1. tf函數(shù)

tensorflow 封裝的工具類函數(shù)

| 操作組 | 操作 |
|:-------------| -----|
|Maths| Add, Sub, Mul, Div, Exp, Log, Greater, Less, Equal|
|Array| Concat, Slice, Split, Constant, Rank, Shape, Shuffle|
|Matrix| MatMul, MatrixInverse, MatrixDeterminant|
|Neuronal| Network SoftMax, Sigmoid, ReLU, Convolution2D, MaxPool|
|Checkpointing| Save, Restore|
|Queues and syncronizations| Enqueue, Dequeue, MutexAcquire, MutexRelease|
|Flow control| Merge, Switch, Enter, Leave, NextIteration|

2.算術(shù)操作

|操作 |描述|
|:-------|------|
|tf.add(x, y, name=None) |求和|
|tf.sub(x, y, name=None) |減法|
|tf.mul(x, y, name=None) |乘法|
|tf.div(x, y, name=None) |除法|
|tf.mod(x, y, name=None) |取模|
|tf.abs(x, name=None) |求絕對(duì)值|
|tf.neg(x, name=None) |取負(fù) (y = -x).|
|tf.sign(x, name=None) |返回符號(hào) y = sign(x) = -1 if x < 0; 0 if x == 0; 1 if x > 0.|
|tf.inv(x, name=None) |取反|
|tf.square(x, name=None) |計(jì)算平方 (y = x * x = x^2).|
|tf.round(x, name=None) |舍入最接近的整數(shù)# ‘a(chǎn)’ is [0.9, 2.5, 2.3, -4.4]
tf.round(a) ==> [ 1.0, 3.0, 2.0, -4.0 ]|
|tf.sqrt(x, name=None) |開根號(hào) (y = \sqrt{x} = x^{1/2}).|
|tf.pow(x, y, name=None) |冪次方 # tensor ‘x’ is [[2, 2], [3, 3]]
# tensor ‘y’ is [[8, 16], [2, 3]]
tf.pow(x, y) ==> [[256, 65536], [9, 27]]|
|tf.exp(x, name=None) |計(jì)算e的次方|
|tf.log(x, name=None) |計(jì)算log，一個(gè)輸入計(jì)算e的ln，兩輸入以第二輸入為底|
|tf.maximum(x, y, name=None) |返回最大值 (x > y ? x : y)|
|tf.minimum(x, y, name=None) |返回最小值 (x < y ? x : y)|
|tf.cos(x, name=None) |三角函數(shù)cosine|

|tf.sin(x, name=None) |三角函數(shù)sine|
|tf.tan(x, name=None) |三角函數(shù)tan|
|tf.atan(x, name=None) |三角函數(shù)ctan|

3.張量（矩陣）操作

• 數(shù)據(jù)類型轉(zhuǎn)換

|操作 |描述|
|:-------|------|
|tf.string_to_number(string_tensor, out_type=None, name=None) | 字符串轉(zhuǎn)為數(shù)字 |
|tf.to_double(x, name=’ToDouble’) | 轉(zhuǎn)為64位浮點(diǎn)類型–float64 |
|tf.to_float(x, name=’ToFloat’) | 轉(zhuǎn)為32位浮點(diǎn)類型–float32 |
|tf.to_int32(x, name=’ToInt32’) | 轉(zhuǎn)為32位整型–int32 |
|tf.to_int64(x, name=’ToInt64’) | 轉(zhuǎn)為64位整型–int64 |
|tf.cast(x, dtype, name=None) | 將x或者x.values轉(zhuǎn)換為dtype
# tensor a is [1.8, 2.2], dtype=tf.float
tf.cast(a, tf.int32) ==> [1, 2] # dtype=tf.int32 |

• 矩陣的變形操作

|操作 |描述|
|:-------|------|
|tf.shape(input, name=None) | 返回?cái)?shù)據(jù)的shape
# ‘t’ is [[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]]
shape(t) ==> [2, 2, 3] |
|tf.size(input, name=None) | 返回?cái)?shù)據(jù)的元素?cái)?shù)量
# ‘t’ is [[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]]]
size(t) ==> 12 |
|tf.rank(input, name=None) | 返回tensor的rank
注意：此rank不同于矩陣的rank，tensor的rank表示一個(gè)tensor需要的索引數(shù)目來唯一表示任何一個(gè)元素也就是通常所說的 “order”, “degree”或”ndims”
#’t’ is [[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]]
# shape of tensor ‘t’ is [2, 2, 3]
rank(t) ==> 3
|tf.reshape(tensor, shape, name=None) | 改變tensor的形狀
# tensor ‘t’ is [1, 2, 3, 4, 5, 6, 7, 8, 9]
# tensor ‘t’ has shape [9]
reshape(t, [3, 3]) ==>
[[1, 2, 3],
[4, 5, 6],
[7, 8, 9]]
#如果shape有元素[-1]

,表示在該維度打平至一維
# -1 將自動(dòng)推導(dǎo)得為 9:
reshape(t, [2, -1]) ==>
[[1, 1, 1, 2, 2, 2, 3, 3, 3],
[4, 4, 4, 5, 5, 5, 6, 6, 6]] |
|tf.expand_dims(input, dim, name=None) | 插入維度1進(jìn)入一個(gè)tensor中
#該操作要求-1-input.dims()
# ‘t’ is a tensor of shape [2]
shape(expand_dims(t, 0)) ==> [1, 2]
shape(expand_dims(t, 1)) ==> [2, 1]
shape(expand_dims(t, -1)) ==> [2, 1] <= dim <= input.dims() |

• 矩陣切片、合并

|操作|描述|
|:------|-------|
| tf.slice(input_, begin, size, name=None) | 對(duì)tensor進(jìn)行切片操作
其中size[i] = input.dim_size(i) - begin[i]
該操作要求 0 <= begin[i] <= begin[i] + size[i] <=
Di for i in [0, n]
#’input’ is
#[[[1, 1, 1], [2, 2, 2]],[[3, 3, 3], [4, 4, 4]],[[5, 5, 5], [6, 6, 6]]]
tf.slice(input, [1, 0, 0], [1, 1, 3]) ==> [[[3, 3, 3]]]
tf.slice(input, [1, 0, 0], [1, 2, 3]) ==>
[[[3, 3, 3],
[4, 4, 4]]]
tf.slice(input, [1, 0, 0], [2, 1, 3]) ==>
[[[3, 3, 3]],
[[5, 5, 5]]] |
| tf.split(split_dim, num_split, value, name=’split’) | 沿著某一維度將tensor分離為num_split tensors
# ‘value’ is a tensor with shape [5, 30]
# Split ‘value’ into 3 tensors along dimension 1
split0, split1, split2 = tf.split(1, 3, value)
tf.shape(split0) ==> [5, 10] |
| tf.concat(concat_dim, values, name=’concat’) | 沿著某一維度連結(jié)tensor
t1 = [[1, 2, 3], [4, 5, 6]]
t2 = [[7, 8, 9], [10, 11, 12]]
tf.concat(0, [t1, t2]) ==> [[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]]
tf.concat(1, [t1, t2]) ==> [[1, 2, 3, 7, 8, 9], [4, 5, 6, 10, 11, 12]]
如果想沿著tensor一新軸連結(jié)打包,那么可以：
tf.concat(axis, [tf.expand_dims(t, axis) for t in tensors])
等同于tf.pack(tensors, axis=axis) |
| tf.pack(values, axis=0, name=’pack’) | 將一系列rank-R的tensor打包為一個(gè)rank-(R+1)的tensor
# ‘x’ is [1, 4], ‘y’ is [2, 5], ‘z’ is [3, 6]
pack([x, y, z]) => [[1, 4], [2, 5], [3, 6]]
# 沿著第一維pack
pack([x, y, z], axis=1) => [[1, 2, 3], [4, 5, 6]]
等價(jià)于tf.pack([x, y, z]) = np.asarray([x, y, z]) |
| tf.reverse(tensor, dims, name=None) | 沿著某維度進(jìn)行序列反轉(zhuǎn)
其中dim為列表，元素為bool型，size等于rank(tensor)
# tensor ‘t’ is
[[[[ 0, 1, 2, 3],
#[ 4, 5, 6, 7],
<p>
#[ 8, 9, 10, 11]],
#[[12, 13, 14, 15],
#[16, 17, 18, 19],
#[20, 21, 22, 23]]]]
# tensor ‘t’ shape is [1, 2, 3, 4]
# ‘dims’ is [False, False, False, True]
reverse(t, dims) ==>
[[[[ 3, 2, 1, 0],
[ 7, 6, 5, 4],
[ 11, 10, 9, 8]],
[[15, 14, 13, 12],
[19, 18, 17, 16],
[23, 22, 21, 20]]]]|
| tf.transpose(a, perm=None, name=’transpose’) | 調(diào)換tensor的維度順序
按照列表perm的維度排列調(diào)換tensor順序，
如為定義，則perm為(n-1…0)
# ‘x’ is [[1 2 3],[4 5 6]]
# ‘x’ is [[1 2 3],[4 5 6]]
# Equivalently
tf.transpose(x, perm=[1, 0]) ==> [[1 4],[2 5], [3 6]] |
| tf.gather(params, indices, validate_indices=None, name=None) | 合并索引indices所指示params中的切片

|
| tf.one_hot(indices, depth, on_value=None, off_value=None,axis=None, dtype=None, name=None) | indices = [0, 2, -1, 1]
depth = 3
on_value = 5.0
off_value = 0.0
axis = -1
#Then output is [4 x 3]:
output =
[5.0 0.0 0.0] // one_hot(0)
[0.0 0.0 5.0] // one_hot(2)
[0.0 0.0 0.0] // one_hot(-1)
[0.0 5.0 0.0] // one_hot(1) |

• 矩陣運(yùn)算相關(guān)函數(shù)

|操作|描述|
|:------|-------|
| tf.diag(diagonal, name=None) | 返回一個(gè)給定對(duì)角值的對(duì)角tensor
# ‘diagonal’ is [1, 2, 3, 4]
tf.diag(diagonal) ==>
[[1, 0, 0, 0]
[0, 2, 0, 0]
[0, 0, 3, 0]
[0, 0, 0, 4]] |
| tf.diag_part(input, name=None) | tf.diag_part(input, name=None) |
| tf.trace(x, name=None) | 求一個(gè)2維tensor足跡，即對(duì)角值diagonal之和 |
| tf.transpose(a, perm=None, name=’transpose’) | 調(diào)換tensor的維度順序
按照列表perm的維度排列調(diào)換tensor順序，
如為定義，則perm為(n-1…0)
# ‘x’ is [[1 2 3],[4 5 6]]
tf.transpose(x) ==> [[1 4], [2 5],[3 6]]
# Equivalently
tf.transpose(x, perm=[1, 0]) ==> [[1 4],[2 5], [3 6]] |
| tf.matmul(a, b, transpose_a=False, transpose_b=False, a_is_sparse=False, b_is_sparse=False, name=None) | 矩陣相乘 |
| tf.matrix_determinant(input, name=None) | 返回方陣的行列式 |
| tf.matrix_inverse(input, adjoint=None, name=None) | 求方陣的逆矩陣，adjoint為True時(shí)，計(jì)算輸入共軛矩陣的逆矩陣 |
| tf.cholesky(input, name=None) | 對(duì)輸入方陣cholesky分解，
即把一個(gè)對(duì)稱正定的矩陣表示成一個(gè)下三角矩陣L和其轉(zhuǎn)置的乘積的分解A=LL^T |
| tf.matrix_solve(matrix, rhs, adjoint=None, name=None) | 求解tf.matrix_solve(matrix, rhs, adjoint=None, name=None)
matrix為方陣shape為[M,M],rhs的shape為[M,K]，output為[M,K] |

• 復(fù)數(shù)計(jì)算

|操作|描述|

|:------|-------|
| tf.complex(real, imag, name=None) | 將兩實(shí)數(shù)轉(zhuǎn)換為復(fù)數(shù)形式
# tensor ‘real’ is [2.25, 3.25]
# tensor imag is [4.75, 5.75]
tf.complex(real, imag) ==> [[2.25 + 4.75j], [3.25 + 5.75j]] |
| tf.complex_abs(x, name=None) | 計(jì)算復(fù)數(shù)的絕對(duì)值，即長度。
# tensor ‘x’ is [[-2.25 + 4.75j], [-3.25 + 5.75j]]
tf.complex_abs(x) ==> [5.25594902, 6.60492229] |
| tf.conj(input, name=None) | 計(jì)算共軛復(fù)數(shù) |
| tf.imag(input, name=None)
tf.real(input, name=None) | 提取復(fù)數(shù)的虛部和實(shí)部 |
| tf.fft(input, name=None) | 計(jì)算一維的離散傅里葉變換，輸入數(shù)據(jù)類型為complex64 |

• 規(guī)約計(jì)算

|操作|描述|
|:------|-------|
| tf.reduce_sum(input_tensor, reduction_indices=None, keep_dims=False, name=None) | 計(jì)算輸入tensor元素的和，或者安照reduction_indices指定的軸進(jìn)行求和
# ‘x’ is [[1, 1, 1]
# [1, 1, 1]]
tf.reduce_sum(x) ==> 6
tf.reduce_sum(x, 0) ==> [2, 2, 2]
tf.reduce_sum(x, 1) ==> [3, 3]
tf.reduce_sum(x, 1, keep_dims=True) ==> [[3], [3]]
tf.reduce_sum(x, [0, 1]) ==> 6 |
| tf.reduce_prod(input_tensor, reduction_indices=None, keep_dims=False, name=None) | 計(jì)算輸入tensor元素的乘積，或者安照reduction_indices指定的軸進(jìn)行求乘積 |
| tf.reduce_min(input_tensor, reduction_indices=None, keep_dims=False, name=None) | 求tensor中最小值 |

| tf.reduce_max(input_tensor, reduction_indices=None,keep_dims=False, name=None) | 求tensor中最大值 |
| tf.reduce_mean(input_tensor, reduction_indices=None, keep_dims=False, name=None) | 求tensor中平均值 |
| tf.reduce_all(input_tensor, reduction_indices=None, keep_dims=False, name=None) | 對(duì)tensor中各個(gè)元素求邏輯’與’ # ‘x’ is
# [[True, True]
# [False, False]]
tf.reduce_all(x) ==> False
tf.reduce_all(x, 0) ==> [False, False]
tf.reduce_all(x, 1) ==> [True, False] |
| tf.reduce_any(input_tensor, reduction_indices=None, keep_dims=False, name=None) | 對(duì)tensor中各個(gè)元素求邏輯’或’|
| tf.accumulate_n(inputs, shape=None, tensor_dtype=None, name=None) | 計(jì)算一系列tensor的和
# tensor ‘a(chǎn)’ is [[1, 2], [3, 4]]
# tensor b is [[5, 0], [0, 6]]
tf.accumulate_n([a, b, a]) ==> [[7, 4], [6, 14]] |
| tf.cumsum(x, axis=0, exclusive=False, reverse=False, name=None) | 求累積和
tf.cumsum([a, b, c]) ==> [a, a + b, a + b + c]
tf.cumsum([a, b, c], exclusive=True) ==> [0, a, a + b]
tf.cumsum([a, b, c], reverse=True) ==> [a + b + c, b + c, c]
tf.cumsum([a, b, c], exclusive=True, reverse=True) ==> [b + c, c, 0] |

• 分割(Segmentation)

|操作|描述|
|:------|-------|
| tf.segment_sum(data, segment_ids, name=None) | 根據(jù)segment_ids的分段計(jì)算各個(gè)片段的和
其中segment_ids為一個(gè)size與data第一維相同的tensor
其中id為int型數(shù)據(jù)，最大id不大于size
c = tf.constant([[1,2,3,4], [-1,-2,-3,-4], [5,6,7,8]])
tf.segment_sum(c, tf.constant([0, 0, 1]))
==>[[0 0 0 0]
[5 6 7 8]]
上面例子分為[0,1]兩id,對(duì)相同id的data相應(yīng)數(shù)據(jù)進(jìn)行求和,
并放入結(jié)果的相應(yīng)id中，
且segment_ids只升不降 |

| tf.segment_prod(data, segment_ids, name=None) | 根據(jù)segment_ids的分段計(jì)算各個(gè)片段的積 |
| tf.segment_min(data, segment_ids, name=None) | 根據(jù)segment_ids的分段計(jì)算各個(gè)片段的最小值 |
| tf.segment_max(data, segment_ids, name=None) | 根據(jù)segment_ids的分段計(jì)算各個(gè)片段的最大值 |
| tf.segment_mean(data, segment_ids, name=None) | 根據(jù)segment_ids的分段計(jì)算各個(gè)片段的平均值 |
| tf.unsorted_segment_sum(data, segment_ids,num_segments, name=None) | 與tf.segment_sum函數(shù)類似，不同在于segment_ids中id順序可以是無序的 |
| tf.sparse_segment_sum(data, indices, segment_ids, name=None) | 輸入進(jìn)行稀疏分割求和
c = tf.constant([[1,2,3,4], [-1,-2,-3,-4], [5,6,7,8]])
# Select two rows, one segment.
tf.sparse_segment_sum(c, tf.constant([0, 1]), tf.constant([0, 0]))
==> [[0 0 0 0]]
對(duì)原data的indices為[0,1]位置的進(jìn)行分割，
并按照segment_ids的分組進(jìn)行求和 |

• 序列比較與索引提取(Sequence Comparison and Indexing)

|操作|描述|
|:------|-------|
| tf.argmin(input, dimension, name=None) | 返回input最小值的索引index |
| tf.argmax(input, dimension, name=None) | 返回input最大值的索引index |
| tf.listdiff(x, y, name=None) | 返回x，y中不同值的索引 |
| tf.where(input, name=None) | 返回bool型tensor中為True的位置
# ‘input’ tensor is
#[[True, False]
#[True, False]]
# ‘input’ 有兩個(gè)’True’,那么輸出兩個(gè)坐標(biāo)值.
# ‘input’的rank為2, 所以每個(gè)坐標(biāo)為具有兩個(gè)維度.
where(input) ==>
[[0, 0],
[1, 0]] |
| tf.unique(x, name=None) | 返回一個(gè)元組tuple(y,idx)，y為x的列表的唯一化數(shù)據(jù)列表，
idx為x數(shù)據(jù)對(duì)應(yīng)y元素的index
# tensor ‘x’ is [1, 1, 2, 4, 4, 4, 7, 8, 8]
y, idx = unique(x)
y ==> [1, 2, 4, 7, 8]
idx ==> [0, 0, 1, 2, 2, 2, 3, 4, 4] |
| tf.invert_permutation(x, name=None) | 置換x數(shù)據(jù)與索引的關(guān)系
# tensor x is [3, 4, 0, 2, 1]
invert_permutation(x) ==> [2, 4, 3, 0, 1] |

4.神經(jīng)網(wǎng)絡(luò)(Neural Network)

• 激活函數(shù)（Activation Functions）

|操作|描述|
|:------|-------|
| tf.nn.relu(features, name=None) | 整流函數(shù)：max(features, 0) |
| tf.nn.relu6(features, name=None) | 以6為閾值的整流函數(shù)：min(max(features, 0), 6) |
| tf.nn.elu(features, name=None) | elu函數(shù)，exp(features) - 1 if < 0,否則featuresExponential Linear Units (ELUs) |
| tf.nn.softplus(features, name=None) | 計(jì)算softplus：log(exp(features) + 1) |
| tf.nn.dropout(x, keep_prob, noise_shape=None, seed=None, name=None) | 計(jì)算dropout，keep_prob為keep概率
noise_shape為噪聲的shape |
| tf.nn.bias_add(value, bias, data_format=None, name=None) | 對(duì)value加一偏置量
此函數(shù)為tf.add的特殊情況，bias僅為一維，
函數(shù)通過廣播機(jī)制進(jìn)行與value求和,
數(shù)據(jù)格式可以與value不同，返回為與value相同格式 |
| tf.sigmoid(x, name=None) | y = 1 / (1 + exp(-x))
|
| tf.tanh(x, name=None) | tf.tanh(x, name=None) |

• 卷積函數(shù)（Convolution）

|操作|描述|
|:------|-------|
| tf.nn.conv2d(input, filter, strides, padding, use_cudnn_on_gpu=None, data_format=None, name=None) | 在給定的4D input與 filter下計(jì)算2D卷積
輸入shape為 [batch, height, width, in_channels] |
| tf.nn.conv3d(input, filter, strides, padding, name=None) | 在給定的5D input與 filter下計(jì)算3D卷積
輸入shape為[batch, in_depth, in_height, in_width, in_channels]|

• 池化函數(shù)（Pooling）

|操作|描述|
|:------|-------|
| tf.nn.avg_pool(value, ksize, strides, padding, data_format=’NHWC’, name=None) | 平均方式池化 |
| tf.nn.max_pool(value, ksize, strides, padding, data_format=’NHWC’, name=None) | 最大值方法池化 |
| tf.nn.max_pool_with_argmax(input, ksize, strides,padding, Targmax=None, name=None) | 返回一個(gè)二維元組(output,argmax),最大值pooling，返回最大值及其相應(yīng)的索引 |
| tf.nn.avg_pool3d(input, ksize, strides, padding, name=None) | 3D平均值pooling |
| tf.nn.max_pool3d(input, ksize, strides, padding, name=None) | 3D最大值pooling |

• 數(shù)據(jù)標(biāo)準(zhǔn)化（Normalization）

|操作|描述|

|:------|-------|
| tf.nn.l2_normalize(x, dim, epsilon=1e-12, name=None) | 對(duì)維度dim進(jìn)行L2范式標(biāo)準(zhǔn)化
output = x / sqrt(max(sum(x2), epsilon)) |
| tf.nn.sufficient_statistics(x, axes, shift=None, keep_dims=False, name=None) | 計(jì)算與均值和方差有關(guān)的完全統(tǒng)計(jì)量
返回4維元組,元素個(gè)數(shù)，元素總和，元素的平方和，shift結(jié)果
參見算法介紹 |
| tf.nn.normalize_moments(counts, mean_ss, variance_ss, shift, name=None) | 基于完全統(tǒng)計(jì)量計(jì)算均值和方差 |
| tf.nn.moments(x, axes, shift=None, name=None, keep_dims=False) | 直接計(jì)算均值與方差 |

• 損失函數(shù)（Losses）

|操作|描述|

|:------|-------|
| tf.nn.l2_loss(t, name=None) | output = sum(t ** 2) / 2 |

• 分類函數(shù)（Classification）

|操作|描述|
|:------|-------|
| tf.nn.sigmoid_cross_entropy_with_logits(logits, targets, name=None)* | 計(jì)算輸入logits, targets的交叉熵
|
| tf.nn.softmax(logits, name=None) | 計(jì)算softmax
softmax[i, j] = exp(logits[i, j]) / sum_j(exp(logits[i, j])) |
| tf.nn.log_softmax(logits, name=None) | logsoftmax[i, j] = logits[i, j] - log(sum(exp(logits[i])))
|
| tf.nn.softmax_cross_entropy_with_logits(logits, labels, name=None) | 計(jì)算logits和labels的softmax交叉熵
logits, labels必須為相同的shape與數(shù)據(jù)類型 |
| tf.nn.sparse_softmax_cross_entropy_with_logits(logits, labels, name=None) | 計(jì)算logits和labels的softmax交叉熵
|
| tf.nn.weighted_cross_entropy_with_logits(logits, targets, pos_weight, name=None) | 與sigmoid_cross_entropy_with_logits()相似，但給正向樣本損失加了權(quán)重pos_weight |

• 符號(hào)嵌入（Embeddings）

|操作|描述|
|:------|-------|
| tf.nn.embedding_lookup(params, ids, partition_strategy=’mod’, name=None, validate_indices=True) | 根據(jù)索引ids查詢embedding列表params中的tensor值
如果len(params) > 1，id將會(huì)安照partition_strategy策略進(jìn)行分割
1、如果partition_strategy為”mod”，
id所分配到的位置為p = id % len(params)
比如有13個(gè)ids，分為5個(gè)位置，那么分配方案為：
[[0, 5, 10], [1, 6, 11], [2, 7, 12], [3, 8], [4, 9]]
2、如果partition_strategy為”div”,那么分配方案為：
[[0, 1, 2], [3, 4, 5], [6, 7, 8], [9, 10], [11, 12]] |
| tf.nn.embedding_lookup_sparse(params, sp_ids, sp_weights, partition_strategy=’mod’, name=None, combiner=’mean’) | 對(duì)給定的ids和權(quán)重查詢embedding
1、sp_ids為一個(gè)N x M的稀疏tensor，
N為batch大小，M為任意，數(shù)據(jù)類型int64
2、sp_weights的shape與sp_ids的稀疏tensor權(quán)重，
浮點(diǎn)類型，若為None，則權(quán)重為全’1’ |

• 循環(huán)神經(jīng)網(wǎng)絡(luò)（Recurrent Neural Networks）

|操作|描述|
|:------|-------|
| tf.nn.rnn(cell, inputs, initial_state=None, dtype=None, sequence_length=None, scope=None) | 基于RNNCell類的實(shí)例cell建立循環(huán)神經(jīng)網(wǎng)絡(luò) |
| tf.nn.dynamic_rnn(cell, inputs, sequence_length=None, initial_state=None, dtype=None, parallel_iterations=None, swap_memory=False, time_major=False, scope=None) | 基于RNNCell類的實(shí)例cell建立動(dòng)態(tài)循環(huán)神經(jīng)網(wǎng)絡(luò)
與一般rnn不同的是，該函數(shù)會(huì)根據(jù)輸入動(dòng)態(tài)展開
返回(outputs,state) |
| tf.nn.state_saving_rnn(cell, inputs, state_saver, state_name, sequence_length=None, scope=None) | 可儲(chǔ)存調(diào)試狀態(tài)的RNN網(wǎng)絡(luò) |
| tf.nn.bidirectional_rnn(cell_fw, cell_bw, inputs,initial_state_fw=None, initial_state_bw=None, dtype=None,sequence_length=None, scope=None) | 雙向RNN, 返回一個(gè)3元組tuple
(outputs, output_state_fw, output_state_bw) |

-— tf.nn.rnn簡要介紹—
cell: 一個(gè)RNNCell實(shí)例
inputs: 一個(gè)shape為[batch_size, input_size]的tensor
initial_state: 為RNN的state設(shè)定初值，可選
sequence_length：制定輸入的每一個(gè)序列的長度，size為[batch_size],值范圍為[0, T)的int型數(shù)據(jù)
其中T為輸入數(shù)據(jù)序列的長度
@
@針對(duì)輸入batch中序列長度不同，所設(shè)置的動(dòng)態(tài)計(jì)算機(jī)制
@對(duì)于在時(shí)間t，和batch的b行，有
(output, state)(b, t) = ? (zeros(cell.output_size), states(b, sequence_length(b) - 1)) : cell(input(b, t), state(b, t - 1))-

• 求值網(wǎng)絡(luò)（Evaluation）

|操作|描述|
|:------|-------|
| tf.nn.top_k(input, k=1, sorted=True, name=None) | 返回前k大的值及其對(duì)應(yīng)的索引 |

| tf.nn.in_top_k(predictions, targets, k, name=None) | 返回判斷是否targets索引的predictions相應(yīng)的值是否在在predictions前k個(gè)位置中，返回?cái)?shù)據(jù)類型為bool類型，len與predictions同 |

• 監(jiān)督候選采樣網(wǎng)絡(luò)（Candidate Sampling）
  對(duì)于有巨大量的多分類與多標(biāo)簽?zāi)Ｐ?，如果使用全連接softmax將會(huì)占用大量的時(shí)間與空間資源，所以采用候選采樣方法僅使用一小部分類別與標(biāo)簽作為監(jiān)督以加速訓(xùn)練

|操作|描述|
|:------|-------|
| Sampled Loss Functions | |
| tf.nn.nce_loss(weights, biases, inputs, labels, num_sampled,num_classes, num_true=1, sampled_values=None,remove_accidental_hits=False, partition_strategy=’mod’,name=’nce_loss’) | 返回noise-contrastive的訓(xùn)練損失結(jié)果 |
| tf.nn.sampled_softmax_loss(weights, biases, inputs, labels, num_sampled, num_classes, num_true=1, sampled_values=None,remove_accidental_hits=True, partition_strategy=’mod’, name=’sampled_softmax_loss’) | 返回sampled softmax的訓(xùn)練損失 參考- Jean et al., 2014第3部分|
| Candidate Samplers | |
| tf.nn.uniform_candidate_sampler(true_classes, num_true, num_sampled, unique, range_max, seed=None, name=None) | 通過均勻分布的采樣集合
返回三元tuple
1、sampled_candidates 候選集合。
2、期望的true_classes個(gè)數(shù)，為浮點(diǎn)值
3、期望的sampled_candidates個(gè)數(shù)，為浮點(diǎn)值 |
| tf.nn.log_uniform_candidate_sampler(true_classes, num_true,num_sampled, unique, range_max, seed=None, name=None) | 通過log均勻分布的采樣集合，返回三元tuple |
| tf.nn.learned_unigram_candidate_sampler(true_classes, num_true, num_sampled, unique, range_max, seed=None, name=None) | 根據(jù)在訓(xùn)練過程中學(xué)習(xí)到的分布狀況進(jìn)行采樣
返回三元tuple |
| tf.nn.fixed_unigram_candidate_sampler(true_classes, num_true,num_sampled, unique, range_max, vocab_file=”, distortion=1.0, num_reserved_ids=0, num_shards=1, shard=0, unigrams=(), seed=None, name=None) | 基于所提供的基本分布進(jìn)行采樣 |

• 保存與恢復(fù)變量

|操作|描述|
|:------|-------|
| 類tf.train.Saver(Saving and Restoring Variables) | |
| tf.train.Saver.init(var_list=None, reshape=False, sharded=False, max_to_keep=5, keep_checkpoint_every_n_hours=10000.0, name=None, restore_sequentially=False,saver_def=None, builder=None) | 創(chuàng)建一個(gè)存儲(chǔ)器Savervar_list定義需要存儲(chǔ)和恢復(fù)的變量 |
| tf.train.Saver.save(sess, save_path, global_step=None, latest_filename=None, meta_graph_suffix=’meta’,write_meta_graph=True) | 保存變量 |
| tf.train.Saver.restore(sess, save_path) | tf.train.Saver.restore(sess, save_path) |
| tf.train.Saver.last_checkpoints | 列出最近未刪除的checkpoint 文件名 |
| tf.train.Saver.set_last_checkpoints(last_checkpoints) | 設(shè)置checkpoint文件名列表 |
| tf.train.Saver.set_last_checkpoints_with_time(last_checkpoints_with_time) | 設(shè)置checkpoint文件名列表和時(shí)間戳 |

引用：

本文絕大部分內(nèi)容參考自：http://blog.csdn.net/lenbow/article/details/52152766
官方API文檔：https://www.tensorflow.org/versions/r0.11/api_docs/python/index.html