def __init__(self, imagesize, flatten_size, num_labels, lr=1e-3):
self.N, self.C, self.W, self.H = imagesize
self.num_labels = num_labels
self.lr = lr
self.graph = nn.Graph()
self.X = nn.Placeholder()
W0 = nn.Variable(nn.UniformInit([5, 5, self.C, 32]), lr=lr)
self.graph.add_var(W0)
b0 = nn.Variable(nn.UniformInit([32, 1]), lr=lr)
self.graph.add_var(b0)
W1 = nn.Variable(nn.UniformInit([5, 5, 32, 64]), lr=lr)
self.graph.add_var(W1)
b1 = nn.Variable(nn.UniformInit([64, 1]), lr=lr)
self.graph.add_var(b1)
WFC0 = nn.Variable(nn.UniformInit([flatten_size, 1024]), lr=lr)
self.graph.add_var(WFC0)
bFC0 = nn.Variable(nn.UniformInit([1024, 1]), lr=lr)
self.graph.add_var(bFC0)
WFC1 = nn.Variable(nn.UniformInit([1024, num_labels]), lr=lr)
self.graph.add_var(WFC1)
bFC1 = nn.Variable(nn.UniformInit([num_labels, 1]), lr=lr)
self.graph.add_var(bFC1)
conv0 = nn.Op(nn.Conv2d(padding='valid'), self.X, [W0, b0])
self.graph.add_op(conv0)
act0 = nn.Layer(nn.ReluActivator(), conv0)
self.graph.add_layer(act0)
pool0 = nn.Op(nn.MaxPooling(filter_h=2, filter_w=2, stride=2), act0)
self.graph.add_op(pool0)
conv1 = nn.Op(nn.Conv2d(padding='valid'), pool0, [W1, b1])
self.graph.add_op(conv1)
act1 = nn.Layer(nn.ReluActivator(), conv1)
self.graph.add_layer(act1)
pool1 = nn.Op(nn.MaxPooling(filter_h=2, filter_w=2, stride=2), act1)
self.graph.add_op(pool1)
fla = nn.Op(nn.Flatten(), pool1)
self.graph.add_op(fla)
FC0 = nn.Op(nn.Dot(), fla, [WFC0, bFC0])
self.graph.add_op(FC0)
Fact0 = nn.Layer(nn.ReluActivator(), FC0)
self.graph.add_layer(Fact0)
dp0 = nn.Op(nn.Dropout(0.3), Fact0)
self.graph.add_op(dp0)
FC1 = nn.Op(nn.Dot(), dp0, [WFC1, bFC1])
self.graph.add_op(FC1)
# Fact1 = nn.Layer(nn.IdentityActivator(), FC1)
# self.graph.add_layer(Fact1)
self.graph.add_loss(nn.Loss(nn.Softmax()))
self.graph.add_optimizer((nn.AdamOptimizer()))
def __init__(self, imagesize, flatten_size, num_labels, lr=1e-4):
self.N, self.C, self.W, self.H = imagesize
self.num_labels = num_labels
self.lr = lr
self.graph = nn.Graph()
self.X = nn.Placeholder()
W0 = nn.Variable(nn.UniformInit([3, 3, self.C, 64]), lr=self.lr)
self.graph.add_var(W0)
b0 = nn.Variable(nn.UniformInit([64, 1]), lr=lr)
self.graph.add_var(b0)
conv0 = nn.Op(nn.Conv2d(padding='same'), self.X, [W0, b0])
self.graph.add_op(conv0)
act0 = nn.Layer(nn.ReluActivator(), conv0)
self.graph.add_layer(act0)
W1 = nn.Variable(nn.UniformInit([3, 3, 64, 64]), lr=self.lr)
self.graph.add_var(W1)
b1 = nn.Variable(nn.UniformInit([64, 1]), lr=lr)
self.graph.add_var(b1)
conv1 = nn.Op(nn.Conv2d(padding='same'), act0, [W1, b1])
self.graph.add_op(conv1)
act1 = nn.Layer(nn.ReluActivator(), conv1)
self.graph.add_layer(act1)
pool1 = nn.Op(nn.MaxPooling(filter_h=2, filter_w=2, stride=2), act1)
self.graph.add_op(pool1)
W2 = nn.Variable(nn.UniformInit([3, 3, 64, 128]), lr=self.lr)
self.graph.add_var(W2)
b2 = nn.Variable(nn.UniformInit([128, 1]), lr=lr)
self.graph.add_var(b2)
conv2 = nn.Op(nn.Conv2d(padding='same'), pool1, [W2, b2])
self.graph.add_op(conv2)
act2 = nn.Layer(nn.ReluActivator(), conv2)
self.graph.add_layer(act2)
W3 = nn.Variable(nn.UniformInit([3, 3, 128, 128]), lr=self.lr)
self.graph.add_var(W3)
b3 = nn.Variable(nn.UniformInit([128, 1]), lr=lr)
self.graph.add_var(b3)
conv3 = nn.Op(nn.Conv2d(padding='same'), act2, [W3, b3])
self.graph.add_op(conv3)
act3 = nn.Layer(nn.ReluActivator(), conv3)
self.graph.add_layer(act3)
pool3 = nn.Op(nn.MaxPooling(filter_h=2, filter_w=2, stride=2), act3)
self.graph.add_op(pool3)
W4 = nn.Variable(nn.UniformInit([3, 3, 128, 256]), lr=self.lr)
self.graph.add_var(W4)
b4 = nn.Variable(nn.UniformInit([256, 1]), lr=lr)
self.graph.add_var(b4)
conv4 = nn.Op(nn.Conv2d(padding='same'), pool3, [W4, b4])
self.graph.add_op(conv4)
act4 = nn.Layer(nn.ReluActivator(), conv4)
self.graph.add_layer(act4)
W5 = nn.Variable(nn.UniformInit([3, 3, 256, 256]), lr=self.lr)
self.graph.add_var(W5)
b5 = nn.Variable(nn.UniformInit([256, 1]), lr=lr)
self.graph.add_var(b5)
conv5 = nn.Op(nn.Conv2d(padding='same'), act4, [W5, b5])
self.graph.add_op(conv5)
act5 = nn.Layer(nn.ReluActivator(), conv5)
self.graph.add_layer(act5)
W6 = nn.Variable(nn.UniformInit([3, 3, 256, 256]), lr=self.lr)
self.graph.add_var(W6)
b6 = nn.Variable(nn.UniformInit([256, 1]), lr=lr)
self.graph.add_var(b6)
conv6 = nn.Op(nn.Conv2d(padding='same'), act5, [W6, b6])
self.graph.add_op(conv6)
act6 = nn.Layer(nn.ReluActivator(), conv6)
self.graph.add_layer(act6)
W7 = nn.Variable(nn.UniformInit([3, 3, 256, 256]), lr=self.lr)
self.graph.add_var(W7)
b7 = nn.Variable(nn.UniformInit([256, 1]), lr=lr)
self.graph.add_var(b7)
conv7 = nn.Op(nn.Conv2d(padding='same'), act6, [W7, b7])
self.graph.add_op(conv7)
act7 = nn.Layer(nn.ReluActivator(), conv7)
self.graph.add_layer(act7)
pool8 = nn.Op(nn.MaxPooling(filter_h=2, filter_w=2, stride=2), act7)
self.graph.add_op(pool8)
W9 = nn.Variable(nn.UniformInit([3, 3, 256, 512]), lr=self.lr)
self.graph.add_var(W9)
b9 = nn.Variable(nn.UniformInit([512, 1]), lr=lr)
self.graph.add_var(b9)
conv9 = nn.Op(nn.Conv2d(padding='same'), pool8, [W9, b9])
self.graph.add_op(conv9)
act9 = nn.Layer(nn.ReluActivator(), conv9)
self.graph.add_layer(act9)
W10 = nn.Variable(nn.UniformInit([3, 3, 512, 512]), lr=self.lr)
self.graph.add_var(W10)
b10 = nn.Variable(nn.UniformInit([512, 1]), lr=lr)
self.graph.add_var(b10)
conv10 = nn.Op(nn.Conv2d(padding='same'), act9, [W10, b10])
self.graph.add_op(conv10)
act10 = nn.Layer(nn.ReluActivator(), conv10)
self.graph.add_layer(act10)
W11 = nn.Variable(nn.UniformInit([3, 3, 512, 512]), lr=self.lr)
self.graph.add_var(W11)
b11 = nn.Variable(nn.UniformInit([512, 1]), lr=lr)
self.graph.add_var(b11)
conv11 = nn.Op(nn.Conv2d(padding='same'), act10, [W11, b11])
self.graph.add_op(conv11)
act11 = nn.Layer(nn.ReluActivator(), conv11)
self.graph.add_layer(act11)
pool12 = nn.Op(nn.MaxPooling(filter_h=2, filter_w=2, stride=2), act11)
self.graph.add_op(pool12)
W13 = nn.Variable(nn.UniformInit([3, 3, 512, 512]), lr=self.lr)
self.graph.add_var(W13)
b13 = nn.Variable(nn.UniformInit([512, 1]), lr=lr)
self.graph.add_var(b13)
conv13 = nn.Op(nn.Conv2d(padding='same'), pool12, [W13, b13])
self.graph.add_op(conv13)
act13 = nn.Layer(nn.ReluActivator(), conv13)
self.graph.add_layer(act13)
W14 = nn.Variable(nn.UniformInit([3, 3, 512, 512]), lr=self.lr)
self.graph.add_var(W14)
b14 = nn.Variable(nn.UniformInit([512, 1]), lr=lr)
self.graph.add_var(b14)
conv14 = nn.Op(nn.Conv2d(padding='same'), act13, [W14, b14])
self.graph.add_op(conv14)
act14 = nn.Layer(nn.ReluActivator(), conv14)
self.graph.add_layer(act14)
W15 = nn.Variable(nn.UniformInit([3, 3, 512, 512]), lr=self.lr)
self.graph.add_var(W15)
b15 = nn.Variable(nn.UniformInit([512, 1]), lr=lr)
self.graph.add_var(b15)
conv15 = nn.Op(nn.Conv2d(padding='same'), act14, [W15, b15])
self.graph.add_op(conv15)
act15 = nn.Layer(nn.ReluActivator(), conv15)
self.graph.add_layer(act15)
pool16 = nn.Op(nn.MaxPooling(filter_h=2, filter_w=2, stride=2), act15)
self.graph.add_op(pool16)
fla = nn.Op(nn.Flatten(), pool16)
self.graph.add_op(fla)
WFC0 = nn.Variable(nn.UniformInit([flatten_size, 4096]), lr=lr)
self.graph.add_var(WFC0)
bFC0 = nn.Variable(nn.UniformInit([4096, 1]), lr=lr)
self.graph.add_var(bFC0)
FC0 = nn.Op(nn.Dot(), fla, [WFC0, bFC0])
self.graph.add_op(FC0)
Fact0 = nn.Layer(nn.ReluActivator(), FC0)
self.graph.add_layer(Fact0)
WFC1 = nn.Variable(nn.UniformInit([4096, 4096]), lr=lr)
self.graph.add_var(WFC1)
bFC1 = nn.Variable(nn.UniformInit([4096, 1]), lr=lr)
self.graph.add_var(bFC1)
FC1 = nn.Op(nn.Dot(), Fact0, [WFC1, bFC1])
self.graph.add_op(FC1)
Fact1 = nn.Layer(nn.ReluActivator(), FC1)
self.graph.add_layer(Fact1)
WFC2 = nn.Variable(nn.UniformInit([4096, num_labels]), lr=lr)
self.graph.add_var(WFC2)
bFC2 = nn.Variable(nn.UniformInit([num_labels, 1]), lr=lr)
self.graph.add_var(bFC2)
FC2 = nn.Op(nn.Dot(), Fact1, [WFC2, bFC2])
self.graph.add_op(FC2)
# Fact2 = nn.Layer(nn.ReluActivator(), FC2)
# self.graph.add_layer(Fact2)
self.graph.add_loss(nn.Loss(nn.Softmax()))
self.graph.add_optimizer((nn.AdamOptimizer()))
def __init__(self, imagesize, flatten_size, num_labels, lr=1e-3):
self.N, self.C, self.W, self.H = imagesize
self.num_labels = num_labels
self.lr = lr
self.graph = nn.Graph()
self.X = nn.Placeholder()
W0 = nn.Variable(nn.UniformInit([11, 11, self.C, 96]), lr=lr)
b0 = nn.Variable(nn.UniformInit([96, 1]), lr=lr)
self.graph.add_vars([W0, b0])
conv0 = nn.Op(nn.Conv2d(padding='same', stride=4), self.X, [W0, b0])
self.graph.add_op(conv0)
act0 = nn.Layer(nn.ReluActivator(), conv0)
self.graph.add_layer(act0)
pool0 = nn.Op(nn.MaxPooling(filter_h=3, filter_w=3, stride=2), act0)
self.graph.add_op(pool0)
W1 = nn.Variable(nn.UniformInit([5, 5, 96, 256]), lr=lr)
b1 = nn.Variable(nn.UniformInit([256, 1]), lr=lr)
self.graph.add_vars([W1, b1])
conv1 = nn.Op(nn.Conv2d(padding='same', stride=1), pool0, [W1, b1])
self.graph.add_op(conv1)
act1 = nn.Layer(nn.ReluActivator(), conv1)
self.graph.add_layer(act1)
pool1 = nn.Op(nn.MaxPooling(filter_h=3, filter_w=3, stride=2), act1)
self.graph.add_op(pool1)
W2 = nn.Variable(nn.UniformInit([5, 5, 256, 384]), lr=lr)
b2 = nn.Variable(nn.UniformInit([384, 1]), lr=lr)
self.graph.add_vars([W2, b2])
conv2 = nn.Op(nn.Conv2d(padding='same', stride=1), pool1, [W2, b2])
self.graph.add_op(conv2)
act2 = nn.Layer(nn.ReluActivator(), conv2)
self.graph.add_layer(act2)
W3 = nn.Variable(nn.UniformInit([5, 5, 384, 384]), lr=lr)
b3 = nn.Variable(nn.UniformInit([384, 1]), lr=lr)
self.graph.add_vars([W3, b3])
conv3 = nn.Op(nn.Conv2d(padding='same', stride=1), act2, [W3, b3])
self.graph.add_op(conv3)
act3 = nn.Layer(nn.ReluActivator(), conv3)
self.graph.add_layer(act3)
W4 = nn.Variable(nn.UniformInit([5, 5, 384, 256]), lr=lr)
b4 = nn.Variable(nn.UniformInit([256, 1]), lr=lr)
self.graph.add_vars([W4, b4])
conv4 = nn.Op(nn.Conv2d(padding='same', stride=1), act3, [W4, b4])
self.graph.add_op(conv4)
act4 = nn.Layer(nn.ReluActivator(), conv4)
self.graph.add_layer(act4)
pool5 = nn.Op(nn.MaxPooling(filter_h=3, filter_w=3, stride=2), act4)
self.graph.add_op(pool5)
fla = nn.Op(nn.Flatten(), pool5)
self.graph.add_op(fla)
WFC0 = nn.Variable(nn.UniformInit([flatten_size, 4096]), lr=lr)
bFC0 = nn.Variable(nn.UniformInit([4096, 1]), lr=lr)
FC0 = nn.Op(nn.Dot(), fla, [WFC0, bFC0])
self.graph.add_op(FC0)
Fact0 = nn.Layer(nn.ReluActivator(), FC0)
self.graph.add_layer(Fact0)
dp0 = nn.Op(nn.Dropout(0.3), Fact0)
self.graph.add_op(dp0)
WFC1 = nn.Variable(nn.UniformInit([4096, 4096]), lr=lr)
bFC1 = nn.Variable(nn.UniformInit([4096, 1]), lr=lr)
FC1 = nn.Op(nn.Dot(), dp0, [WFC1, bFC1])
self.graph.add_op(FC1)
Fact1 = nn.Layer(nn.ReluActivator(), FC1)
self.graph.add_layer(Fact1)
dp1 = nn.Op(nn.Dropout(0.3), Fact1)
self.graph.add_op(dp1)
WFC2 = nn.Variable(nn.UniformInit([4096, 1000]), lr=lr)
bFC2 = nn.Variable(nn.UniformInit([1000, 1]), lr=lr)
FC2 = nn.Op(nn.Dot(), dp1, [WFC2, bFC2])
self.graph.add_op(FC2)
Fact2 = nn.Layer(nn.ReluActivator(), FC2)
self.graph.add_layer(Fact2)
dp2 = nn.Op(nn.Dropout(0.3), Fact2)
self.graph.add_op(dp2)
WFC3 = nn.Variable(nn.UniformInit([1000, num_labels]), lr=lr)
bFC3 = nn.Variable(nn.UniformInit([num_labels, 1]), lr=lr)
FC3 = nn.Op(nn.Dot(), dp2, [WFC3, bFC3])
self.graph.add_op(FC3)
Fact3 = nn.Layer(nn.ReluActivator(), FC3)
self.graph.add_layer(Fact3)
dp3 = nn.Op(nn.Dropout(0.3), Fact3)
self.graph.add_op(dp3)
self.graph.add_loss(nn.Loss(nn.Softmax()))
self.graph.add_optimizer((nn.AdamOptimizer()))
te_Y += label
tr_X, te_X = np.vstack(tr_X), np.vstack(te_X)
tr_Y, te_Y = np.array(tr_Y), np.array(te_Y)
tr_X = nn.scaleallimage(tr_X, (112, 112))
te_X = nn.scaleallimage(te_X, (112, 112))
tr_X, te_X = tr_X.astype(np.float32) / 127.5 - 1.0, te_X.astype(
np.float32) / 127.5 - 1.0
return tr_X, tr_Y, te_X, te_Y
X_train, Y_train, X_test, Y_test = readdata()
print(X_train.shape, Y_train.shape, X_test.shape, Y_test.shape)
X_train = nn.Variable(X_train, lr=0)
Y_train = np.expand_dims(Y_train, axis=1)
Y_train = nn.onehot(Y_train)
Y_train = nn.Variable(Y_train, lr=0)
X_test = nn.Variable(X_test, lr=0)
Y_test = np.expand_dims(Y_test, axis=1)
Y_test = nn.onehot(Y_test)
Y_test = nn.Variable(Y_test, lr=0)
# lenet = LeNet(X_train.value.shape, 512, 10)
# lenet.train(X_train, Y_train, X_test, Y_test, epochs=10, batchsize=10)
# res18 = ResNet18(X_train.value.shape, 25088, 10)
def __init__(self, imagesize, flatten_size, num_labels, lr=1e-3):
self.N, self.C, self.W, self.H = imagesize
self.num_labels = num_labels
self.lr = lr
self.graph = nn.Graph()
X = nn.Placeholder()
W0 = nn.Variable(nn.UniformInit([3, 3, self.C, 64]), lr=lr)
b0 = nn.Variable(nn.UniformInit([64, 1]), lr=lr)
gamma0 = nn.Variable(nn.UniformInit([1, 64, self.W, self.H]), lr=lr)
beta0 = nn.Variable(nn.UniformInit([1, 64, self.W, self.H]), lr=lr)
self.graph.add_vars([W0, b0, gamma0, beta0])
conv0 = nn.Op(nn.Conv2d(), X, [W0, b0])
bn0 = nn.Op(nn.BatchNorm(gamma0, beta0), conv0)
act0 = nn.Layer(nn.ReluActivator(), bn0)
self.graph.add_ops([conv0, bn0])
self.graph.add_layer(act0)
# Block 1
W1_1 = nn.Variable(nn.UniformInit([3, 3, 64, 64]), lr=lr)
b1_1 = nn.Variable(nn.UniformInit([64, 1]), lr=lr)
gamma1_1 = nn.Variable(nn.UniformInit([1, 64, self.W, self.H]), lr=lr)
beta1_1 = nn.Variable(nn.UniformInit([1, 64, self.W, self.H]), lr=lr)
W1_2 = nn.Variable(nn.UniformInit([3, 3, 64, 64]), lr=lr)
b1_2 = nn.Variable(nn.UniformInit([64, 1]), lr=lr)
gamma1_2 = nn.Variable(nn.UniformInit([1, 64, self.W, self.H]), lr=lr)
beta1_2 = nn.Variable(nn.UniformInit([1, 64, self.W, self.H]), lr=lr)
self.graph.add_vars([W1_1, b1_1, gamma1_1, beta1_1, \
W1_2, b1_2, gamma1_2, beta1_2])
pamas1 = {'w1': W1_1, 'b1': b1_1, \
'gamma1': gamma1_1, 'beta1': beta1_1, \
'w2': W1_2, 'b2': b1_2, \
'gamma2': gamma1_2, 'beta2': beta1_2}
B1 = nn.ResBlock(act0, pamas1)
self.graph.add_block(B1)
# Block 2
W2_1 = nn.Variable(nn.UniformInit([3, 3, 64, 64]), lr=lr)
b2_1 = nn.Variable(nn.UniformInit([64, 1]), lr=lr)
gamma2_1 = nn.Variable(nn.UniformInit([1, 64, self.W, self.H]), lr=lr)
beta2_1 = nn.Variable(nn.UniformInit([1, 64, self.W, self.H]), lr=lr)
W2_2 = nn.Variable(nn.UniformInit([3, 3, 64, 64]), lr=lr)
b2_2 = nn.Variable(nn.UniformInit([64, 1]), lr=lr)
gamma2_2 = nn.Variable(nn.UniformInit([1, 64, self.W, self.H]), lr=lr)
beta2_2 = nn.Variable(nn.UniformInit([1, 64, self.W, self.H]), lr=lr)
self.graph.add_vars([W2_1, b2_1, gamma2_1, beta2_1, \
W2_2, b2_2, gamma2_2, beta2_2])
pamas2 = {'w1': W2_1, 'b1': b2_1, \
'gamma1': gamma2_1, 'beta1': beta2_1, \
'w2': W2_2, 'b2': b2_2, \
'gamma2': gamma2_2, 'beta2': beta2_2}
B2 = nn.ResBlock(B1, pamas2)
self.graph.add_block(B2)
# pool2
pool2 = nn.Op(nn.MaxPooling(2, 2), B2)
self.graph.add_layer(pool2)
# Block 3
W3_1 = nn.Variable(nn.UniformInit([3, 3, 64, 128]), lr=lr)
b3_1 = nn.Variable(nn.UniformInit([128, 1]), lr=lr)
gamma3_1 = nn.Variable(nn.UniformInit([1, 128, self.W / 2,
self.H / 2]),
lr=lr)
beta3_1 = nn.Variable(nn.UniformInit([1, 128, self.W / 2, self.H / 2]),
lr=lr)
W3_2 = nn.Variable(nn.UniformInit([3, 3, 128, 128]), lr=lr)
b3_2 = nn.Variable(nn.UniformInit([128, 1]), lr=lr)
gamma3_2 = nn.Variable(nn.UniformInit([1, 128, self.W / 2,
self.H / 2]),
lr=lr)
beta3_2 = nn.Variable(nn.UniformInit([1, 128, self.W / 2, self.H / 2]),
lr=lr)
w_sc3 = nn.Variable(nn.UniformInit([3, 3, 64, 128]), lr=lr)
b_sc3 = nn.Variable(nn.UniformInit([128, 1]), lr=lr)
gamma_sc3 = nn.Variable(nn.UniformInit(
[1, 128, self.W / 2, self.H / 2]),
lr=lr)
beta3_sc3 = nn.Variable(nn.UniformInit(
[1, 128, self.W / 2, self.H / 2]),
lr=lr)
self.graph.add_vars([W3_1, b3_1, gamma3_1, beta3_1, \
W3_2, b3_2, gamma3_2, beta3_2, \
w_sc3, b_sc3, gamma_sc3, beta3_sc3])
pamas3 = {'w1': W3_1, 'b1': b3_1, \
'gamma1': gamma3_1, 'beta1': beta3_1, \
'w2': W3_2, 'b2': b3_2, \
'gamma2': gamma3_2, 'beta2': beta3_2}
sc3 = {'w': w_sc3, 'b': b_sc3, 'gamma': gamma_sc3, 'beta': beta3_sc3}
B3 = nn.ResBlock(pool2, pamas3, sc3)
self.graph.add_block(B3)
# Block 4
W4_1 = nn.Variable(nn.UniformInit([3, 3, 128, 128]), lr=lr)
b4_1 = nn.Variable(nn.UniformInit([128, 1]), lr=lr)
gamma4_1 = nn.Variable(nn.UniformInit([1, 128, self.W / 2,
self.H / 2]),
lr=lr)
beta4_1 = nn.Variable(nn.UniformInit([1, 128, self.W / 2, self.H / 2]),
lr=lr)
W4_2 = nn.Variable(nn.UniformInit([3, 3, 128, 128]), lr=lr)
b4_2 = nn.Variable(nn.UniformInit([128, 1]), lr=lr)
gamma4_2 = nn.Variable(nn.UniformInit([1, 128, self.W / 2,
self.H / 2]),
lr=lr)
beta4_2 = nn.Variable(nn.UniformInit([1, 128, self.W / 2, self.H / 2]),
lr=lr)
self.graph.add_vars([W4_1, b4_1, gamma4_1, beta4_1, \
W4_2, b4_2, gamma4_2, beta4_2])
pamas4 = {'w1': W4_1, 'b1': b4_1, \
'gamma1': gamma4_1, 'beta1': beta4_1, \
'w2': W4_2, 'b2': b4_2, \
'gamma2': gamma4_2, 'beta2': beta4_2}
B4 = nn.ResBlock(B3, pamas4)
self.graph.add_block(B4)
# pool4
pool4 = nn.Op(nn.MaxPooling(2, 2), B4)
self.graph.add_layer(pool4)
# Block 5
W5_1 = nn.Variable(nn.UniformInit([3, 3, 128, 256]), lr=lr)
b5_1 = nn.Variable(nn.UniformInit([256, 1]), lr=lr)
gamma5_1 = nn.Variable(nn.UniformInit([1, 256, self.W / 4,
self.H / 4]),
lr=lr)
beta5_1 = nn.Variable(nn.UniformInit([1, 256, self.W / 4, self.H / 4]),
lr=lr)
W5_2 = nn.Variable(nn.UniformInit([3, 3, 256, 256]), lr=lr)
b5_2 = nn.Variable(nn.UniformInit([256, 1]), lr=lr)
gamma5_2 = nn.Variable(nn.UniformInit([1, 256, self.W / 4,
self.H / 4]),
lr=lr)
beta5_2 = nn.Variable(nn.UniformInit([1, 256, self.W / 4, self.H / 4]),
lr=lr)
w_sc5 = nn.Variable(nn.UniformInit([3, 3, 128, 256]), lr=lr)
b_sc5 = nn.Variable(nn.UniformInit([256, 1]), lr=lr)
gamma_sc5 = nn.Variable(nn.UniformInit(
[1, 256, self.W / 4, self.H / 4]),
lr=lr)
beta3_sc5 = nn.Variable(nn.UniformInit(
[1, 256, self.W / 4, self.H / 4]),
lr=lr)
self.graph.add_vars([W5_1, b5_1, gamma5_1, beta5_1, \
W5_2, b5_2, gamma5_2, beta5_2, \
w_sc5, b_sc5, gamma_sc5, beta3_sc5])
pamas5 = {'w1': W5_1, 'b1': b5_1, \
'gamma1': gamma5_1, 'beta1': beta5_1, \
'w2': W5_2, 'b2': b5_2, \
'gamma2': gamma5_2, 'beta2': beta5_2}
sc5 = {'w': w_sc5, 'b': b_sc5, 'gamma': gamma_sc5, 'beta': beta3_sc5}
B5 = nn.ResBlock(pool4, pamas5, sc5)
self.graph.add_block(B5)
# Block 6
W6_1 = nn.Variable(nn.UniformInit([3, 3, 256, 256]), lr=lr)
b6_1 = nn.Variable(nn.UniformInit([256, 1]), lr=lr)
gamma6_1 = nn.Variable(nn.UniformInit([1, 256, self.W / 4,
self.H / 4]),
lr=lr)
beta6_1 = nn.Variable(nn.UniformInit([1, 256, self.W / 4, self.H / 4]),
lr=lr)
W6_2 = nn.Variable(nn.UniformInit([3, 3, 256, 256]), lr=lr)
b6_2 = nn.Variable(nn.UniformInit([256, 1]), lr=lr)
gamma6_2 = nn.Variable(nn.UniformInit([1, 256, self.W / 4,
self.H / 4]),
lr=lr)
beta6_2 = nn.Variable(nn.UniformInit([1, 256, self.W / 4, self.H / 4]),
lr=lr)
self.graph.add_vars([W6_1, b6_1, gamma6_1, beta6_1, \
W6_2, b6_2, gamma6_2, beta6_2])
pamas6 = {'w1': W6_1, 'b1': b6_1, \
'gamma1': gamma6_1, 'beta1': beta6_1, \
'w2': W6_2, 'b2': b6_2, \
'gamma2': gamma6_2, 'beta2': beta6_2}
B6 = nn.ResBlock(B5, pamas6)
self.graph.add_block(B6)
# pool6
pool6 = nn.Op(nn.MaxPooling(2, 2), B6)
self.graph.add_layer(pool6)
# Block 7
W7_1 = nn.Variable(nn.UniformInit([3, 3, 256, 512]), lr=lr)
b7_1 = nn.Variable(nn.UniformInit([512, 1]), lr=lr)
gamma7_1 = nn.Variable(nn.UniformInit([1, 512, self.W / 8,
self.H / 8]),
lr=lr)
beta7_1 = nn.Variable(nn.UniformInit([1, 512, self.W / 8, self.H / 8]),
lr=lr)
W7_2 = nn.Variable(nn.UniformInit([3, 3, 512, 512]), lr=lr)
b7_2 = nn.Variable(nn.UniformInit([512, 1]), lr=lr)
gamma7_2 = nn.Variable(nn.UniformInit([1, 512, self.W / 8,
self.H / 8]),
lr=lr)
beta7_2 = nn.Variable(nn.UniformInit([1, 512, self.W / 8, self.H / 8]),
lr=lr)
w_sc7 = nn.Variable(nn.UniformInit([3, 3, 256, 512]), lr=lr)
b_sc7 = nn.Variable(nn.UniformInit([512, 1]), lr=lr)
gamma_sc7 = nn.Variable(nn.UniformInit(
[1, 512, self.W / 8, self.H / 8]),
lr=lr)
beta3_sc7 = nn.Variable(nn.UniformInit(
[1, 512, self.W / 8, self.H / 8]),
lr=lr)
self.graph.add_vars([W7_1, b7_1, gamma7_1, beta7_1, \
W7_2, b7_2, gamma7_2, beta7_2, \
w_sc7, b_sc7, gamma_sc7, beta3_sc7])
pamas7 = {'w1': W7_1, 'b1': b7_1, \
'gamma1': gamma7_1, 'beta1': beta7_1, \
'w2': W7_2, 'b2': b7_2, \
'gamma2': gamma7_2, 'beta2': beta7_2}
sc7 = {'w': w_sc7, 'b': b_sc7, 'gamma': gamma_sc7, 'beta': beta3_sc7}
B7 = nn.ResBlock(pool6, pamas7, sc7)
self.graph.add_block(B7)
# Block 8
W8_1 = nn.Variable(nn.UniformInit([3, 3, 512, 512]), lr=lr)
b8_1 = nn.Variable(nn.UniformInit([512, 1]), lr=lr)
gamma8_1 = nn.Variable(nn.UniformInit([1, 512, self.W / 8,
self.H / 8]),
lr=lr)
beta8_1 = nn.Variable(nn.UniformInit([1, 512, self.W / 8, self.H / 8]),
lr=lr)
W8_2 = nn.Variable(nn.UniformInit([3, 3, 512, 512]), lr=lr)
b8_2 = nn.Variable(nn.UniformInit([512, 1]), lr=lr)
gamma8_2 = nn.Variable(nn.UniformInit([1, 512, self.W / 8,
self.H / 8]),
lr=lr)
beta8_2 = nn.Variable(nn.UniformInit([1, 512, self.W / 8, self.H / 8]),
lr=lr)
self.graph.add_vars([W8_1, b8_1, gamma8_1, beta8_1, \
W8_2, b8_2, gamma8_2, beta8_2])
pamas8 = {'w1': W8_1, 'b1': b8_1, \
'gamma1': gamma8_1, 'beta1': beta8_1, \
'w2': W8_2, 'b2': b8_2, \
'gamma2': gamma8_2, 'beta2': beta8_2}
B8 = nn.ResBlock(B7, pamas8)
self.graph.add_block(B8)
# pool9
pool9 = nn.Op(nn.MaxPooling(2, 2), B8)
self.graph.add_layer(pool9)
# flatten
fla = nn.Op(nn.Flatten(), pool9)
self.graph.add_op(fla)
# Fc10
WFC10 = nn.Variable(nn.UniformInit([flatten_size, num_labels]), lr=lr)
self.graph.add_var(WFC10)
bFC10 = nn.Variable(nn.UniformInit([num_labels, 1]), lr=lr)
self.graph.add_var(bFC10)
FC10 = nn.Op(nn.Dot(), fla, [WFC10, bFC10])
self.graph.add_op(FC10)
# loss and optimizer
self.graph.add_loss(nn.Loss(nn.Softmax()))
self.graph.add_optimizer((nn.AdamOptimizer()))