEChain(10, 20), # in: 256
NA_.CAEChain(20, 20), # in: 128
NA_.CAEChain(20, 20), # in: 64
NA_.CAEChain(20, 20), # in: 32
NA_.CAEChain(20, 20), # in: 16
NA_.CAEChain(20, 20), # in: 8
NA_.CAEChain(20, 20), # in: 4
NA_.LAEChain(None, 10), # in: 10*3*3
NA_.LAEChain(10, 2, activation=(None, F.relu)))
loss = L.Classifier(model, lossfun=F.mean_squared_error)
if C_.DEVICE >= 0:
loss.to_gpu(C_.DEVICE)
return loss
def get_model_case9_4():
''' VAE
z: 64dim
パラメータ数: ((3*8+8*8+8*16+16*16+16*32+32*32+32*64+64*64+64*64+64*64*128+128*128+128*128+128*128)*3*3+
1152*384+384*128+128*64)*2=11541808
'''
model = NA_.CAEList(
NA_.CAEChain(3, 8, activation=(F.sigmoid, None), ksize=3),
NA_.CAEChain(None, 16, ksize=3),
NA_.CAEChain(None, 32, ksize=3),
NA_.CAEChain(None, 64, ksize=3),
NA_.CAEChain(None, 64, ksize=3),
NA_.CAEChain(None, 128, ksize=3),
NA_.CAEChain(None, 128, ksize=3), # => (128, 3, 3)
NA_.LAEChain(None, 384), # 1152 -> 384 (1 / 3)
NA_.LAEChain(None, 128), # 384 -> 128 (1 / 3)
NV_.VAEChain(None, 64)) # 128 -> 64 (1 / 2)
loss = NV_.VAELoss(model)
return loss
def get_model_case7():
''' AE
'''
# 入出力チャンネル数を指定
model = NA_.CAEList(
NA_.CAEChain(3, 8, activation=(F.relu, None), batch_norm='re',
ksize=5, padding=True),
NA_.CAEChain(8, 16, batch_norm='re', ksize=5, padding=True),
NA_.CAEChain(16, 32, batch_norm='re', ksize=5, padding=True),
NA_.CAEChain(32, 64, batch_norm='re', ksize=5, padding=True),
NA_.CAEChain(64, 64, batch_norm='re', ksize=5, padding=True),
NA_.CAEChain(64, 128, batch_norm='re', ksize=5, padding=True),
NA_.CAEChain(128, 256, batch_norm='re', ksize=5, padding=True),
NA_.LAEChain(None, 64, activation=(None, F.relu)))
loss = L.Classifier(model, lossfun=F.mean_squared_error)
if C_.DEVICE >= 0:
loss.to_gpu(C_.DEVICE)
return loss
def get_model_case6():
''' VAE
1 4 10 22 46 94 190 382
2 6 14 30 62 126 254 510
3 8 18 38 78 158 318
'''
# 入出力チャンネル数を指定
model = NA_.CAEList(
NA_.CAEChain(3, 8, activation=(F.relu, None)),
NA_.CAEChain(8, 16),
NA_.CAEChain(16, 32),
NA_.CAEChain(32, 64),
NA_.CAEChain(64, 64),
NA_.CAEChain(64, 128),
NA_.CAEChain(128, 256),
NV_.VAEChain(None, 64))
loss = NV_.VAELoss(model, beta=1.0, k=1)
return loss
def get_model_case4_1():
''' VAE
1 4 10 22 46 94 190 382
2 6 14 30 62 126 254 510
3 8 18 38 78 158 318
'''
# 入出力チャンネル数を指定
model = NA_.CAEList(
NA_.CAEChain(2, 10, activation=(F.relu, None)),
NA_.CAEChain(10, 20),
NA_.CAEChain(20, 30),
NA_.CAEChain(30, 30),
NA_.CAEChain(30, 30),
NA_.CAEChain(30, 30),
NA_.CAEChain(30, 20),
NV_.VAEChain(None, 10))
loss = NV_.VAELoss(model, beta=1.0, k=1)
return loss
def get_model_case5_1():
''' AE
1 4 10 22 46 94 190 382
2 6 14 30 62 126 254 510
3 8 18 38 78 158 318
'''
# 入出力チャンネル数を指定
model = NA_.CAEList(
NA_.CAEChain(2, 10, activation=(F.relu, None)),
NA_.CAEChain(10, 20),
NA_.CAEChain(20, 30),
NA_.CAEChain(30, 30),
NA_.CAEChain(30, 30),
NA_.CAEChain(30, 30),
NA_.CAEChain(30, 20),
NA_.LAEChain(None, 10, activation=(None, F.relu)))
loss = L.Classifier(model, lossfun=F.mean_squared_error)
if C_.DEVICE >= 0:
loss.to_gpu(C_.DEVICE)
return loss
