This commit is contained in:
Connor Olding 2017-01-09 22:19:28 -08:00
parent 8baa7a267a
commit d299520fd9

View File

@ -230,6 +230,23 @@ class Affine(Layer):
def dF(self, dY):
return dY * self.a
class Sigmoid(Layer): # aka Logistic
def F(self, X):
from scipy.special import expit as sigmoid
self.sig = sigmoid(X)
return X * self.sig
def dF(self, dY):
return dY * self.sig * (1 - self.sig)
class Tanh(Layer):
def F(self, X):
self.sig = np.tanh(X)
return X * self.sig
def dF(self, dY):
return dY * (1 - self.sig * self.sig)
class Relu(Layer):
def F(self, X):
self.cond = X >= 0
@ -239,7 +256,8 @@ class Relu(Layer):
return np.where(self.cond, dY, 0)
class GeluApprox(Layer):
# refer to https://www.desmos.com/calculator/ydzgtccsld
# paper: https://arxiv.org/abs/1606.08415
# plot: https://www.desmos.com/calculator/ydzgtccsld
def F(self, X):
from scipy.special import expit as sigmoid
self.a = 1.704 * X
@ -385,26 +403,32 @@ if __name__ == '__main__':
config = DotMap(
fn = 'ml/cie_mlp_min.h5',
batch_size = 64,
# multi-residual network parameters
res_width = 12,
res_depth = 3,
res_block = 2, # normally 2
res_multi = 4, # normally 1
res_block = 2, # normally 2 for plain resnet
res_multi = 4, # normally 1 for plain resnet
# style of resnet
# only one is implemented so far
parallel_style = 'batchless',
activation = 'gelu',
optim = 'adam',
nesterov = False, # only used with SGD or Adam
momentum = 0.33, # only used with SGD
epochs = 6, # 6
# learning parameters: SGD with restarts
LR = 1e-2,
restarts = 3, # 3
epochs = 6,
LR_halve_every = 2,
restarts = 3,
LR_restart_advance = 3,
# misc
batch_size = 64,
init = 'he_normal',
loss = 'mse',
parallel_style = 'batchless',
)
# toy CIE-2000 data
@ -429,7 +453,7 @@ if __name__ == '__main__':
y = x
last_size = input_samples
activations = dict(relu=Relu, gelu=GeluApprox)
activations = dict(sigmoid=Sigmoid, tanh=Tanh, relu=Relu, gelu=GeluApprox)
activation = activations[config.activation]
for blah in range(config.res_depth):