第一步把tensorflow儲存的.ckpt模型轉為pb模型,並記下模型的輸入輸出名字.

第二步去ncnn的github上把倉庫clone下來,按照上面的要求裝好依賴並make.

第三步是修改ncnn的CMakeList,具體修改的位置有:

ncnn/CMakeList.txt 檔案,在檔案開頭處加入add_definitions(-std=c++11),末尾處加上add_subdirectory(examples),如果ncnn沒有examples資料夾,就新建一個,並加上CMakeList.txt檔案.

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

原版的tools/tensorflow/tensorflow2ncnn.cpp裡,不支援tensorflow的elu,FusedBathNormalization,Conv2dBackpropback操作,其實elu是支援的,只需要仿照relu的格式,在.cpp檔案里加上就行. FusedBatchNormalization就是ncnn/layer/裡實現的batchnorm.cpp,只是`tensorflow2ncnn裡沒有寫上,可以增加下面的內容：

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,bp);
   }
   for (int i=0; i< channels; ++i)
   {
    fwrite(var_data+i,bp);
   }
   for (int i=0; i< channels; ++i)
   {
    fwrite(b_data+i,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其實就是ncnn裡的反捲積操作,只不過ncnn實現反捲積的操作和tensorflow內部實現反捲積的操作過程不一樣,但結果是一致的,需要仿照普通卷積的寫法加上去.

ncnn同樣支援空洞卷積,但無法識別tensorflow的空洞卷積,具體原理可以看tensorflow空洞卷積的原理,tensorflow是改變featuremap做空洞卷積,而ncnn是改變kernel做空洞卷積,結果都一樣. 需要對.proto檔案修改即可完成空洞卷積.

總之ncnn對tensorflow的支援很不友好,有的層還需要自己手動去實現,還是很麻煩.

補充知識：pytorch模型轉mxnet

介紹

gluon把mxnet再進行封裝，封裝的風格非常接近pytorch

使用gluon的好處是非常容易把pytorch模型向mxnet轉化

唯一的問題是gluon封裝還不成熟，封裝好的layer不多，很多常用的layer 如concat,upsampling等layer都沒有

這裡關注如何把pytorch 模型快速轉換成 mxnet基於symbol 和 exector設計的網路

pytorch轉mxnet module

關鍵點:

mxnet 設計網路時symbol 名稱要和pytorch初始化中各網路層名稱對應

torch.load()讀入pytorch模型checkpoint 字典，取當中的'state_dict'元素，也是一個字典

pytorch state_dict 字典中key是網路層引數的名稱，val是引數ndarray

pytorch 的引數名稱的組織形式和mxnet一樣，但是連線符號不同，pytorch是'.'，而mxnet是'_'比如：

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

pytorch 的引數array 和mxnet 的引數array 完全一樣，只要名稱對上，直接賦值即可初始化mxnet模型

需要做的有以下幾點：

設計和pytorch網路對應的mxnet網路

載入pytorch checkpoint

調整pytorch checkpoint state_dict 的key名稱和mxnet命名格式一致

FlowNet2S PytorchToMxnet

pytorch flownet2S 的checkpoint 可以在github上搜到

import mxnet as mx
from 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,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,kernel=(5,5),num_filter=128,name=name + 'conv2_0')
 conv2 = mx.sym.LeakyReLU(data=conv2,slope=0.1)
 
 conv3a = mx.sym.Convolution(conv2,num_filter=256,name=name + 'conv3_0')
 conv3a = mx.sym.LeakyReLU(data=conv3a,slope=0.1)
 
 conv3b = mx.sym.Convolution(conv3a,pad=(1,1),kernel=(3,stride=(1,name=name + 'conv3_1_0')
 conv3b = mx.sym.LeakyReLU(data=conv3b,slope=0.1)
 
 conv4a = mx.sym.Convolution(conv3b,num_filter=512,name=name + 'conv4_0')
 conv4a = mx.sym.LeakyReLU(data=conv4a,slope=0.1)
 
 conv4b = mx.sym.Convolution(conv4a,name=name + 'conv4_1_0')
 conv4b = mx.sym.LeakyReLU(data=conv4b,slope=0.1)
 
 conv5a = mx.sym.Convolution(conv4b,name=name + 'conv5_0')
 conv5a = mx.sym.LeakyReLU(data=conv5a,slope=0.1)
 
 conv5b = mx.sym.Convolution(conv5a,name=name + 'conv5_1_0')
 conv5b = mx.sym.LeakyReLU(data=conv5b,slope=0.1)
 
 conv6a = mx.sym.Convolution(conv5b,num_filter=1024,name=name + 'conv6_0')
 conv6a = mx.sym.LeakyReLU(data=conv6a,slope=0.1)
 
 conv6b = mx.sym.Convolution(conv6a,name=name + 'conv6_1_0')
 conv6b = mx.sym.LeakyReLU(data=conv6b,slope=0.1,)
 
 #predict flow
 pr6 = mx.sym.Convolution(conv6b,num_filter=2,name=name + 'predict_flow6')
 prediction['loss6'] = pr6
 
 upsample_pr6to5 = mx.sym.Deconvolution(pr6,kernel=(4,4),name=name + 'upsampled_flow6_to_5',no_bias=True)
 upconv5 = mx.sym.Deconvolution(conv6b,name=name + 'deconv5_0',no_bias=False)
 upconv5 = mx.sym.LeakyReLU(data=upconv5,slope=0.1)
 iconv5 = mx.sym.Concat(conv5b,upconv5,upsample_pr6to5,dim=1)
 
 
 pr5 = mx.sym.Convolution(iconv5,name=name + 'predict_flow5')
 prediction['loss5'] = pr5
 
 upconv4 = mx.sym.Deconvolution(iconv5,name=name + 'deconv4_0',no_bias=False)
 upconv4 = mx.sym.LeakyReLU(data=upconv4,slope=0.1)
 
 upsample_pr5to4 = mx.sym.Deconvolution(pr5,name=name + 'upsampled_flow5_to_4',no_bias=True)
 
 iconv4 = mx.sym.Concat(conv4b,upconv4,upsample_pr5to4)
 
 pr4 = mx.sym.Convolution(iconv4,name=name + 'predict_flow4')
 prediction['loss4'] = pr4
 
 upconv3 = mx.sym.Deconvolution(iconv4,name=name + 'deconv3_0',no_bias=False)
 upconv3 = mx.sym.LeakyReLU(data=upconv3,slope=0.1)
 
 upsample_pr4to3 = mx.sym.Deconvolution(pr4,name= name + 'upsampled_flow4_to_3',no_bias=True)
 iconv3 = mx.sym.Concat(conv3b,upconv3,upsample_pr4to3)
 
 pr3 = mx.sym.Convolution(iconv3,name=name + 'predict_flow3')
 prediction['loss3'] = pr3
 
 upconv2 = mx.sym.Deconvolution(iconv3,name=name + 'deconv2_0',no_bias=False)
 upconv2 = mx.sym.LeakyReLU(data=upconv2,slope=0.1)
 
 upsample_pr3to2 = mx.sym.Deconvolution(pr3,name=name + 'upsampled_flow3_to_2',no_bias=True)
 iconv2 = mx.sym.Concat(conv2,upconv2,upsample_pr3to2)
 
 pr2 = mx.sym.Convolution(iconv2,name=name + 'predict_flow2')
 prediction['loss2'] = pr2
 flow = mx.sym.UpSampling(arg0=pr2,scale=4,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,is_sparse=is_sparse,type='flow'))
  loss.append(mx.sym.MAERegressionOutput(data=prediction[key] * 20,label=labels[key],grad_scale=loss_scale[key]))
 # print('loss: ',loss)
 #group 暫時不知道為嘛要group
 loss_group =mx.sym.Group(loss)
 # print('net: ',loss_group)
 return loss_group,flow
 
import gluonbook as gb
import torch
from utils.frame_utils import *
import numpy as np
if __name__ == '__main__':
 checkpoint = torch.load("C:/Users/junjie.huang/PycharmProjects/flownet2_mxnet/flownet2_pytorch/FlowNet2-S_checkpoint.pth.tar")
 # # checkpoint是一個字典
 print(isinstance(checkpoint['state_dict'],dict))
 # # 列印checkpoint字典中的key名
 print('keys of checkpoint:')
 for i in checkpoint:
  print(i)
 print('')
 # # pytorch 模型引數儲存在一個key名為'state_dict'的元素中
 state_dict = checkpoint['state_dict']
 # # state_dict也是一個字典
 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']))
 #檢視某個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,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 模型引數'''
 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(0,2).data
 exe.arg_dict['img2'][:] = pim2[np.newaxis,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)
 
 #來自pytroch flownet2
 from visualize import flow2color
 # color = flow2color(exe.outputs[0].asnumpy()[0].transpose(1,0))
 flow_color = flow2color(exe.outputs[0].asnumpy()[0].transpose(1,0))
 print('color type:',type(flow_color))
 import matplotlib.pyplot as plt
 #來自pytorch
 from torchvision.transforms import ToPILImage
 TF = ToPILImage()
 images = TF(flow_color)
 images.show()
 # plt.imshow(color)

以上這篇tensorflow模型轉ncnn的操作方式就是小編分享給大家的全部內容了，希望能給大家一個參考，也希望大家多多支援我們。