dl框架支持的算子(欢迎补充更正)

最新推荐文章于 2025-09-01 15:21:03 发布

原创最新推荐文章于 2025-09-01 15:21:03 发布 · 630 阅读

1 ·

CC 4.0 BY-SA版权

torch 同时被 3 个专栏收录

18 篇文章

订阅专栏

caffe

7 篇文章

订阅专栏

mxnet

3 篇文章

订阅专栏

本文详细对比了Caffe、mxnet、onnx、Tensorflow及tf_lite等深度学习框架所支持的算子集，涵盖了从基本数学运算到复杂神经网络层的各种算子，为开发者选择合适的框架提供参考。

支持的算子
● Caffe
Accuracy, BatchNorm, Clip, Concat, Convolution, Data, Deconvolution, DepthwiseConvolution, DetectionOutput, Dropout, Eltwise, Flatten, InnerProduct, Input, LRN, Normalize, PReLU, Permute, Pooling, Power, PriorBox, ROIPooling, RPN, ReLU, ReLU6, Region, Reorg, Reshape, Resize, Scale, Sigmoid, Slice, Softmax, SoftmaxWithLoss, Split, TanH, Tile, Upsample

● mxnet
Activation, BatchNorm, Concat, Convolution, Copy, Crop, Deconvolution, Dropout, FullyConnected, LeakyReLU, Pooling, RNN, Reshape, SoftmaxActivation, SoftmaxOutput, SwapAxis, UpSampling, _minus_scalar, _mul_scalar, add_n, clip, elemwise_add, transpose

● onnx
Add, AveragePool, BatchNormalization, Concat, Conv, Dropout, Flatten, Gemm, GlobalAveragePool, MaxPool, Relu, Softmax

● Tensorflow
Add, AddN, ArgMax, ArgMin, AudioSpectrogram, AvgPool, ComposedBN, ConcatV2, Conv2D, Conv2DBackpropInput, DecodeWav, DepthwiseConv2dNative, Dropout, Exp, FIFOQueueV2, Flatten, Floor, FusedBatchNorm, GRU, LRN, LSTM, Log, MatMul, MaxPool, Maximum, Mean, Mfcc, Minimum, MirrorPad, Mul, Pad, Pow, RNN, RealDiv, Relu, Relu6, Reshape, ResizeNearestNeighbor, ReverseV2, Rsqrt, Sigmoid, Softmax, Split, Sqrt, StridedSlice, Sub, Sum, Tanh, TopKV2

● tf_lite
ADD, AVERAGE_POOL_2D, CONCATENATION, CONV_2D, DEPTHWISE_CONV_2D, LOGISTIC, MAX_POOL_2D, RESHAPE, SOFTMAX, SQUEEZE, TFLite_Detection_PostProcess