def mknet_gpu(*outlayers):
return df.Sequential( # 3@46
df.SpatialConvolutionCUDNN(3, 24, 3, 3), # -> 24@44
df.BatchNormalization(24),
df.ReLU(),
df.SpatialConvolutionCUDNN(24, 24, 3, 3), # -> 24@42
df.BatchNormalization(24),
df.SpatialMaxPoolingCUDNN(2, 2), # -> 24@21
df.ReLU(),
df.SpatialConvolutionCUDNN(24, 48, 3, 3), # -> 48@19
df.BatchNormalization(48),
df.ReLU(),
df.SpatialConvolutionCUDNN(48, 48, 3, 3), # -> 48@17
df.BatchNormalization(48),
df.SpatialMaxPooling(2, 2), # -> 48@9
df.ReLU(),
df.SpatialConvolutionCUDNN(48, 64, 3, 3), # -> 48@7
df.BatchNormalization(64),
df.ReLU(),
df.SpatialConvolutionCUDNN(64, 64, 3, 3), # -> 48@5
df.BatchNormalization(64),
df.ReLU(),
df.Dropout(0.2),
Flatten(),
df.Linear(64 * 5 * 5, 512),
df.ReLU(),
df.Dropout(0.5),
*outlayers)
def lenet():
model = df.Sequential()
model.add(df.Reshape(-1, 1, 28, 28))
model.add(df.SpatialConvolutionCUDNN(1, 32, 5, 5, 1, 1, 2, 2, with_bias=False))
model.add(df.BatchNormalization(32))
model.add(df.ReLU())
model.add(df.SpatialMaxPoolingCUDNN(2, 2))
model.add(df.SpatialConvolutionCUDNN(32, 64, 5, 5, 1, 1, 2, 2, with_bias=False))
model.add(df.BatchNormalization(64))
model.add(df.ReLU())
model.add(df.SpatialMaxPoolingCUDNN(2, 2))
model.add(df.Reshape(-1, 7*7*64))
model.add(df.Linear(7*7*64, 100, with_bias=False))
model.add(df.BatchNormalization(100))
model.add(df.ReLU())
model.add(df.Dropout(0.5))
model.add(df.Linear(100, 10))
model.add(df.SoftMax())
return model
def model_head(fully_conv=True):
if fully_conv:
return [
df.SpatialConvolutionCUDNN(512, 4096, 7, 7, border='valid'),
df.ReLU(),
df.Dropout(0.5),
df.SpatialConvolutionCUDNN(4096, 4096, 1, 1, border='valid'),
df.ReLU(),
df.Dropout(0.5),
df.SpatialConvolutionCUDNN(4096, 1000, 1, 1, border='valid'),
df.SpatialSoftMaxCUDNN(),
]
else:
return [
df.Reshape(-1, 512 * 7 * 7),
df.Linear(512 * 7 * 7, 4096),
df.ReLU(),
df.Dropout(0.5),
df.Linear(4096, 4096),
df.ReLU(),
df.Dropout(0.5),
df.Linear(4096, 1000),
df.SoftMax()
]
def model(fully_conv=True):
conv3 = lambda nin, nout: df.SpatialConvolutionCUDNN(
nin, nout, 3, 3, border='same')
return df.Sequential(conv3(3, 64), df.ReLU(), conv3(64, 64), df.ReLU(),
df.SpatialMaxPoolingCUDNN(2, 2), conv3(64, 128),
df.ReLU(), conv3(128, 128), df.ReLU(),
df.SpatialMaxPoolingCUDNN(2, 2), conv3(128, 256),
df.ReLU(), conv3(256,
256), df.ReLU(), conv3(256, 256),
df.ReLU(), df.SpatialMaxPoolingCUDNN(2, 2),
conv3(256, 512), df.ReLU(), conv3(512,
512), df.ReLU(),
conv3(512, 512), df.ReLU(),
df.SpatialMaxPoolingCUDNN(2, 2), conv3(512, 512),
df.ReLU(), conv3(512, 512), df.ReLU(),
conv3(512, 512), df.ReLU(),
df.SpatialMaxPoolingCUDNN(2, 2),
*_vgg.model_head(fully_conv))
