def resnet_block(input, strides, out_channels, name):
t = make_conv2d(input_tensor=input,
filter_shape=(1, 1, input.shape[1].value, out_channels),
strides=(1, 1, 1, 1),
padding="SAME",
actimode="RELU",
name=name + "_conv1")
t = make_conv2d(input_tensor=t,
filter_shape=(3, 3, out_channels, out_channels),
strides=strides,
padding="SAME",
actimode="RELU",
name=name + "_conv2")
t = make_conv2d(input_tensor=t,
filter_shape=(1, 1, out_channels, out_channels * 4),
strides=(1, 1, 1, 1),
padding="SAME",
actimode="NONE",
name=name + "_conv3")
if (strides[2] > 1) or (input.shape[1].value != out_channels * 4):
input = make_conv2d(input_tensor=input,
filter_shape=(1, 1, input.shape[1].value,
out_channels * 4),
strides=strides,
padding="SAME",
actimode="RELU",
name=name + "_conv4")
return tf.nn.relu(tf.add(input, t))
def resnext_block(input, strides, out_channels, groups, name):
t = make_conv2d(input_tensor=input,
filter_shape=(1, 1, input.shape[1].value, out_channels),
strides=(1, 1, 1, 1),
padding="SAME",
actimode="RELU",
name=name + "_conv1")
t = tf.split(t, groups, axis=1, name=name + "_split")
assert (len(t) == groups)
for i in range(groups):
t[i] = make_conv2d(input_tensor=t[i],
filter_shape=(3, 3, t[i].shape[1].value,
out_channels // groups),
strides=strides,
padding="SAME",
actimode="RELU",
name=name + "_conv2_".format(i))
output = tf.concat(t, axis=1, name=name + "_concat")
t = make_conv2d(input_tensor=output,
filter_shape=(1, 1, output.shape[1].value,
2 * out_channels),
strides=(1, 1, 1, 1),
padding="SAME",
actimode="NONE",
name=name + "_conv3")
if (strides[2] > 1) or (input.shape[1].value != out_channels * 2):
input = make_conv2d(input_tensor=input,
filter_shape=(1, 1, input.shape[1].value,
out_channels * 2),
strides=strides,
padding="SAME",
actimode="RELU",
name=name + "_conv4")
return tf.nn.relu(tf.add(input, t))
def squeeze(out_channels, input):
return make_conv2d(input_tensor=input,
filter_shape=(1, 1, input.shape[1].value, out_channels),
strides=(1, 1, 1, 1),
padding="SAME",
actimode="RELU",
name="squeeze")
def fit(current, input):
if (input.shape[2].value == current.shape[2].value):
return squeeze(current.shape[1].value, input)
else:
return make_conv2d(input_tensor=input,
filter_shape=(3, 3, input.shape[1].value,
current.shape[1].value),
strides=(1, 1, 2, 2),
padding="SAME",
actimode="RELU",
name="fit")