def __init__(self, model_conf: ModelConfig, mode: RunMode, outputs):
self.model_conf = model_conf
self.utils = NetworkUtils(mode)
self.max_label_num = self.model_conf.max_label_num
if self.max_label_num == -1:
exception(
text=
"The scene must set the maximum number of label (MaxLabelNum)",
code=-998)
self.category_num = self.model_conf.category_num
flatten = tf.keras.layers.Flatten()(outputs)
shape_list = flatten.get_shape().as_list()
# print(shape_list[1], self.max_label_num)
outputs = tf.keras.layers.Reshape(
[self.max_label_num,
int(shape_list[1] / self.max_label_num)])(flatten)
self.outputs = tf.keras.layers.Dense(
input_shape=outputs.shape,
units=self.category_num,
)(inputs=outputs)
print("output to reshape ----------- ", self.outputs.shape)
self.outputs = tf.keras.layers.Reshape(
[self.max_label_num, self.category_num])(self.outputs)
def __init__(self, model_conf: ModelConfig, mode: RunMode,
backbone: CNNNetwork, recurrent: RecurrentNetwork):
"""
:param model_conf: 模型配置
:param mode: 运行模式 (Trains/Validation/Predict)
:param backbone:
:param recurrent:
"""
self.model_conf = model_conf
self.decoder = Decoder(self.model_conf)
self.mode = mode
self.network = backbone
self.recurrent = recurrent
self.inputs = tf.keras.Input(dtype=tf.float32,
shape=self.input_shape,
name='input')
self.labels = tf.keras.Input(dtype=tf.int32,
shape=[None],
sparse=True,
name='labels')
self.utils = NetworkUtils(mode)
self.merged_summary = None
self.optimizer = None
self.dataset_size = None
def __init__(self, model_conf: ModelConfig, mode: RunMode, cnn: CNNNetwork,
recurrent: RecurrentNetwork):
"""
:param model_conf: 模型配置
:param mode: 运行模式 (Trains/Validation/Predict), 其实[Validation]是没有真正被使用的，验证集使用的是[Trains]模式
对于静态图可以定义一个占位符[is_training]通过Feed值来实现这种控制, 但是预测
:param cnn:
:param recurrent:
"""
self.model_conf = model_conf
self.decoder = Decoder(self.model_conf)
self.mode = mode
self.network = cnn
self.recurrent = recurrent
self.inputs = tf.keras.Input(dtype=tf.float32,
shape=self.input_shape,
name='input')
self.labels = tf.keras.Input(dtype=tf.int32,
shape=[None],
sparse=True,
name='labels')
self.utils = NetworkUtils(mode)
self.merged_summary = None
self.optimizer = None
def __init__(self, mode, cnn: CNNNetwork, recurrent: RecurrentNetwork):
self.mode = mode
self.utils = NetworkUtils(mode)
self.network = cnn
self.recurrent = recurrent
self.inputs = tf.placeholder(tf.float32, [None, None, RESIZE[1], IMAGE_CHANNEL], name='input')
self.labels = tf.sparse_placeholder(tf.int32, name='labels')
self.seq_len = None
self.merged_summary = None
def __init__(self, model_conf: ModelConfig, mode: RunMode, cnn: CNNNetwork, recurrent: RecurrentNetwork):
self.model_conf = model_conf
self.mode = mode
self.decoder = Decoder(self.model_conf)
self.utils = NetworkUtils(mode)
self.network = cnn
self.recurrent = recurrent
self.inputs = tf.keras.Input(dtype=tf.float32, shape=self.input_shape, name='input')
self.labels = tf.keras.Input(dtype=tf.int32, shape=[None], sparse=True, name='labels')
self.merged_summary = None
def __init__(self, model_conf: ModelConfig, mode: RunMode, outputs):
self.model_conf = model_conf
self.utils = NetworkUtils(mode)
self.dense = tf.keras.layers.Dense(
units=self.model_conf.category_num + 2,
kernel_initializer=tf.keras.initializers.he_normal(seed=None),
bias_initializer='zeros',
)
self.outputs = self.dense(outputs)
self.predict = tf.keras.backend.permute_dimensions(self.outputs, pattern=(1, 0, 2))
def __init__(self, model_conf: ModelConfig, mode: RunMode, outputs):
self.model_conf = model_conf
self.utils = NetworkUtils(mode)
self.dense = tf.keras.layers.Dense(
units=self.model_conf.category_num + 2,
kernel_initializer=tf.keras.initializers.he_normal(seed=None),
kernel_regularizer=l1_l2(l1=0.01, l2=0.001),
bias_initializer='zeros',
)
self.time_distributed = lambda: tf.keras.layers.TimeDistributed(
layer=self.dense,
name='predict',
)(inputs=outputs, training=self.utils.training)
self.outputs = self.time_distributed()
self.predict = tf.keras.backend.permute_dimensions(self.outputs,
pattern=(1, 0, 2))