第一步把tensorflow保存的.ckpt模型转为pb模型, 并记下模型的输入输出名字.

第二步去ncnn的github上把仓库clone下来, 按照上面的要求装好依赖并make.

第三步是修改ncnn的CMakeList, 具体修改的位置有:

ncnn/CMakeList.txt 文件, 在文件开头处加入add_definitions(-std=c++11), 末尾处加上add_subdirectory(examples), 如果ncnn没有examples文件夹,就新建一个, 并加上CMakeList.txt文件.

ncnn/tools/CMakeList.txt 文件, 加入add_subdirectory(tensorflow)

原版的tools/tensorflow/tensorflow2ncnn.cpp里, 不支持tensorflow的elu, FusedBathNormalization, Conv2dBackpropback操作, 其实elu是支持的,只需要仿照relu的格式, 在.cpp文件里加上就行. FusedBatchNormalization就是ncnn/layer/里实现的batchnorm.cpp, 只是`tensorflow2ncnn里没有写上, 可以增加下面的内容：

else if (node.op() == "FusedBatchNorm"){ fprintf(pp, "%-16s", "BatchNorm");}...else if (node.op() == "FusedBatchNorm"){ std::cout << "node name is FusedBatchNorm" << std::endl; tensorflow::TensorProto tensor; find_tensor_proto(weights, node, tensor); const tensorflow::TensorShapeProto& shape = tensor.tensor_shape(); const tensorflow::TensorProto& gamma = weights[node.input(1)]; const tensorflow::TensorProto& Beta = weights[node.input(2)]; const tensorflow::TensorProto& mean = weights[node.input(3)]; const tensorflow::TensorProto& var = weights[node.input(4)]; int channels = gamma.tensor_shape().dim(0).size(); // data size int dtype = gamma.dtype(); switch (dtype){ case 1: { const float * gamma_tensor = reinterpret_cast<const float *>(gamma.tensor_content().c_str()); const float * mean_data = reinterpret_cast<const float *>(mean.tensor_content().c_str()); const float * var_data = reinterpret_cast<const float *>(var.tensor_content().c_str()); const float * b_data = reinterpret_cast<const float *>(Beta.tensor_content().c_str()); for (int i=0; i< channels; ++i) { fwrite(gamma_tensor+i, sizeof(float), 1, bp); } for (int i=0; i< channels; ++i) { fwrite(mean_data+i, sizeof(float), 1, bp); } for (int i=0; i< channels; ++i) { fwrite(var_data+i, sizeof(float), 1, bp); } for (int i=0; i< channels; ++i) { fwrite(b_data+i, sizeof(float), 1, bp); } } default: std::cerr << "Type is not supported." << std::endl; } fprintf(pp, " 0=%d", channels); tensorflow::AttrValue value_epsilon; if (find_attr_value(node, "epsilon", value_epsilon)){ float epsilon = value_epsilon.f(); fprintf(pp, " 1=%f", epsilon); }}

同理, Conv2dBackpropback其实就是ncnn里的反卷积操作, 只不过ncnn实现反卷积的操作和tensorflow内部实现反卷积的操作过程不一样, 但结果是一致的, 需要仿照普通卷积的写法加上去.

ncnn同样支持空洞卷积, 但无法识别tensorflow的空洞卷积, 具体原理可以看tensorflow空洞卷积的原理, tensorflow是改变featuremap做空洞卷积, 而ncnn是改变kernel做空洞卷积, 结果都一样. 需要对.proto文件修改即可完成空洞卷积.

总之ncnn对tensorflow的支持很不友好, 有的层还需要自己手动去实现, 还是很麻烦.

补充知识：pytorch模型转mxnet

介绍

gluon把mxnet再进行封装，封装的风格非常接近pytorch

使用gluon的好处是非常容易把pytorch模型向mxnet转化

唯一的问题是gluon封装还不成熟，封装好的layer不多，很多常用的layer 如concat,upsampling等layer都没有

这里关注如何把pytorch 模型快速转换成 mxnet基于symbol 和 exector设计的网络

pytorch转mxnet module

关键点:

mxnet 设计网络时symbol 名称要和pytorch初始化中各网络层名称对应

torch.load()读入pytorch模型checkpoint 字典，取当中的'state_dict'元素，也是一个字典

pytorch state_dict 字典中key是网络层参数的名称，val是参数ndarray

pytorch 的参数名称的组织形式和mxnet一样，但是连接符号不同，pytorch是'.'，而mxnet是'_'比如：

pytorch '0.conv1.0.weight'
mxnet '0_conv1_0_weight'

pytorch 的参数array 和mxnet 的参数array 完全一样，只要名称对上，直接赋值即可初始化mxnet模型

需要做的有以下几点：

设计和pytorch网络对应的mxnet网络

加载pytorch checkpoint

调整pytorch checkpoint state_dict 的key名称和mxnet命名格式一致

FlowNet2S PytorchToMxnet

pytorch flownet2S 的checkpoint 可以在github上搜到

import mxnet as mxfrom symbol_util import *import pickle def get_loss(data, label, loss_scale, name, get_input=False, is_sparse = False, type='stereo'): if type == 'stereo': data = mx.sym.Activation(data=data, act_type='relu',name=name+'relu') # loss if is_sparse: loss =mx.symbol.Custom(data=data, label=label, name=name, loss_scale= loss_scale, is_l1=True, op_type='SparseRegressionLoss') else: loss = mx.sym.MAERegressionOutput(data=data, label=label, name=name, grad_scale=loss_scale) return (loss,data) if get_input else loss def flownet_s(loss_scale, is_sparse=False, name=''): img1 = mx.symbol.Variable('img1') img2 = mx.symbol.Variable('img2') data = mx.symbol.concat(img1,img2,dim=1) labels = {'loss{}'.format(i): mx.sym.Variable('loss{}_label'.format(i)) for i in range(0, 7)} # print('labels: ',labels) prediction = {}# a dict for loss collection loss = []#a list #normalize data = (data-125)/255 # extract featrue conv1 = mx.sym.Convolution(data, pad=(3, 3), kernel=(7, 7), stride=(2, 2), num_filter=64, name=name + 'conv1_0') conv1 = mx.sym.LeakyReLU(data=conv1, act_type='leaky', slope=0.1) conv2 = mx.sym.Convolution(conv1, pad=(2, 2), kernel=(5, 5), stride=(2, 2), num_filter=128, name=name + 'conv2_0') conv2 = mx.sym.LeakyReLU(data=conv2, act_type='leaky', slope=0.1) conv3a = mx.sym.Convolution(conv2, pad=(2, 2), kernel=(5, 5), stride=(2, 2), num_filter=256, name=name + 'conv3_0') conv3a = mx.sym.LeakyReLU(data=conv3a, act_type='leaky', slope=0.1) conv3b = mx.sym.Convolution(conv3a, pad=(1, 1), kernel=(3, 3), stride=(1, 1), num_filter=256, name=name + 'conv3_1_0') conv3b = mx.sym.LeakyReLU(data=conv3b, act_type='leaky', slope=0.1) conv4a = mx.sym.Convolution(conv3b, pad=(1, 1), kernel=(3, 3), stride=(2, 2), num_filter=512, name=name + 'conv4_0') conv4a = mx.sym.LeakyReLU(data=conv4a, act_type='leaky', slope=0.1) conv4b = mx.sym.Convolution(conv4a, pad=(1, 1), kernel=(3, 3), stride=(1, 1), num_filter=512, name=name + 'conv4_1_0') conv4b = mx.sym.LeakyReLU(data=conv4b, act_type='leaky', slope=0.1) conv5a = mx.sym.Convolution(conv4b, pad=(1, 1), kernel=(3, 3), stride=(2, 2), num_filter=512, name=name + 'conv5_0') conv5a = mx.sym.LeakyReLU(data=conv5a, act_type='leaky', slope=0.1) conv5b = mx.sym.Convolution(conv5a, pad=(1, 1), kernel=(3, 3), stride=(1, 1), num_filter=512, name=name + 'conv5_1_0') conv5b = mx.sym.LeakyReLU(data=conv5b, act_type='leaky', slope=0.1) conv6a = mx.sym.Convolution(conv5b, pad=(1, 1), kernel=(3, 3), stride=(2, 2), num_filter=1024, name=name + 'conv6_0') conv6a = mx.sym.LeakyReLU(data=conv6a, act_type='leaky', slope=0.1) conv6b = mx.sym.Convolution(conv6a, pad=(1, 1), kernel=(3, 3), stride=(1, 1), num_filter=1024, name=name + 'conv6_1_0') conv6b = mx.sym.LeakyReLU(data=conv6b, act_type='leaky', slope=0.1, ) #predict flow pr6 = mx.sym.Convolution(conv6b, pad=(1, 1), kernel=(3, 3), stride=(1, 1), num_filter=2, name=name + 'predict_flow6') prediction['loss6'] = pr6 upsample_pr6to5 = mx.sym.Deconvolution(pr6, pad=(1, 1), kernel=(4, 4), stride=(2, 2), num_filter=2, name=name + 'upsampled_flow6_to_5', no_bias=True) upconv5 = mx.sym.Deconvolution(conv6b, pad=(1, 1), kernel=(4, 4), stride=(2, 2), num_filter=512, name=name + 'deconv5_0', no_bias=False) upconv5 = mx.sym.LeakyReLU(data=upconv5, act_type='leaky', slope=0.1) iconv5 = mx.sym.Concat(conv5b, upconv5, upsample_pr6to5, dim=1) pr5 = mx.sym.Convolution(iconv5, pad=(1, 1), kernel=(3, 3), stride=(1, 1), num_filter=2, name=name + 'predict_flow5') prediction['loss5'] = pr5 upconv4 = mx.sym.Deconvolution(iconv5, pad=(1, 1), kernel=(4, 4), stride=(2, 2), num_filter=256, name=name + 'deconv4_0', no_bias=False) upconv4 = mx.sym.LeakyReLU(data=upconv4, act_type='leaky', slope=0.1) upsample_pr5to4 = mx.sym.Deconvolution(pr5, pad=(1, 1), kernel=(4, 4), stride=(2, 2), num_filter=2, name=name + 'upsampled_flow5_to_4', no_bias=True) iconv4 = mx.sym.Concat(conv4b, upconv4, upsample_pr5to4) pr4 = mx.sym.Convolution(iconv4, pad=(1, 1), kernel=(3, 3), stride=(1, 1), num_filter=2, name=name + 'predict_flow4') prediction['loss4'] = pr4 upconv3 = mx.sym.Deconvolution(iconv4, pad=(1, 1), kernel=(4, 4), stride=(2, 2), num_filter=128, name=name + 'deconv3_0', no_bias=False) upconv3 = mx.sym.LeakyReLU(data=upconv3, act_type='leaky', slope=0.1) upsample_pr4to3 = mx.sym.Deconvolution(pr4, pad=(1, 1), kernel=(4, 4), stride=(2, 2), num_filter=2, name= name + 'upsampled_flow4_to_3', no_bias=True) iconv3 = mx.sym.Concat(conv3b, upconv3, upsample_pr4to3) pr3 = mx.sym.Convolution(iconv3, pad=(1, 1), kernel=(3, 3), stride=(1, 1), num_filter=2, name=name + 'predict_flow3') prediction['loss3'] = pr3 upconv2 = mx.sym.Deconvolution(iconv3, pad=(1, 1), kernel=(4, 4), stride=(2, 2), num_filter=64, name=name + 'deconv2_0', no_bias=False) upconv2 = mx.sym.LeakyReLU(data=upconv2, act_type='leaky', slope=0.1) upsample_pr3to2 = mx.sym.Deconvolution(pr3, pad=(1, 1), kernel=(4, 4), stride=(2, 2), num_filter=2, name=name + 'upsampled_flow3_to_2', no_bias=True) iconv2 = mx.sym.Concat(conv2, upconv2, upsample_pr3to2) pr2 = mx.sym.Convolution(iconv2, pad=(1, 1), kernel=(3, 3), stride=(1, 1), num_filter=2, name=name + 'predict_flow2') prediction['loss2'] = pr2 flow = mx.sym.UpSampling(arg0=pr2,scale=4,num_filter=2,num_args = 1,sample_type='nearest', name='upsample_flow2_to_1') # ignore the loss functions with loss scale of zero keys = loss_scale.keys() # keys.sort() #obtain the symbol of the losses for key in keys: # loss.append(get_loss(prediction[key] * 20, labels[key], loss_scale[key], name=key + name,get_input=False, is_sparse=is_sparse, type='flow')) loss.append(mx.sym.MAERegressionOutput(data=prediction[key] * 20, label=labels[key], name=key + name, grad_scale=loss_scale[key])) # print('loss: ',loss) #group 暂时不知道为嘛要group loss_group =mx.sym.Group(loss) # print('net: ',loss_group) return loss_group,flow import gluonbook as gbimport torchfrom utils.frame_utils import *import numpy as npif __name__ == '__main__': checkpoint = torch.load("C:/Users/junjie.huang/PycharmProjects/flownet2_mxnet/flownet2_pytorch/FlowNet2-S_checkpoint.pth.tar") # # checkpoint是一个字典 print(isinstance(checkpoint['state_dict'], dict)) # # 打印checkpoint字典中的key名 print('keys of checkpoint:') for i in checkpoint: print(i) print('') # # pytorch 模型参数保存在一个key名为'state_dict'的元素中 state_dict = checkpoint['state_dict'] # # state_dict也是一个字典 print('keys of state_dict:') for i in state_dict: print(i) # print(state_dict[i].size()) print('') # print(state_dict) #字典的value是torch.tensor print(torch.is_tensor(state_dict['conv1.0.weight'])) #查看某个value的size print(state_dict['conv1.0.weight'].size()) #flownet-mxnet init loss_scale={'loss2': 1.00, 'loss3': 1.00, 'loss4': 1.00, 'loss5': 1.00, 'loss6': 1.00} loss,flow = flownet_s(loss_scale=loss_scale,is_sparse=False) print('loss information: ') print('loss:',loss) print('type:',type(loss)) print('list_arguments:',loss.list_arguments()) print('list_outputs:',loss.list_outputs()) print('list_inputs:',loss.list_inputs()) print('') print('flow information: ') print('flow:',flow) print('type:',type(flow)) print('list_arguments:',flow.list_arguments()) print('list_outputs:',flow.list_outputs()) print('list_inputs:',flow.list_inputs()) print('') name_mxnet = symbol.list_arguments() print(type(name_mxnet)) for key in name_mxnet: print(key) name_mxnet.sort() for key in name_mxnet: print(key) print(name_mxnet) shapes = (1, 3, 384, 512) ctx = gb.try_gpu() # exe = symbol.simple_bind(ctx=ctx, img1=shapes,img2=shapes) exe = flow.simple_bind(ctx=ctx, img1=shapes, img2=shapes) print('exe type: ',type(exe)) print('exe: ',exe) #module # mod = mx.mod.Module(flow) # print('mod type: ', type(exe)) # print('mod: ', exe) pim1 = read_gen("C:/Users/junjie.huang/PycharmProjects/flownet2_mxnet/data/0000007-img0.ppm") pim2 = read_gen("C:/Users/junjie.huang/PycharmProjects/flownet2_mxnet/data/0000007-img1.ppm") print(pim1.shape) '''使用pytorch 的state_dict 初始化 mxnet 模型参数''' for key in state_dict: # print(type(key)) k_split = key.split('.') key_mx = '_'.join(k_split) # print(key,key_mx) try: exe.arg_dict[key_mx][:]=state_dict[key].data except: print(key,exe.arg_dict[key_mx].shape,state_dict[key].data.shape) exe.arg_dict['img1'][:] = pim1[np.newaxis, :, :, :].transpose(0, 3, 1, 2).data exe.arg_dict['img2'][:] = pim2[np.newaxis, :, :, :].transpose(0, 3, 1, 2).data result = exe.forward() print('result: ',type(result)) # for tmp in result: # print(type(tmp)) # print(tmp.shape) # color = flow2color(exe.outputs[0].asnumpy()[0].transpose(1, 2, 0)) outputs = exe.outputs print('output type: ',type(outputs)) # for tmp in outputs: # print(type(tmp)) # print(tmp.shape) #来自pytroch flownet2 from visualize import flow2color # color = flow2color(exe.outputs[0].asnumpy()[0].transpose(1,2,0)) flow_color = flow2color(exe.outputs[0].asnumpy()[0].transpose(1, 2, 0)) print('color type:',type(flow_color)) import matplotlib.pyplot as plt #来自pytorch from torchvision.transforms import ToPILImage TF = ToPILImage() images = TF(flow_color) images.show() # plt.imshow(color)

以上这篇tensorflow模型转ncnn的操作方式就是小编分享给大家的全部内容了，希望能给大家一个参考，也希望大家多多支持。