Teach yourself to train pytorch model to Caffe (2)

The key point

Transform the training model , My model file is google_checkpoint_ep60

 Reference resources ：https://github.com/inisis/brocolli
https://www.bilibili.com/video/BV1fb4y177Ry

The elder brother built an environment for us , So you don't have to build your own environment , But we have to download docker

 Watch this video docker  packed  https://www.bilibili.com/video/BV11L411g7U1

Then we download the image provided by brother

Concrete , At the command line, follow my brother's github Guide knock

  The first line is to download the image （ It can be understood as a file in the virtual machine ）, The second line is to create a container （ Create a virtual machine and boot it , And login ）, I recommend that you put --rm Went to the , Otherwise cmd After exiting, the container is automatically deleted , The third line is to enter the container and execute the command

This command is to transform the model , Last saved in tmp In the folder .

  If there is no error, let's start to convert our own model

I use docker cp Copy between host and container

First of all, put our cats and dogs GoogLeNet.py Copy in .

docker cp E:/workspace/brocolli/custom_models/GoogLeNet.py ee132b3c68f8b98f204de3bbb8872573e9fb00db319a3dd8bb2c3674f6d4a776:/root/brocolli/custom_models 

The front is the host file or folder , Behind it is the container id ,id It can be downloaded from docker desktop Copy

 

  Finally, the container folder

Then we write a script to transform our model , Brother readme plagiarism/copycat

  My file is called run.py

import torchvision
import torch

from custom_models import GoogLeNet
model = GoogLeNet.GoogLeNet(num_classes=2) # Here, you should use your ownd model
model.load_state_dict(torch.load("./google_checkpoint_ep60.pth",map_location=torch.device('cpu')),)

# input  = torch.rand(1,3,224,224)
# output = model(input)
# aaa = 100

from bin.pytorch2caffe import Runner # if caffe, use bin.pytorch2caffe, if TensorRT use bin.pytorch2trt;
runner = Runner("googlenet_dog", model, [1, 3, 224, 224], 13,True)

runner.pyotrch_inference()
runner.convert()
runner.caffe_inference() # if caffe, use caffe_inference, if TensorRT use trt_inference;
runner.check_result()

We put run.py and google_checkpoint_ep60.pth Use both docker cp Copied to the /root/brocolli in

And then run , Conversion completed , hold tmp Below googlenet_dog.prototxt and googlenet_dog.caffemodel Copy is over

Split line ---------------------------------

docker Some orders

docker run Create a container

docker run -it  yaphets4desmond/brocolli:v1.0 bash

 If you used rm  Deleted , Use the above command to create a , Or so 

  If you handle it carefully bash The window is closed , But he is still running

  You can use the command to enter bash, The long one is the container id

docker exec -it  e132bd2faff7adb3596068d0f53316fc54d20309f2aed6464984206f3a811fe4 bash