Mastering PyTorch Visualization: TensorBoard and Visdom Guide

PyTorch Visualization Tools: TensorBoard and Visdom

1. TensorBoard

TensorBoard is primarily a visualization tool for TensorFlow, but it can also be used independently with PyTorch to display computation graphs, scalar metrics, and custom data.

1. Install

pip install tensorboard

2. Launch

tensorboard --logdir="log directory"

When launching TensorBoard you can specify logdir, port (default 6006), host (default localhost) and other parameters.

usage: tensorboard [-h] [--helpfull] [--logdir PATH] [--logdir_spec PATH_SPEC]
                   [--host ADDR] [--bind_all] [--port PORT]
                   [--purge_orphaned_data BOOL] [--db URI] [--db_import]
                   [--inspect] [--version_tb] [--tag TAG] [--event_file PATH]
                   [--path_prefix PATH] [--window_title TEXT]
                   [--max_reload_threads COUNT] [--reload_interval SECONDS]
                   [--reload_task TYPE] [--reload_multifile BOOL]
                   [--reload_multifile_inactive_secs SECONDS]
                   [--generic_data TYPE]
                   [--samples_per_plugin SAMPLES_PER_PLUGIN]
                   [--debugger_data_server_grpc_port PORT]
                   [--debugger_port PORT] [--master_tpu_unsecure_channel ADDR]

3. TensorBoard demo (PyTorch)

During model training, use SummaryWriter to log loss, accuracy, etc.

# Import SummaryWriter
from torch.utils.tensorboard import SummaryWriter
# Create writer instance
summaryWriter = SummaryWriter(log_dir="/path/to/logs")
# Log training metrics
summaryWriter.add_scalars("loss", {"train_loss_avg": train_loss_avg, "test_loss_avg": test_loss_avg}, epoch)
summaryWriter.add_scalar("score", score, epoch)

Launch TensorBoard to visualize the training process: tensorboard --logdir=/path/to/logs Successful launch screenshot:

Resulting visualizations:

2. Visdom

Visdom, developed by Facebook for PyTorch, supports remote data visualization and works with Torch and NumPy.

1. Install

pip install visdom

2. Launch

python -m visdom.server

– start the server as a module nohup python -m visdom.server & – run in background on Linux/macOS

You can specify port (default 8097) and hostname (default localhost) when launching.

usage: server.py [-h] [-port port] [--hostname hostname] [-base_url base_url]
                [-env_path env_path] [-logging_level logger_level]
                [-readonly] [-enable_login] [-force_new_cookie]
                [-use_frontend_client_polling]

3. Visdom demo

Start Visdom server: python -m visdom.server -port 8097 Successful launch screenshot:

Training visualization code:

# Import visdom package
import visdom
# Create Visdom object, connect to server, specify environment
viz = visdom.Visdom(server='http://localhost', port=8097, env='liyunfei')
# Initialize lines
viz.line([0.], [0], win='train_loss', opts=dict(title='train_loss'))
viz.line([0.], [0], win='accuracy', opts=dict(title='accuracy'))
# During training, update loss and accuracy
viz.line([train_loss_avg], [epoch], win='train_loss', update='append')
viz.line([accuracy], [epoch], win='accuracy', update='append')

Resulting visualizations:

4. Visualizing images

Example of displaying a single random image and a batch of random images:

import visdom
import numpy as np
viz = visdom.Visdom(server='http://localhost', port=8097, env='liyunfei')
# Single image
viz.image(np.random.rand(3, 512, 256), opts=dict(title='Random!', caption='How random.'))
# Multiple images
viz.images(np.random.randn(20, 3, 64, 64), nrow=5,
           opts=dict(title='Random images', caption='How random.'))

Result:

5. Common Visdom APIs

vis.scatter : 2D or 3D scatter plot
vis.line    : line plot
vis.stem    : stem plot
vis.heatmap : heat map
vis.bar     : bar chart
vis.histogram : histogram
vis.boxplot : box plot
vis.surf    : surface plot
vis.contour : contour plot
vis.quiver  : 2D vector field
vis.image   : image display
vis.text    : text display
vis.mesh    : mesh plot
vis.save    : serialize state