def conv2d_input(x,
filter=32,
format="NHWC",
name="conv_input",
batch_normalization=True,
layer_list=None):
"""
conv2d_input
### 引数
x: object 入力Tensor
filter： string 最初のconv2dレイヤーフィルターサイズ。
Noneの場合、入力テンソルチャネルサイズの半分になる。
format: string "NHWC"(チャネル先)and "NCHW" for チャネル最後. default は 'NHWC'
name: string 名前 デフォルトはconv_input
batch_normalization: boolean batch_normalization:するがどうか
layer_list: list 存在する場合戻り値追加する
### 戻り値
y: object「tensor」。 変換したtensor
"""
with tf.name_scope(name) as scope:
x = tf.convert_to_tensor(x)
input_shape = x.get_shape()
N, H, W, C = (0, 0, 0, 0)
if format == "NHWC":
N, H, W, C = input_shape[0].value, input_shape[
1].value, input_shape[2].value, input_shape[3].value
elif format == "NCHW":
N, C, H, W = input_shape[0].value, input_shape[
1].value, input_shape[2].value, input_shape[3].value
filter_1 = filter if filter is not None else C
y = tf_helper.add_conv2d(x,
filter_1,
h_kernel=3,
w_kernel=3,
name=scope,
h_stride=1,
w_stride=1,
format=format,
batch_normalization=batch_normalization,
activation="leaky_relu",
leaky_relu_alpha=0.1,
padding="SAME")
print("conv_input : ", y.shape)
if layer_list is not None:
layer_list.append(y)
return y
def feature_out(x,
classes_num,
format="NHWC",
name="feature_out",
batch_normalization=True,
layer_list=None):
"""
feature_out
### 引数
x: object 入力Tensor
filter： string 最初のconv2dレイヤーフィルターサイズ。 Noneの場合、入力テンソルチャネルサイズの半分になる。
format: string "NHWC"(チャネル先)and "NCHW" for チャネル最後. default は 'NHWC'
name: string 名前デフォルトはfeature_out
batch_normalization: boolean batch_normalizationするがどうか
layer_list: list 存在する場合戻り値追加する
### 戻り値
feature : object「tensor」。 変換したtensor
"""
with tf.name_scope(name) as scope:
x = tf.convert_to_tensor(x)
input_shape = x.get_shape()
if format == "NHWC":
N, H, W, C = input_shape[0].value, input_shape[
1].value, input_shape[2].value, input_shape[3].value
elif format == "NCHW":
N, C, H, W = input_shape[0].value, input_shape[
1].value, input_shape[2].value, input_shape[3].value
filter_1 = (classes_num + 5) * 3
feature = tf_helper.add_conv2d(x,
filter_1,
h_kernel=1,
w_kernel=1,
name=scope,
h_stride=1,
w_stride=1,
format=format,
batch_normalization=batch_normalization,
activation="linear",
leaky_relu_alpha=0.1,
padding="SAME")
print("conv_feature : ", feature.shape)
if layer_list is not None:
layer_list.append(feature)
return feature
def __init__(self,
input_size=784,
mid_units=100,
output_size=10,
lr=1e-4,
keep_prob=0.5,
output_op_name='output'):
super(CONV, self).__init__()
self.input_size = input_size
self.mid_units = mid_units
self.output_size = output_size
self.lr = lr
self.output_node_name = output_op_name
mlp_graph = tf.Graph()
with mlp_graph.as_default():
x = tf.placeholder(tf.float32, [None, self.input_size],
name='input')
y_t = tf.placeholder(tf.float32, [None, self.output_size],
name='valid')
x_image = tf.reshape(x, [-1, 28, 28, 1], name='reshape')
conv1 = tf_ext.add_conv2d(x_image,
output_size=32,
h_kernel=5,
w_kernel=5,
activation='relu',
name='conv1')
pool1 = tf_ext.add_pool(conv1, name='pool1')
conv2 = tf_ext.add_conv2d(pool1,
output_size=64,
h_kernel=5,
w_kernel=5,
activation='relu',
name='conv2')
pool2 = tf_ext.add_pool(conv2, name='pool2')
flat = tf_ext.add_flatten(pool2, name='flatten')
fc1 = tf_ext.add_fc(flat,
output_size=1024,
activation='relu',
name='fc1')
drop = tf_ext.add_dropout(fc1,
keep_prob=keep_prob,
name='dropout1')
fc2 = tf_ext.add_fc(drop,
self.output_size,
activation='relu',
name='fc2')
y = tf.nn.softmax(fc2, name=self.output_node_name)
with tf.name_scope('train'):
cross_entropy = -tf.reduce_sum(y_t * tf.log(y))
train_op = tf.train.AdamOptimizer(
learning_rate=self.lr).minimize(cross_entropy)
with tf.name_scope('predict'):
correct_prediction = tf.equal(tf.argmax(y, 1),
tf.argmax(y_t, 1))
predict_op = tf.reduce_mean(
tf.cast(correct_prediction, tf.float32))
# version V2 ファイル新規作成する際にエラーになります。
# max_to_keep 出力ckpt フィアル数
saver = tf.train.Saver(write_version=tf.train.SaverDef.V1,
max_to_keep=1)
pb_saver = tf.train
init = tf.global_variables_initializer()
self.graph = mlp_graph
self.train_op = train_op
self.predict_op = predict_op
self.loss_op = cross_entropy
self.x = x
self.y = y
self.y_t = y_t
self.init = init
self.saver = saver
self.pb_saver = pb_saver
def conv2d_5l_block(x,
filter=None,
format="NHWC",
name="conv_5l",
batch_normalization=True,
layer_list=None):
"""
conv2d_5l_block
### 引数
x: object 入力Tensor
sub_tensor: string 存在する場合tensorを連結する
route_dim: object `0-D `int32`` Tensor`。連結するtensor。
format: string "NHWC"(チャネル先) and "NCHW" for チャネル最後. default は 'NHWC'
name: string 名前
batch_normalization: string Trueまたfalse
layer_list: list 存在する場合block_conv_1から5まで追加する
### 戻り値
y object「tensor」。 変換したtensor
"""
with tf.name_scope(name) as scope:
x = tf.convert_to_tensor(x)
input_shape = x.get_shape()
N, H, W, C = (0, 0, 0, 0)
if format == "NHWC":
N, H, W, C = input_shape[0].value, input_shape[
1].value, input_shape[2].value, input_shape[3].value
elif format == "NCHW":
N, C, H, W = input_shape[0].value, input_shape[
1].value, input_shape[2].value, input_shape[3].value
filter_1 = filter if filter is not None else int(C / 2)
filter_2 = int(filter_1 * 2)
filter_3 = filter_1
filter_4 = int(filter_1 * 2)
filter_5 = filter_1
block_conv_1 = tf_helper.add_conv2d(
x,
filter_1,
h_kernel=1,
w_kernel=1,
name="layer_1",
h_stride=1,
w_stride=1,
format=format,
batch_normalization=batch_normalization,
activation="leaky_relu",
leaky_relu_alpha=0.1,
padding="SAME")
print("conv5l : ", block_conv_1.shape)
block_conv_2 = tf_helper.add_conv2d(
block_conv_1,
filter_2,
h_kernel=3,
w_kernel=3,
name="layer_2",
h_stride=1,
w_stride=1,
format=format,
batch_normalization=batch_normalization,
activation="leaky_relu",
leaky_relu_alpha=0.1,
padding="SAME")
print("conv5l : ", block_conv_2.shape)
block_conv_3 = tf_helper.add_conv2d(
block_conv_2,
filter_3,
h_kernel=1,
w_kernel=1,
name="layer_3",
h_stride=1,
w_stride=1,
format=format,
batch_normalization=batch_normalization,
activation="leaky_relu",
leaky_relu_alpha=0.1,
padding="SAME")
print("conv5l : ", block_conv_3.shape)
block_conv_4 = tf_helper.add_conv2d(
block_conv_3,
filter_4,
h_kernel=3,
w_kernel=3,
name="layer_4",
h_stride=1,
w_stride=1,
format=format,
batch_normalization=batch_normalization,
activation="leaky_relu",
leaky_relu_alpha=0.1,
padding="SAME")
print("conv5l : ", block_conv_4.shape)
block_conv_5 = tf_helper.add_conv2d(
block_conv_4,
filter_5,
h_kernel=1,
w_kernel=1,
name="layer_5",
h_stride=1,
w_stride=1,
format=format,
batch_normalization=batch_normalization,
activation="leaky_relu",
leaky_relu_alpha=0.1,
padding="SAME")
print("conv5l : ", block_conv_5.shape)
y = block_conv_5
if layer_list is not None:
layer_list.append(block_conv_1)
layer_list.append(block_conv_2)
layer_list.append(block_conv_3)
layer_list.append(block_conv_4)
layer_list.append(block_conv_5)
return y