tensorflow模型转ncnn的操作方式
作者:MatthewHsw 时间:2022-10-29 15:55:19
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 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, 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 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)
来源:https://blog.csdn.net/Murdock_C/article/details/88043784