tensorflow中神经网络相关函数算法

    xiaoxiao2021-03-25  116

    Neural Network

    激活算法

    The activation ops provide different types of nonlinearities for use in neural networks. These include smooth nonlinearities (sigmoid, tanh, elu, softplus, and softsign), continuous but not everywhere differentiable functions (relu, relu6, crelu and relu_x), and random regularization (dropout).

    All activation ops apply componentwise, and produce a tensor of the same shape as the input tensor.

    tf.nn.relutf.nn.relu6tf.nn.crelutf.nn.elutf.nn.softplustf.nn.softsigntf.nn.dropouttf.nn.bias_addtf.sigmoidtf.tanh

    卷积算法

    The convolution ops sweep a 2-D filter over a batch of images, applying the filter to each window of each image of the appropriate size. The different ops trade off between generic vs. specific filters:

    conv2d: Arbitrary filters that can mix channels together.depthwise_conv2d: Filters that operate on each channel independently.separable_conv2d: A depthwise spatial filter followed by a pointwise filter.

    Note that although these ops are called "convolution", they are strictly speaking "cross-correlation" since the filter is combined with an input window without reversing the filter. For details, see the properties of cross-correlation.

    The filter is applied to image patches of the same size as the filter and strided according to the strides argument.strides = [1, 1, 1, 1] applies the filter to a patch at every offset, strides = [1, 2, 2, 1] applies the filter to every other image patch in each dimension, etc.

    Ignoring channels for the moment, and assume that the 4-D input has shape [batch, in_height, in_width, ...]and the 4-D filter has shape [filter_height, filter_width, ...], then the spatial semantics of the convolution ops are as follows: first, according to the padding scheme chosen as 'SAME' or 'VALID', the output size and the padding pixels are computed. For the 'SAME' padding, the output height and width are computed as:

    out_height = ceil(float(in_height) / float(strides[1]))out_width  = ceil(float(in_width) / float(strides[2]))

    and the padding on the top and left are computed as:

    pad_along_height = max((out_height - 1) * strides[1] +                    filter_height - in_height, 0)pad_along_width = max((out_width - 1) * strides[2] +                   filter_width - in_width, 0)pad_top = pad_along_height // 2pad_bottom = pad_along_height - pad_toppad_left = pad_along_width // 2pad_right = pad_along_width - pad_left

    Note that the division by 2 means that there might be cases when the padding on both sides (top vs bottom, right vs left) are off by one. In this case, the bottom and right sides always get the one additional padded pixel. For example, when pad_along_height is 5, we pad 2 pixels at the top and 3 pixels at the bottom. Note that this is different from existing libraries such as cuDNN and Caffe, which explicitly specify the number of padded pixels and always pad the same number of pixels on both sides.

    For the 'VALID' padding, the output height and width are computed as:

    out_height = ceil(float(in_height - filter_height + 1) / float(strides[1]))out_width  = ceil(float(in_width - filter_width + 1) / float(strides[2]))

    and the padding values are always zero. The output is then computed as

    output[b, i, j, :] =    sum_{di, dj} input[b, strides[1] * i + di - pad_top,                       strides[2] * j + dj - pad_left, ...] *                 filter[di, dj, ...]

    where any value outside the original input image region are considered zero ( i.e. we pad zero values around the border of the image).

    Since input is 4-D, each input[b, i, j, :] is a vector. For conv2d, these vectors are multiplied by the filter[di, dj, :, :] matrices to produce new vectors. For depthwise_conv_2d, each scalar component input[b, i, j, k] is multiplied by a vector filter[di, dj, k], and all the vectors are concatenated.

    tf.nn.convolutiontf.nn.conv2dtf.nn.depthwise_conv2dtf.nn.depthwise_conv2d_nativetf.nn.separable_conv2dtf.nn.atrous_conv2dtf.nn.atrous_conv2d_transposetf.nn.conv2d_transposetf.nn.conv1dtf.nn.conv3dtf.nn.conv3d_transposetf.nn.conv2d_backprop_filtertf.nn.conv2d_backprop_inputtf.nn.conv3d_backprop_filter_v2tf.nn.depthwise_conv2d_native_backprop_filtertf.nn.depthwise_conv2d_native_backprop_input

    池化算法

    The pooling ops sweep a rectangular window over the input tensor, computing a reduction operation for each window (average, max, or max with argmax). Each pooling op uses rectangular windows of size ksize separated by offsetstrides. For example, if strides is all ones every window is used, if strides is all twos every other window is used in each dimension, etc.

    In detail, the output is

    output[i] = reduce(value[strides * i:strides * i + ksize])

    where the indices also take into consideration the padding values. Please refer to the Convolution section for details about the padding calculation.

    tf.nn.avg_pooltf.nn.max_pooltf.nn.max_pool_with_argmaxtf.nn.avg_pool3dtf.nn.max_pool3dtf.nn.fractional_avg_pooltf.nn.fractional_max_pooltf.nn.pool

    形态过滤算法

    Morphological operators are non-linear filters used in image processing.

    Greyscale morphological dilation is the max-sum counterpart of standard sum-product convolution:

    output[b, y, x, c] =    max_{dy, dx} input[b,                       strides[1] * y + rates[1] * dy,                       strides[2] * x + rates[2] * dx,                       c] +                 filter[dy, dx, c]

    The filter is usually called structuring function. Max-pooling is a special case of greyscale morphological dilation when the filter assumes all-zero values (a.k.a. flat structuring function).

    Greyscale morphological erosion is the min-sum counterpart of standard sum-product convolution:

    output[b, y, x, c] =    min_{dy, dx} input[b,                       strides[1] * y - rates[1] * dy,                       strides[2] * x - rates[2] * dx,                       c] -                 filter[dy, dx, c]

    Dilation and erosion are dual to each other. The dilation of the input signal f by the structuring signal g is equal to the negation of the erosion of -f by the reflected g, and vice versa.

    Striding and padding is carried out in exactly the same way as in standard convolution. Please refer to the Convolutionsection for details.

    tf.nn.dilation2dtf.nn.erosion2dtf.nn.with_space_to_batch

    正则化算法

    Normalization is useful to prevent neurons from saturating when inputs may have varying scale, and to aid generalization.

    tf.nn.l2_normalizetf.nn.local_response_normalizationtf.nn.sufficient_statisticstf.nn.normalize_momentstf.nn.momentstf.nn.weighted_momentstf.nn.fused_batch_normtf.nn.batch_normalizationtf.nn.batch_norm_with_global_normalization

    loss相关算法

    The loss ops measure error between two tensors, or between a tensor and zero. These can be used for measuring accuracy of a network in a regression task or for regularization purposes (weight decay).

    tf.nn.l2_losstf.nn.log_poisson_loss

    分类常用算法

    TensorFlow provides several operations that help you perform classification.

    tf.nn.sigmoid_cross_entropy_with_logitstf.nn.softmaxtf.nn.log_softmaxtf.nn.softmax_cross_entropy_with_logitstf.nn.sparse_softmax_cross_entropy_with_logitstf.nn.weighted_cross_entropy_with_logits

    嵌入算法

    TensorFlow provides library support for looking up values in embedding tensors.

    tf.nn.embedding_lookuptf.nn.embedding_lookup_sparse

    RNN

    TensorFlow provides a number of methods for constructing Recurrent Neural Networks. Most accept an RNNCell-subclassed object (see the documentation for tf.contrib.rnn).

    tf.nn.dynamic_rnntf.nn.bidirectional_dynamic_rnntf.nn.raw_rnn

    Connectionist Temporal Classification (CTC)

    tf.nn.ctc_losstf.nn.ctc_greedy_decodertf.nn.ctc_beam_search_decoder

    Evaluation

    The evaluation ops are useful for measuring the performance of a network. They are typically used at evaluation time.

    tf.nn.top_ktf.nn.in_top_k

    Candidate Sampling

    Do you want to train a multiclass or multilabel model with thousands or millions of output classes (for example, a language model with a large vocabulary)? Training with a full Softmax is slow in this case, since all of the classes are evaluated for every training example. Candidate Sampling training algorithms can speed up your step times by only considering a small randomly-chosen subset of contrastive classes (called candidates) for each batch of training examples.

    See our Candidate Sampling Algorithms Reference

    Sampled Loss Functions

    TensorFlow provides the following sampled loss functions for faster training.

    tf.nn.nce_losstf.nn.sampled_softmax_loss

    Candidate Samplers

    TensorFlow provides the following samplers for randomly sampling candidate classes when using one of the sampled loss functions above.

    tf.nn.uniform_candidate_samplertf.nn.log_uniform_candidate_samplertf.nn.learned_unigram_candidate_samplertf.nn.fixed_unigram_candidate_sampler

    Miscellaneous candidate sampling utilities

    tf.nn.compute_accidental_hits

    Quantization ops

    tf.nn.quantized_conv2dtf.nn.quantized_relu_xtf.nn.quantized_max_pooltf.nn.quantized_avg_pool
    转载请注明原文地址: https://ju.6miu.com/read-18559.html

    最新回复(0)