Python sugartensor.sg_context方法代码示例

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def discriminator(tensor):
    # reuse flag
    reuse = len([t for t in tf.global_variables() if t.name.startswith('discriminator')]) > 0
    with tf.sg_context(name='discriminator', size=4, stride=2, act='leaky_relu', bn=True, reuse=reuse):
        res = (tensor
               .sg_conv(dim=64, name='conv1')
               .sg_conv(dim=128, name='conv2')
               .sg_flatten()
               .sg_dense(dim=1024, name='fc1')
               .sg_dense(dim=1, act='linear', bn=False, name='fc2')
               .sg_squeeze())
        return res


#
# inputs
#

# MNIST input tensor ( with QueueRunner ) 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def generator(tensor):
    # reuse flag
    reuse = len([t for t in tf.global_variables() if t.name.startswith('generator')]) > 0
    with tf.sg_context(name='generator', size=4, stride=2, act='leaky_relu', bn=True, reuse=reuse):

        # generator network
        res = (tensor
               .sg_dense(dim=1024, name='fc1')
               .sg_dense(dim=7*7*128, name='fc2')
               .sg_reshape(shape=(-1, 7, 7, 128))
               .sg_upconv(dim=64, name='conv1')
               .sg_upconv(dim=1, act='sigmoid', bn=False, name='conv2'))

        return res


# random normal seed 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def generator(tensor):

    # reuse flag
    reuse = len([t for t in tf.global_variables() if t.name.startswith('generator')]) > 0

    with tf.sg_context(name='generator', size=4, stride=2, act='leaky_relu', bn=True, reuse=reuse):
        res = (tensor
               .sg_dense(dim=1024, name='fc1')
               .sg_dense(dim=7*7*128, name='fc2')
               .sg_reshape(shape=(-1, 7, 7, 128))
               .sg_upconv(dim=64, name='conv1')
               .sg_upconv(dim=1, act='sigmoid', bn=False, name='conv2'))
    return res


#
# inputs
#

# target_number 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def get_loss(opt):

    # conv layers
    with tf.sg_context(name='convs', act='relu', bn=True):
        conv = (opt.input[opt.gpu_index]
                .sg_conv(dim=16, name='conv1')
                .sg_pool()
                .sg_conv(dim=32, name='conv2')
                .sg_pool()
                .sg_conv(dim=32, name='conv3')
                .sg_pool())

    # fc layers
    with tf.sg_context(name='fcs', act='relu', bn=True):
        logit = (conv
                 .sg_flatten()
                 .sg_dense(dim=256, name='fc1')
                 .sg_dense(dim=10, act='linear', bn=False, name='fc2'))

        # cross entropy loss with logit
        return logit.sg_ce(target=opt.target[opt.gpu_index])

# parallel training ( same as single GPU training ) 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def sg_res_block(tensor, opt):

    # default rate
    opt += tf.sg_opt(size=3, rate=1, causal=False, is_first=False)

    # input dimension
    in_dim = tensor.get_shape().as_list()[-1]

    with tf.sg_context(name='block_%d_%d' % (opt.block, opt.rate)):

        # reduce dimension
        input_ = (tensor
                  .sg_bypass(act='relu', ln=(not opt.is_first), name='bypass')  # do not
                  .sg_conv1d(size=1, dim=in_dim/2, act='relu', ln=True, name='conv_in'))

        # 1xk conv dilated
        out = (input_
               .sg_aconv1d(size=opt.size, rate=opt.rate, causal=opt.causal, act='relu', ln=True, name='aconv'))

        # dimension recover and residual connection
        out = out.sg_conv1d(size=1, dim=in_dim, name='conv_out') + tensor

    return out

# inject residual multiplicative block 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def encode(x):

    with tf.sg_context(name='encoder'):
        res = x
        # loop dilated conv block
        for i in range(num_blocks):
            res = (res
                   .sg_res_block(size=5, block=i, rate=1, is_first=True)
                   .sg_res_block(size=5, block=i, rate=2)
                   .sg_res_block(size=5, block=i, rate=4)
                   .sg_res_block(size=5, block=i, rate=8)
                   .sg_res_block(size=5, block=i, rate=16))

    return res


#
# decode graph ( causal convolution )
# 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def decode(x, voca_size):

    with tf.sg_context(name='decoder'):
        res = x
        # loop dilated causal conv block
        for i in range(num_blocks):
            res = (res
                   .sg_res_block(size=3, block=i, rate=1, causal=True, is_first=True)
                   .sg_res_block(size=3, block=i, rate=2, causal=True)
                   .sg_res_block(size=3, block=i, rate=4, causal=True)
                   .sg_res_block(size=3, block=i, rate=8, causal=True)
                   .sg_res_block(size=3, block=i, rate=16, causal=True))

        # final fully convolution layer for softmax
        res = res.sg_conv1d(size=1, dim=voca_size, name='conv_final')

    return res 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def generator(tensor):

    # reuse flag
    reuse = len([t for t in tf.global_variables() if t.name.startswith('generator')]) > 0

    with tf.sg_context(name='generator', size=4, stride=2, act='relu', bn=True, reuse=reuse):
        res = (tensor
               .sg_dense(dim=1024, name='fc1')
               .sg_dense(dim=7*7*128, name='fc2')
               .sg_reshape(shape=(-1, 7, 7, 128))
               .sg_upconv(dim=64, name='conv1')
               .sg_upconv(dim=1, act='sigmoid', bn=False, name='conv2'))
    return res

#
# inputs
#

# target_number 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def generator(x):

    reuse = len([t for t in tf.global_variables() if t.name.startswith('generator')]) > 0
    with tf.sg_context(name='generator', size=4, stride=2, act='leaky_relu', bn=True, reuse=reuse):

        # generator network
        res = (x.sg_dense(dim=1024, name='fc_1')
               .sg_dense(dim=7*7*128, name='fc_2')
               .sg_reshape(shape=(-1, 7, 7, 128))
               .sg_upconv(dim=64, name='conv_1')
               .sg_upconv(dim=1, act='sigmoid', bn=False, name='conv_2'))
    return res


#
# create discriminator
# 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def generator(tensor):
    # reuse flag
    reuse = len([t for t in tf.global_variables() if t.name.startswith('generator')]) > 0
    with tf.sg_context(name='generator', size=4, stride=2, act='leaky_relu', bn=True, reuse=reuse):

        # generator network
        res = (tensor
               .sg_dense(dim=1024, name='fc1')
               .sg_dense(dim=7*7*128, name='fc2')
               .sg_reshape(shape=(-1, 7, 7, 128))
               .sg_upconv(dim=64, name='conv1')
               .sg_upconv(dim=1, act='sigmoid', bn=False, name='conv2'))

        return res 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def generator(tensor):

    # reuse flag
    reuse = len([t for t in tf.global_variables() if t.name.startswith('generator')]) > 0

    with tf.sg_context(name='generator', size=4, stride=2, act='leaky_relu', bn=True, reuse=reuse):
        res = (tensor
               .sg_dense(dim=1024, name='fc1')
               .sg_dense(dim=7*7*128, name='fc2')
               .sg_reshape(shape=(-1, 7, 7, 128))
               .sg_upconv(dim=64, name='conv1')
               .sg_upconv(dim=1, act='sigmoid', bn=False, name='conv2'))
    return res 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def _resnet_graph(x, opt, nums):

    # default option
    opt += tf.sg_opt(num_class=1000, conv_only=False, squeeze=True, act='relu')

    # convolution layers ( residual net v2 arch )
    with tf.sg_context(name=opt.name):
        conv = (x
                .sg_conv(dim=64, size=7, stride=2, bias=False, reuse=opt.reuse, name='conv1')
                .sg_pool(size=3, stride=2, pad='SAME')
                .sg_resnet_layer(dim=64, num=nums[0], stride=1, act=opt.act, reuse=opt.reuse, name='conv2')
                .sg_resnet_layer(dim=128, num=nums[1], stride=2, act=opt.act, reuse=opt.reuse, name='conv3')
                .sg_resnet_layer(dim=256, num=nums[2], stride=2, act=opt.act, reuse=opt.reuse, name='conv4')
                .sg_resnet_layer(dim=512, num=nums[3], stride=2, act=opt.act, reuse=opt.reuse, name='conv5')
                .sg_bypass(act=opt.act, bn=True, reuse=opt.reuse, name='final_act')  # final activation
                .sg_pool(size=7, stride=1, avg=True))  # global average pool

        # fully convolution layers
        fc = conv.sg_conv(dim=opt.num_class, size=1, act='linear', bn=False, reuse=opt.reuse, name='fc')

    # return selectively
    if opt.conv_only:
        return conv
    else:
        if opt.squeeze:
            return fc.sg_squeeze(axis=(1, 2))
        else:
            return fc


#
# Dense net utility
# 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def sg_parallel(func):
    r"""Decorates function as multiple gpu support towers.
    Args:
        func: function to decorate
    """
    @wraps(func)
    def wrapper(**kwargs):
        r"""Manages arguments of `tf.sg_opt`.

        Args:
          kwargs: keyword arguments. The wrapped function will be provided with gpu_index argument.
        """
        # parse option
        opt = tf.sg_opt(kwargs)

        # loop for all available GPUs
        res = []
        for i in range(sg_gpus()):
            # specify device
            with tf.device('/gpu:%d' % i):
                # give new scope only to operation
                with tf.name_scope('gpu_%d' % i):
                    # save reuse flag
                    with sg_context(reuse=(True if i > 0 else False)):
                        # call function
                        res.append(func(opt * tf.sg_opt(gpu_index=i)))

        return res

    return wrapper


#
# Command line argument util funcs
#

# noinspection PyProtectedMember 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def generator(tensor):

    # reuse flag
    reuse = len([t for t in tf.global_variables() if t.name.startswith('generator')]) > 0

    with tf.sg_context(name='generator', size=4, stride=2, act='relu', bn=True, reuse=reuse):
        res = (tensor
               .sg_dense(dim=1024, name='fc1')
               .sg_dense(dim=7*7*128, name='fc2')
               .sg_reshape(shape=(-1, 7, 7, 128))
               .sg_upconv(dim=64, name='conv1')
               .sg_upconv(dim=1, act='sigmoid', bn=False, name='conv2'))
    return res 

# 需要导入模块: import sugartensor [as 别名]
# 或者: from sugartensor import sg_context [as 别名]
def discriminator(tensor):

    # reuse flag
    reuse = len([t for t in tf.global_variables() if t.name.startswith('discriminator')]) > 0

    with tf.sg_context(name='discriminator', size=4, stride=2, act='leaky_relu', reuse=reuse):
        # shared part
        shared = (tensor
                  .sg_conv(dim=64, name='conv1')
                  .sg_conv(dim=128, name='conv2')
                  .sg_flatten()
                  .sg_dense(dim=1024, name='fc1'))

        # discriminator end
        disc = shared.sg_dense(dim=1, act='linear', name='disc').sg_squeeze()

        # shared recognizer part
        recog_shared = shared.sg_dense(dim=128, name='recog')

        # categorical auxiliary classifier end
        cat = recog_shared.sg_dense(dim=cat_dim, act='linear', name='cat')

        # continuous auxiliary classifier end
        con = recog_shared.sg_dense(dim=con_dim, act='sigmoid', name='con')

        return disc, cat, con


#
# inputs
#

# MNIST input tensor ( with QueueRunner )