Article From:https://www.cnblogs.com/guoyaohua/p/9059605.html

This article introduces the common functions of tensorflow, which is derived from online collation.

TensorFlow Converts graphical definitions into distributed execution operations to make full use of available computing resources, such as CPU or GPU. In general, you do not need to explicitly specify the use of CPU or GPU, and TensorFlow can automatically detect. If GPU is detected, TensOrFlow will make the best use of the first GPU found to perform the operation.Parallel computing can speed up computations of expensive algorithms, and TensorFlow has also made effective improvements to complex operations. Most of the nuclear related operations are device related implementations, such as GPU.

　　Here are some important operations / cores:

 Operating group operation 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

## Arithmetic operation of one, TensorFlow

 description tf.add(x, y, name=None) Summing up tf.sub(x, y, name=None) subtraction tf.mul(x, y, name=None) multiplication tf.div(x, y, name=None) division tf.mod(x, y, name=None) Die tf.abs(x, name=None) Absolute value tf.neg(x, name=None) Negation (y = -x) tf.sign(x, name=None) Return symbol y = sign(x) = -1 if x < 0; 0 if x == 0; 1 if x > 0. tf.inv(x, name=None) Reverse tf.square(x, name=None) Square calculation(y = x * x = x^2) tf.round(x, name=None) Rounding the closest integers# ‘a’ 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) Open root (y = \sqrt{x} = x^{1/2}). tf.pow(x, y, name=None) power（Yuan Suji)# 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) Calculation of the sub square of e tf.log(x, name=None) Calculate log, an input to calculate e’s LN, two input to second input as the bottom. tf.maximum(x, y, name=None) Return to maximum(x > y ? x : y) tf.minimum(x, y, name=None) Return the minimum (x < y ? x : y) tf.cos(x, name=None) Trigonometric function cosine tf.sin(x, name=None) Trigonometric function sine tf.tan(x, name=None) Trigonometric function Tan tf.atan(x, name=None) Trigonometric function ctan

## Two. Tensor operation Tensor Transformations

### 2.1  Data type conversion Casting

 operation describe tf.string_to_number(string_tensor, out_type=None, name=None) String to numeric tf.to_double(x, name=’ToDouble’) Turn to 64 – bit floating-point type – float64 tf.to_float(x, name=’ToFloat’) Turn to 32 – bit floating-point type – float32 tf.to_int32(x, name=’ToInt32’) Turn to 32 – bit integer – int32 tf.to_int64(x, name=’ToInt64’) Turn to 64 – bit integer – Int64 tf.cast(x, dtype, name=None) Convert X or x.values to dtype# tensor a is [1.8, 2.2], dtype=tf.floattf.cast(a, tf.int32) ==> [1, 2] # dtype=tf.int32

### 2.2  Shape operation Shapes and Shaping

 operation description tf.shape(input, name=None) Return the shape of the data# ‘t’ is [[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]]shape(t) ==> [2, 2, 3] tf.size(input, name=None) The number of elements returned to the data# ‘t’ is [[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]]]size(t) ==> 12 tf.rank(input, name=None) Return to the rank of tensor（Dimension)Note: this rank is different from the rank of the matrix,tensorThe rank represents the number of indexes required by a tensor to uniquely represent any element.Also known as “order”, “degree” or “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) Change the shape of tensor# 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]]#If shape has an element [-1], it means that the dimension is flat to one dimension.# -1 It will be automatically derived as 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) Insert dimension 1 into a tensor#This operation requires -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()

### 2.3  Slicing and merging (Slicing and Joining)

 operation describe tf.slice(input_, begin, size, name=None) Slice the tensor and extract part of the content from input.inputs：It can be list, array, tensor     begin：nDimension list, begin[i] represents the relative initial offset of 0 when extracting data from inputs I dimension, that is, extracting data from begin[i] of I dimension.     size：nDimension list, size[i] represents the number of I dimensional elements to be extracted.     There are several relationships as follows:         （1） i in [0,n]         （2）tf.shape(inputs)[0]=len(begin)=len(size)         （3）begin[i]>=0   The starting position of the extraction of the I dimension element is greater than 0         （4）begin[i]+size[i]<=tf.shape(inputs)[i]#’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’) Tensor is separated into num_split tensors by a certain dimension.# ‘value’ is a tensor with shape [5, 30]# Split ‘value’ into 3 tensors along dimension 1split0, split1, split2 = tf.split(1, 3, value)tf.shape(split0) ==> [5, 10] tf.concat(concat_dim, values, name=’concat’) Link tensor along a certain dimension（The whole dimension is constantt1 = [[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]]If you want to link up with a new axis of tensor, you can:tf.concat(axis, [tf.expand_dims(t, axis) for t in tensors])Equivalent to tf.pack (tensors, axis=axis) tf.pack(values, axis=0, name=’pack’) Pack a series of rank-R tensor into a rank- (R+1) tensor.（Integral dimension plus one)# ‘x’ is [1, 4], ‘y’ is [2, 5], ‘z’ is [3, 6]pack([x, y, z]) => [[1, 4], [2, 5], [3, 6]] # Along the first dimension packpack([x, y, z], axis=1) => [[1, 2, 3], [4, 5, 6]]It is equivalent to tf.pack ([x, y, z]) = np.asarray ([x, y, z]). tf.reverse(tensor, dims, name=None) Proceed along a certain dimensionSequence reversalWhere dim is a list, the element is bool, and size is rank (tensor).# tensor ‘t’ is [[[[ 0, 1, 2, 3],#[ 4, 5, 6, 7],#[ 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’) Changing the dimensional order of tensor（Axis transformation)Change the tensor order according to the dimension of listing perm.As defined, perm is (n-1… 0)# ‘x’ is [[1 2 3],[4 5 6]]tf.transpose(x) ==> [[1 4], [2 5],[3 6]]# Equivalentlytf.transpose(x, perm=[1, 0]) ==> [[1 4],[2 5], [3 6]] tf.gather(params, indices, validate_indices=None, name=None) The section in params that is indicated by the merge index indices tf.one_hot(indices, depth, on_value=None, off_value=None, axis=None, dtype=None, name=None) Single heat encoding (ont-hot encoing)indices = [0, 2, -1, 1]depth = 3on_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)

## Three, matrix correlation operation

 operation describe tf.diag(diagonal, name=None) A diagonal tensor that returns a given diagonal value# ‘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) The function is opposite to the above tf.trace(x, name=None) Find the trace of a 2 dimensional tensor, that is, the sum of the diagonal value diagonal. tf.transpose(a, perm=None, name=’transpose’) Changing the dimensional order of tensor (axis transformation) tf.matmul(a, b, transpose_a=False, transpose_b=False, a_is_sparse=False, b_is_sparse=False, name=None) Multiplication of matrices (which can handle batch data) tf.matrix_determinant(input, name=None) Determinant returning the square matrix tf.matrix_inverse(input, adjoint=None, name=None) The inverse matrix of the input conjugate matrix is calculated when the inverse matrix of the square matrix is adjoint True. tf.cholesky(input, name=None) Cholesky decomposition of the input matrix,That is, the decomposition of a symmetric positive definite matrix into a product of a lower triangular matrix L and its transposed A=LL^T. tf.matrix_solve(matrix, rhs, adjoint=None, name=None) Solving equationmatrixFor square matrix shape [M, M], RHS shape is [M, K], output is [M, K]

## Four, plural operation

 operation describe tf.complex(real, imag, name=None) Convert two real numbers into plural form# 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) The absolute value of the complex number, that is, the length# tensor ‘x’ is [[-2.25 + 4.75j], [-3.25 + 5.75j]]tf.complex_abs(x) ==> [5.25594902, 6.60492229] tf.conj(input, name=None) Conjugate complex number tf.imag(input, name=None)tf.real(input, name=None) Extracting the imaginary part and the real part of the complex number tf.fft(input, name=None) One dimensional discrete Fourier transform is calculated, and the input data type is complex64.

## Five. Reduction calculation (Reduction)

 operation describe tf.reduce_sum(input_tensor, reduction_indices=None, keep_dims=False, name=None) Calculate the sum of the input tensor elements or seek the sum of the reduction_indices specified axis.# ‘x’ is [[1, 1, 1]# [1, 1, 1]]tf.reduce_sum(x) ==> 6tf.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) Calculate the product of the input tensor element, or calculate the product according to the axis specified by reduction_indices. tf.reduce_min(input_tensor, reduction_indices=None, keep_dims=False, name=None) Finding the minimum value in tensor tf.reduce_max(input_tensor, reduction_indices=None, keep_dims=False, name=None) The maximum value in tensor tf.reduce_mean(input_tensor, reduction_indices=None, keep_dims=False, name=None) Mean value in tensor tf.reduce_all(input_tensor, reduction_indices=None, keep_dims=False, name=None) Seeking for each element in tensorLogic ‘and’# ‘x’ is # [[True, True]# [False, False]]tf.reduce_all(x) ==> Falsetf.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) Seeking for each element in tensorLogic ‘or’ tf.accumulate_n(inputs, shape=None, tensor_dtype=None, name=None) Calculating the sum of a series of tensor# tensor ‘a’ 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) Accumulation and accumulationtf.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]

## Six, division (Segmentation)

 operation describe tf.segment_sum(data, segment_ids, name=None) The sum of the fragments calculated along the tensor (first dimension)Function parameter　　　　data：A Tensor.　　　　segment_ids：A Tensor; must be one of the following types: int32, Int64;　　　　The rank of one dimension tensor is equal to the rank of the first dimension of data; the value should be sorted and can be repeated.　　　　name：The name of the operation (optional).Function return value　　　　tf.segment_sumThe function returns a Tensor, which has the same type as data.　It has the same shape as data, except for dimension 0 of K (number of segments). tf.segment_prod(data, segment_ids, name=None) Calculating the product of each fragment according to the subsection of segment_ids tf.segment_min(data, segment_ids, name=None) Calculate the minimum value of each segment according to segment_ids segmentation. tf.segment_max(data, segment_ids, name=None) Calculate the maximum value of each segment according to segment_ids segmentation. tf.segment_mean(data, segment_ids, name=None) Calculate the average value of each segment according to segment_ids segmentation. tf.unsorted_segment_sum(data, segment_ids,num_segments, name=None) Similar to the tf.segment_sum function,The difference is that the order of ID in segment_ids can be disordered. tf.sparse_segment_sum(data, indices, segment_ids, name=None) Sum of input for sparse partitionc = 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]]The original data’s indices is divided into [0,1] positions.And in accordance with the grouping of segment_ids

## Seven, sequence comparison and index extraction (Sequence Comparison and Indexing).

 operation describe tf.argmin(input, dimension, name=None) Index index that returns the input minimum tf.argmax(input, dimension, name=None) Index index that returns the input maximum tf.listdiff(x, y, name=None) Returns the index of different values in X, y tf.where(input, name=None) Return to the location of the bool type tensor for True# ‘input’ tensor is #[[True, False]#[True, False]]# ‘input’ There are two ‘True’, then the output of two coordinate values.# ‘input’The rank is 2, so each coordinate has two dimensions.where(input) ==>[[0, 0],[1, 0]] tf.unique(x, name=None) Returns a tuple tuple (y, IDX) and Y is a list of unique data for X.idxThe index for the X data corresponding to the Y element# 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) The relationship between replacement of X data and indexy[x[i]] = i for i in [0, 1, …, len(x) – 1]# tensor x is [3, 4, 0, 2, 1]invert_permutation(x) ==> [2, 4, 3, 0, 1]

## Eight, neural network (Neural Network)

• ### Activation function (Activation Functions)

 operation describe tf.nn.relu(features, name=None) Rectifying function: Max (features, 0) tf.nn.relu6(features, name=None) A rectification function with a threshold of 6: min (max (features, 0), 6). tf.nn.elu(features, name=None) eluFunction, exp (features) – 1 if < 0, otherwise featuresExponential Linear Units (ELUs) tf.nn.softplus(features, name=None) Calculation of softplus:log (exp (features) + 1) tf.nn.dropout(x, keep_prob, noise_shape=None, seed=None, name=None) Calculate dropout, keep_prob is the probability of keepnoise_shapeShape for noise tf.nn.bias_add(value, bias, data_format=None, name=None) Add a bias to valueThis function is a special case of tf.add, bias is only one dimension.The function is summed up with the value by the broadcast mechanism.The data format can be different from value and return to the same format as value. tf.sigmoid(x, name=None) y = 1 / (1 + exp(-x)) tf.tanh(x, name=None) Hyperbolic tangent activation function
• ## Convolution function (Convolution)

 operation describe tf.nn.conv2d(input, filter, strides, padding, use_cudnn_on_gpu=None, data_format=None, name=None) Calculation of 2D convolution with a given 4D input and filterThe input shape is [batch, height, width, in_channels] tf.nn.conv3d(input, filter, strides, padding, name=None) Calculation of 3D convolution with a given 5D input and filterInput shape is [batch, in_depth, in_height, in_width, in_channels]
• ## Pooling function (Pooling)

 operation describe tf.nn.avg_pool(value, ksize, strides, padding, data_format=’NHWC’, name=None) Averaging pooling tf.nn.max_pool(value, ksize, strides, padding, data_format=’NHWC’, name=None) Maxima method pooling tf.nn.max_pool_with_argmax(input, ksize, strides,padding, Targmax=None, name=None) Return a two-dimensional tuple(output,argmax),Maximum value pooling,And return the maximum value and its corresponding index tf.nn.avg_pool3d(input, ksize, strides, padding, name=None) 3DMean value pooling tf.nn.max_pool3d(input, ksize, strides, padding, name=None) 3DMaximum value pooling
• ## Data standardization (Normalization)

 operation describe tf.nn.l2_normalize(x, dim, epsilon=1e-12, name=None) Standardization of the L2 paradigm for dimension dimoutput = x / sqrt(max(sum(x**2), epsilon)) tf.nn.sufficient_statistics(x, axes, shift=None, keep_dims=False, name=None) Calculation of complete statistics related to mean and varianceReturns the 4 dimensional tuple, the number of * elements, the sum of * elements, the sum of squares of * elements, and *shift results.See the algorithm Introduction tf.nn.normalize_moments(counts, mean_ss, variance_ss, shift, name=None) Calculating mean and variance based on complete statistics tf.nn.moments(x, axes, shift=None, name=None, keep_dims=False) Direct calculation of mean and variance
• ## Loss function (Losses)

 operation describe tf.nn.l2_loss(t, name=None) output = sum(t ** 2) / 2
• ## Classification function (Classification)

 operation describe tf.nn.sigmoid_cross_entropy_with_logits(logits, targets, name=None)* Calculating the cross entropy of input Logits, targets tf.nn.softmax(logits, name=None) Computing softmaxsoftmax[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) Calculating the softmax cross entropy of Logits and labelslogits, labelsThe same shape and data types must be the same tf.nn.sparse_softmax_cross_entropy_with_logits(logits, labels, name=None) Calculating the softmax cross entropy of Logits and labels tf.nn.weighted_cross_entropy_with_logits(logits, targets, pos_weight, name=None) Similar to sigmoid_cross_entropy_with_logits (),But the weight of the positive sample is added to the weight pos_weight
• ## Symbol embedding (Embeddings)

 operation describe tf.nn.embedding_lookup(params, ids, partition_strategy=’mod’, name=None, validate_indices=True) Query tensor value in embedding list params according to index IDSIf len (params) > 1, ID will be divided according to the partition_strategy policy.1、If partition_strategy is “mod”,idThe assigned location is p = ID% len (params)For example, there are 13 IDS, which are divided into 5 locations.[[0, 5, 10], [1, 6, 11], [2, 7, 12], [3, 8], [4, 9]]2、If partition_strategy is “div”, the allocation plan is:[[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’) For a given IDs and weight query embedding1、sp_idsA sparse tensor for a N x M,NFor batch size, M is arbitrary, data type Int642、sp_weightsSparse tensor weights of shape and sp_ids,Floating point type, if it is None, the weight is all ‘1’.
• ## Recurrent neural network (Recurrent Neural Networks)

 operation describe tf.nn.rnn(cell, inputs, initial_state=None, dtype=None, sequence_length=None, scope=None) Building recurrent neural network based on RNNCell class cell 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) Dynamic cell based on RNNCell class cellUnlike general RNN, the function is dynamically expanded according to input.Return (outputs, state) tf.nn.state_saving_rnn(cell, inputs, state_saver, state_name, sequence_length=None, scope=None) RNN network that can store debug status tf.nn.bidirectional_rnn(cell_fw, cell_bw, inputs, initial_state_fw=None, initial_state_bw=None, dtype=None,sequence_length=None, scope=None) Bi-directional RNN, returning a 3 tuple of tuple(outputs, output_state_fw, output_state_bw)
• ## Evaluation network (Evaluation)

operationdescribe
tf.nn.top_k(input, k=1, sorted=True, name=None)Returns the value of the previous K large and its corresponding index
tf.nn.in_top_k(predictions, targets, k, name=None)Returns the corresponding predictions value for targets index.
Whether it is in the K position before predictions,
The data type is bool, and Len is the same as predictions.
• Supervised candidate sampling network (Candidate Sampling)

For a large number of multi – and multi – label models, if full connection softmax will take up a lot of time and space resources, the candidate sampling method uses only a small portion of the category and label as supervision to speed up the training.

 operation describe 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’) Training loss results to return to noise-contrastive 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’) Training loss to return to sampled softmaxReference – Jean et al., 2014 third parts Candidate Samplers tf.nn.uniform_candidate_sampler(true_classes, num_true, num_sampled, unique, range_max, seed=None, name=None) Sampling set by uniform distributionReturn three yuan tuple1、sampled_candidates Candidate sets.2、Expected true_classes number, floating point value3、Expected sampled_candidates number, floating point value tf.nn.log_uniform_candidate_sampler(true_classes, num_true,num_sampled, unique, range_max, seed=None, name=None) Through the uniformly distributed sampling set of log, three yuan tuple is returned. tf.nn.learned_unigram_candidate_sampler(true_classes, num_true, num_sampled, unique, range_max, seed=None, name=None) Sampling based on the distribution learned during training.Return three yuan 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) Sampling based on the basic distribution provided

# Nine. Preservation and recovery variables

 operation describe Class 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) Create a memory Savervar_listDefine variables that need to be stored and restored tf.train.Saver.save(sess, save_path, global_step=None, latest_filename=None, meta_graph_suffix=’meta’,write_meta_graph=True) Save variable tf.train.Saver.restore(sess, save_path) Recovery variable tf.train.Saver.last_checkpoints List the recently undeleted checkpoint file name tf.train.Saver.set_last_checkpoints(last_checkpoints) Setting up a list of checkpoint filenames tf.train.Saver.set_last_checkpoints_with_time(last_checkpoints_with_time) Setting up checkpoint file name list and timestamp
Link of this Article: TensorFlow common function summary