Pytorch uses tensorboard to realize visual summary

One 、 Introduce ：

Tensorboard yes TensorFlow An additional tool for , Can record the number of training process 、 Images, etc , To facilitate researchers to observe the neural network training process . But for PyTorch Other neural network training frameworks have no functions like Tensorboard Similar tools as comprehensive , Some existing tools have limited functions or are difficult to use (tensorboard_logger, visdom etc. ) .TensorboardX This tool makes TensorFlow Other neural network frameworks can also be used Tensorboard The convenience function of . notes ：pytorch 1.1.0 Future native support tensorboard, No installation required tensorboardX

Two 、 Environmental preparation ：

This process is cumbersome , It has gone through the following steps ：

1. The model is in docker Running in , It's set up tensorboardx and pytorch etc. , but docker There is no graphical interface , Can't use google-chrome（tensorboardx Give a link , To view in the default browser , Online recommendation chrome, Be careful tensorboard and chrome Must be used in the same environment !）

2. Try to docker Install on the server chrome, success , The method is to download chrome The installation files （google-chrome-stable_current_amd64.deb） Format to local , Install again , Such as ：

 sudo dpkg -i google-chrome-stable_current_amd64.deb

 open chrome command ：/usr/bin/google-chrome

3. Only python2.7 Environmental Science （ function tensorboardX Will report a mistake ）, Equipped with python3.7、tensorflow、tensorboardX. install python:https://blog.csdn.net/m0_37814112/article/details/103039263, The other two use pip Can be installed （ If it doesn't work, be patient and try several times ） Pay attention to python3.7 The path of maps to the default python route

4. After successful installation , Enter the command tensorboard --logdir yourpath,yourpath It's for storage. tensorboard Directory of output files , It will be a pile events file , If successful, display a website ：http://localhost:6006/,IP And port number can be set , If there is garbled code, it is definitely unsuccessful .

This process should also be possible in windows Of chrome Project to the server environment to operate , But I didn't try .

3、 ... and 、pytorch Specific usage in ：

Here is mainly used to save training and verification loss diagram 、 Input and output pictures at each step .

1. Import summarywriter package

from tensorboardX import SummaryWriter

pytorch1.1.0 above ：

from torch.utils.TensorBoard import SummaryWriter

2. branch train and val Two modes , And set the address （ Multi card training is set to 0 Number cuda Output ）

self.writers = {}
if self.rank == 0:
    for mode in ["train", "val"]:
        self.writers[mode] = SummaryWriter(os.path.join(self.log_path, mode))

3. Set every certain number of batch Save at the end loss（ You can also save others ）, Simultaneous recording step

writer = self.writers[mode]
for l, v in losses.items():
    writer.add_scalar("{}".format(l), v, self.step)

If you need to save pictures , Save as needed

writer.add_image(
    "color_{}_{}/{}".format(frame_id, s, j),
    inputs[("color", frame_id, s)][j].data, self.step)

4. Close it after training summarywriter:

if len(self.writers) != 0:
    self.writers['train'].close()
    self.writers['val'].close()

Four 、 Effect display ：（ Can be displayed separately train and val Visualization ）

loss visualization ： Move the mouse to see each step Specific data , You can also see the saved pictures and other information