Neural Network Visualization Tools
TensorBoard is a powerful visualization tool with two methods of calling it in pytorch:
tensorboardX import SummaryWriter
This method was written by a god on the internet when tensorboard was not officially supported.
import SummaryWriter
This method was officially added in a later update
1.1 Calling methods
1.1.1 Creating the interface SummaryWriter
Function: Create interface
Call method:
writer = SummaryWriter("runs")
Parameters:
log_dir: event file output folder
comment: folder suffix when log_dir is not specified
filename_suffix: event file filename suffix
1.1.2 Record scalars add_scalars()
Function: record scalars add_scalars()
Call method:
writer.add_scalars("name",{"dic":val},epoch)
Parameters:
tag: the tag name of the image
scalar_step: the scalar to be recorded
global_step: number of rounds
1.1.3 Statistical Histogram add_histogram()
Functions: statistical histograms and multi-quartile line graphs
Call method:
writer.add_histogram("weight",,epoch)
Parameters:
tag: the tag name of the image
values: data to be histogrammed
global_step: number of rounds
bins: values like 'tensorflow', 'auto', 'fd', etc.
1.1.4 Batch display image add_image()
Function: Batch display image
Call method:
writer.add_image(“Cifar10”, img_batch, epoch,'CHW')
Parameters:
tag: the tag name of the image
img_tensor: image data, note size
global_step: number of rounds
dataformats: data forms, CHW,HWC,HW
1.1.5 Viewing the model graph add_graph()
Function: View Model Drawing
Call method:
writer.add_graph(model=net,input_to_model=(1,3, 224, 224).to(device))
Parameters:
model: model, must be
input_to_model: data to be output to the model
verbose: whether or not to print information about the structure of the calculation diagram
Remember to write ()
2. View network layer shape, parameters torchsummary
Function: View network layer shape, parameters
Call method:
from torchsummary import summary summary(net, input_size=(3, 224, 224))
Parameters:
model: pytorch model
input_size: model input size
batch_size：batch size
device：“cuda” or “cpu”
3. Start tensorboard
Open a terminal with cmd in the file path and type
tensorboard --logdir="./runs"
Runs is the filename where I saved the file, open the following link
Addendum: pytorch call to tensorboard method attempts
The tensorboard provides a good interface for monitoring training losses and can help us better tune our parameters. The following section describes how to call tensorboard in pytorch.
in the first place
Installing tensorboard, tensorflow, and tensorboardX
secondly
Import SummaryWriter at the beginning of the file.
from tensorboardX import SummaryWriter
tertiary
Same as tensorflow's tensorboard, tensorboardX provides multiple recording methods such as scalar, image, and so on.
writer = SummaryWriter('path')
If you don't add path, it is named after the time by default.
fourth
Adding Monitor Variables
writer.add_scalar('Train/Acc', Acc, iter)
fifthly
Open tensorboard
tensorboard --logdir 'path'
sixthly
Open port 6006 in your browser
The above is a personal experience, I hope it can give you a reference, and I hope you can support me more. If there is any mistake or something that has not been fully considered, please do not hesitate to give me advice.