tensorcv.models package

Submodules

tensorcv.models.base module

class tensorcv.models.base.ModelDes

Bases: object

base model for ModelDes

create_graph()

create_model(inputs=None)

ex_init_model(dataflow, trainer)

get_batch_size()

get_global_step

get_graph_feed()

get_prediction_placeholder()

get_train_placeholder()

model_input

set_batch_size(val)

set_dropout(dropout_placeholder, keep_prob=0.5)

set_is_training(is_training=True)

set_model_input(inputs=None)

set_prediction_placeholder(plhs=None)

set_train_placeholder(plhs=None)

setup_summary()

class tensorcv.models.base.BaseModel

Bases: tensorcv.models.base.ModelDes

Model with single loss and single optimizer

default_collection

get_grads()

get_loss()

get_optimizer()

class tensorcv.models.base.GANBaseModel(input_vec_length, learning_rate)

Bases: tensorcv.models.base.ModelDes

Base model for GANs

d_collection

def_loss(dis_loss_fnc, gen_loss_fnc)

updata definintion of loss functions

g_collection

get_discriminator_grads()

get_discriminator_loss()

get_discriminator_optimizer()

get_gen_data()

get_generator_grads()

get_generator_loss()

get_generator_optimizer()

get_graph_feed()

get_random_vec_placeholder()

get_sample_gen_data()

tensorcv.models.layers module

tensorcv.models.layers.batch_flatten(x)

Flatten the tensor except the first dimension.

tensorcv.models.layers.batch_norm(x, train=True, name='bn')

batch normal

Parameters:

• x (tf.tensor) – a tensor

• name (str) – name scope

• train (bool) – whether training or not

Returns:

tf.tensor with name ‘name’

tensorcv.models.layers.conv(x, filter_size, out_dim, name='conv', stride=1, padding='SAME', nl=<function identity>, data_dict=None, init_w=None, init_b=None, use_bias=True, wd=None, trainable=True)

2D convolution

Parameters:

• x (tf.tensor) – a 4D tensor Input number of channels has to be known

• filter_size (int or list with length 2) – size of filter

• out_dim (int) – number of output channels

• name (str) – name scope of the layer

• stride (int or list) – stride of filter

• padding (str) – ‘VALID’ or ‘SAME’

• init_b (init_w,) – initializer for weight and bias variables. Default to ‘random_normal_initializer’

• nl – a function

Returns:

tf.tensor with name ‘output’

tensorcv.models.layers.dconv(x, filter_size, out_dim=None, out_shape=None, out_shape_by_tensor=None, name='dconv', stride=2, padding='SAME', nl=<function identity>, data_dict=None, init_w=None, init_b=None, wd=None, trainable=True)

2D deconvolution

Parameters:

• x (tf.tensor) – a 4D tensor Input number of channels has to be known

• filter_size (int or list with length 2) – size of filter

• out_dim (int) – number of output channels

• out_shape (list(int)) – shape of output without None

• out_shape_by_tensor (tf.tensor) – a tensor has the same shape of output except the out_dim

• name (str) – name scope of the layer

• stride (int or list) – stride of filter

• padding (str) – ‘VALID’ or ‘SAME’

• init – initializer for variables. Default to ‘random_normal_initializer’

• nl – a function

Returns:

tf.tensor with name ‘output’

tensorcv.models.layers.dropout(x, keep_prob, is_training, name='dropout')

Dropout

Parameters:

• x (tf.tensor) – a tensor

• keep_prob (float) – keep prbability of dropout

• is_training (bool) – whether training or not

• name (str) – name scope

Returns:

tf.tensor with name ‘name’

tensorcv.models.layers.fc(x, out_dim, name='fc', nl=<function identity>, init_w=None, init_b=None, data_dict=None, wd=None, trainable=True, re_dict=False)

Fully connected layer

Parameters:

• x (tf.tensor) – a tensor to be flattened The first dimension is the batch dimension

• num_out (int) – dimension of output

• name (str) – name scope of the layer

• init – initializer for variables. Default to ‘random_normal_initializer’

• nl – a function

Returns:

tf.tensor with name ‘output’

tensorcv.models.layers.get_shape2D(in_val)

Return a 2D shape

Parameters:in_val (int or list with length 2) – Returns:list with length 2

tensorcv.models.layers.get_shape4D(in_val)

Return a 4D shape

Parameters:in_val (int or list with length 2) – Returns:list with length 4

tensorcv.models.layers.global_avg_pool(x, name='global_avg_pool', data_format='NHWC')

tensorcv.models.layers.leaky_relu(x, leak=0.2, name='LeakyRelu')

Allow a small non-zero gradient when the unit is not active

Parameters:

• x (tf.tensor) – a tensor

• leak (float) – Default to 0.2

Returns:

tf.tensor with name ‘name’

tensorcv.models.layers.max_pool(x, name='max_pool', filter_size=2, stride=None, padding='VALID')

Max pooling layer

Parameters:

• x (tf.tensor) – a tensor

• name (str) – name scope of the layer

• filter_size (int or list with length 2) – size of filter

• stride (int or list with length 2) – Default to be the same as shape

• padding (str) – ‘VALID’ or ‘SAME’. Use ‘SAME’ for FCN.

Returns:

tf.tensor with name ‘name’

tensorcv.models.layers.new_biases(name, idx, shape, initializer=None, data_dict=None, trainable=True)

tensorcv.models.layers.new_normal_variable(name, shape=None, trainable=True, stddev=0.002)

tensorcv.models.layers.new_variable(name, idx, shape, initializer=None)

tensorcv.models.layers.new_weights(name, idx, shape, initializer=None, wd=None, data_dict=None, trainable=True)

tensorcv.models.losses module

tensorcv.models.losses.GAN_discriminator_loss(d_real, d_fake, name='d_loss')

tensorcv.models.losses.GAN_generator_loss(d_fake, name='g_loss')

tensorcv.models.losses.comp_loss_fake(discrim_output)

tensorcv.models.losses.comp_loss_real(discrim_output)

Module contents