def add_n_bottleneck_blocks(in_layer, n_outfmaps, n_repeat, is_first_layer):
assert n_outfmaps % 4 == 0
a0 = in_layer
for i1 in xrange(n_repeat):
if i1 == 0:
if is_first_layer is True:
strides = (1, 1)
shortcut = a0
else:
strides = (2, 2)
shortcut = Convolution2D(n_outfmaps=n_outfmaps,
n_row=1,
n_col=1,
act='linear',
border_mode='valid',
strides=strides)(a0)
else:
strides = (1, 1)
shortcut = a0
a1 = BN(axes=(0, 2, 3))(a0)
a2 = Activation('relu')(a1)
a3 = Convolution2D(n_outfmaps=n_outfmaps,
n_row=1,
n_col=1,
act='linear',
border_mode='valid',
strides=strides)(a2)
a4 = BN(axes=(0, 2, 3))(a3)
a5 = Activation('relu')(a4)
a6 = Convolution2D(n_outfmaps=n_outfmaps / 4,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1),
strides=(1, 1))(a5)
a7 = BN(axes=(0, 2, 3))(a6)
a8 = Activation('relu')(a7)
a9 = Convolution2D(n_outfmaps=n_outfmaps,
n_row=1,
n_col=1,
act='linear',
border_mode='valid',
strides=(1, 1))(a8)
a10 = Lambda(add_layers)([shortcut, a9])
a0 = a10
return a0
def add_n_blocks(in_layer, n_outfmaps, n_repeat, is_first_layer):
a0 = in_layer
for i1 in range(n_repeat):
if i1 == 0:
if is_first_layer is True:
strides = (1, 1)
shortcut = a0
else:
strides = (2, 2)
shortcut = Convolution2D(n_outfmaps=n_outfmaps,
n_row=1,
n_col=1,
act='linear',
border_mode='valid',
strides=strides)(a0)
else:
strides = (1, 1)
shortcut = a0
a1 = Convolution2D(n_outfmaps=n_outfmaps,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1),
strides=strides)(a0)
a2 = BN(axes=(0, 2, 3))(a1)
a3 = Activation('relu')(a2)
a4 = Convolution2D(n_outfmaps=n_outfmaps,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1),
strides=(1, 1))(a3)
a5 = BN(axes=(0, 2, 3))(a4)
a6 = Activation('relu')(a5)
a7 = Lambda(add_layers)([shortcut, a6])
a0 = a7
return a0
def add_n_blocks(seq, n_outfmaps, n_repeat, is_first_layer):
for i1 in range(n_repeat):
if i1 == 0:
if is_first_layer is True: strides = (1, 1)
else: strides = (2, 2)
else:
strides = (1, 1)
seq.add(
Convolution2D(n_outfmaps=n_outfmaps,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1),
strides=strides))
seq.add(BN(axes=(0, 2, 3)))
seq.add(Activation('relu'))
return seq
# init params n_in = 784 n_hid = 500 n_out = 10 # sparse label to 1 of K categorical label tr_y = sparse_to_categorical(tr_y, n_out) va_y = sparse_to_categorical(va_y, n_out) te_y = sparse_to_categorical(te_y, n_out) ### Build model act = 'relu' seq = Sequential() seq.add(InputLayer(in_shape=(1, 28, 28))) seq.add(Convolution2D(n_outfmaps=32, n_row=3, n_col=3, act='relu')) seq.add(MaxPool2D(pool_size=(2, 2))) seq.add(Convolution2D(n_outfmaps=32, n_row=3, n_col=3, act='relu')) seq.add(MaxPool2D(pool_size=(2, 2))) seq.add(Dropout(0.2)) seq.add(Flatten()) seq.add(Dense(n_hid, act='relu')) seq.add(Dropout(0.5)) seq.add(Dense(n_hid, act='relu')) seq.add(Dense(n_out, act='softmax')) md = seq.combine() # print summary info of model md.summary() # optimization method
n_out = 10
# sparse label to 1-of-K categorical label
tr_y = sparse_to_categorical(tr_y, n_out)
te_y = sparse_to_categorical(te_y, n_out)
print tr_X.shape
print tr_y.shape
### Build model
seq = Sequential()
seq.add(InputLayer(in_shape=(3, 32, 32)))
seq.add(
Convolution2D(n_outfmaps=64,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1)))
seq.add(BN(axes=(0, 2, 3)))
seq.add(Activation('relu'))
seq.add(
Convolution2D(n_outfmaps=64,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1)))
seq.add(BN(axes=(0, 2, 3)))
seq.add(Activation('relu'))
seq.add(MaxPool2D(pool_size=(2, 2)))
seq.add(
border_mode=(1, 1),
strides=(1, 1))(a3)
a5 = BN(axes=(0, 2, 3))(a4)
a6 = Activation('relu')(a5)
a7 = Lambda(add_layers)([shortcut, a6])
a0 = a7
return a0
x0 = InputLayer(in_shape=(3, 32, 32))
x1 = Convolution2D(n_outfmaps=64,
n_row=3,
n_col=3,
act='relu',
border_mode=(1, 1))(x0)
x2 = add_n_blocks(x1, n_outfmaps=64, n_repeat=3, is_first_layer=True)
x3 = add_n_blocks(x2, n_outfmaps=128, n_repeat=4, is_first_layer=False)
x4 = add_n_blocks(x3, n_outfmaps=256, n_repeat=6, is_first_layer=False)
x5 = add_n_blocks(x4, n_outfmaps=512, n_repeat=3, is_first_layer=False)
y1 = Lambda(mean_pool)(x5)
y2 = Flatten()(y1)
y3 = Dense(n_out, act='softmax')(y2)
md = Model([x0], [y3])
# print summary info of model
md.summary()