def f(x):
in_channels = K.int_shape(x)[channel_index]
if in_channels == filters:
y = x
if batch_norm:
y = BatchNormalization()(y)
y = make_activation(activation, **kwargs)(y)
else:
if batch_norm:
x = BatchNormalization()(x)
x = make_activation(activation, **kwargs)(x)
y = x
y = make_conv(filters, activation=activation, strides=strides, **kwargs)(y)
if batch_norm:
y = BatchNormalization()(y)
y = make_activation(activation, **kwargs)(y)
if dropout != 0:
y = make_dropout(dropout, activation=activation)(y)
y = make_conv(filters, activation=activation, **kwargs)(y)
if not residual:
return y
shortcut = x if in_channels == filters else make_conv(filters, shape=(1, 1), strides=strides, **kwargs)(x)
return Add()([y, shortcut])
def f(x):
y = make_activation(activation, **kwargs)(x)
y = make_conv(filters, strides=strides)(y)
if batch_norm:
y = BatchNormalization()(y)
y = make_activation(activation, **kwargs)(y)
y = make_conv(filters)(y)
if batch_norm:
y = BatchNormalization()(y)
return y
def make_wrn(activation, input_shape, output_size, depth=4, k=2, dropout=True, batch_norm=True, residual=True, **kwargs):
assert((depth - 4) % 6 == 0)
n = (depth - 4) / 6
dropout = 0.3 if dropout else 0
x = Input(shape=input_shape)
y = x
y = make_conv(16, activation=activation, **kwargs)(y)
for _ in range(n):
y = _make_block(16*k, activation=activation, batch_norm=batch_norm, dropout=dropout, residual=residual, **kwargs)(y)
y = _make_block(32*k, activation=activation, batch_norm=batch_norm, dropout=dropout, strides=(2, 2), residual=residual, **kwargs)(y)
for _ in range(n-1):
y = _make_block(32*k, activation=activation, batch_norm=batch_norm, dropout=dropout, residual=residual, **kwargs)(y)
y = _make_block(64*k, activation=activation, batch_norm=batch_norm, dropout=dropout, strides=(2, 2), residual=residual, **kwargs)(y)
for _ in range(n-1):
y = _make_block(64*k, activation=activation, batch_norm=batch_norm, dropout=dropout, residual=residual, **kwargs)(y)
if batch_norm:
y = BatchNormalization()(y)
y = make_activation(activation, **kwargs)(y)
y = GlobalMaxPooling2D()(y)
y = make_dense(output_size, activation='softmax', **kwargs)(y)
return Model(inputs=x, outputs=y)
def make_shakeshake(activation,
input_shape,
output_size,
depth=26,
k=32,
batch_norm=True,
residual=True,
**kwargs):
assert ((depth - 2) % 6 == 0)
n = (depth - 2) / 6
x = Input(batch_shape=(kwargs['batch_size'], ) + input_shape)
y = make_conv(16)(x)
y = BatchNormalization()(y)
for _ in range(n):
y = _make_block(k,
activation=activation,
batch_norm=batch_norm,
residual=residual,
**kwargs)(y)
y = _make_block(2 * k,
activation=activation,
batch_norm=batch_norm,
residual=residual,
strides=(2, 2),
**kwargs)(y)
for _ in range(n - 1):
y = _make_block(2 * k,
activation=activation,
batch_norm=batch_norm,
residual=residual,
**kwargs)(y)
y = _make_block(4 * k,
activation=activation,
batch_norm=batch_norm,
residual=residual,
strides=(2, 2),
**kwargs)(y)
for _ in range(n - 1):
y = _make_block(4 * k,
activation=activation,
batch_norm=batch_norm,
residual=residual,
**kwargs)(y)
y = make_activation(activation, **kwargs)(y)
y = GlobalAveragePooling2D()(y)
y = make_dense(output_size, activation='softmax', **kwargs)(y)
return Model(inputs=x, outputs=y)
def f(x):
in_channels = K.int_shape(x)[channel_index]
if in_channels == filters:
y = x
else:
y = make_activation(activation, **kwargs)(x)
y1 = AveragePooling2D((1, 1), strides=strides, padding='same')(y)
y1 = make_conv(filters / 2, shape=(1, 1))(y1)
y2 = ZeroPadding2D(padding=((1, 0), (1, 0)))(y)
y2 = Cropping2D(cropping=((0, 1), (0, 1)))(y2)
y2 = AveragePooling2D((1, 1), strides=strides, padding='same')(y2)
y2 = make_conv(filters / 2, shape=(1, 1))(y2)
y = Concatenate(axis=channel_index)([y1, y2])
if batch_norm:
y = BatchNormalization()(y)
return y