# 3.18. Logging with tensorboard¶

class ANNarchy.extensions.tensorboard.Logger(logdir='runs/', experiment=None)[source]

Logger class to use tensorboard to visualize ANNarchy simulations. Requires the tensorboardX package (pip install tensorboardX).

The Logger class is a thin wrapper around tensorboardX.SummaryWriter, which you could also use directly. The doc is available at <https://tensorboardx.readthedocs.io/>. Tensorboard can read any logging data, as long as they are saved in the right format (tfevents), so it is not limited to tensorflow. TensorboardX has been developed to allow the use of tensorboard with pytorch.

The extension has to be imported explictly:

from ANNarchy.extensions.tensorboard import Logger


The Logger class has to be closed properly at the end of the script, so it is advised to use a context:

with Logger() as logger:


You can also make sure to close it:

logger = Logger()
logger.close()


By default, the logs will be written in a subfolder of ./runs/ (which will be created in the current directory). The subfolder is a combination of the current datetime and of the hostname, e.g. ./runs/Apr22_12-11-22_machine. You can control these two elements by passing arguments to Logger():

with Logger(logdir="/tmp/annarchy", experiment="trial1"): # logs in /tmp/annarchy/trial1


The add_* methods allow you to log various structures, such as scalars, images, histograms, figures, etc.

A tag should be given to each plot. In the example above, the figure with the accuracy will be labelled “Accuracy” in tensorboard. You can also group plots together with tags such as “Global performance/Accuracy”, “Global performance/Error rate”, “Neural activity/Population 1”, etc.

After (or while) logging data within your simulation, run tensorboard in the terminal by specifying the log directory:

> tensorboard --logdir runs


TensorboardX enqueues the data in memory before writing to disk. You can force flushing with:

logger.flush()

Parameters: logdir – path (absolute or relative) to the logging directory. Subfolders will be created for each individual run. The default is “runs/” experiment – name of the subfolder for the current run. By default, it is a combination of the current time and the hostname (e.g. Apr22_12-11-22_machine). If you reuse an experiment name, the data will be appended.
add_figure(tag, figure, step=None, close=True)[source]

Logs a Matplotlib figure.

Parameters: tag – name of the image in tensorboard. figure – a list or 1D numpy array of values. step – time index. close – whether the logger will close the figure when done (default: True).

Example:

with Logger() as logger:
for trial in range(100):
simulate(1000.0)
fig = plt.figure()
plt.plot(pop.r)

add_histogram(tag, hist, step=None)[source]

Logs an histogram.

Parameters: tag – name of the figure in tensorboard. hist – a list or 1D numpy array of values. step – time index.

Example:

with Logger() as logger:
for trial in range(100):
simulate(1000.0)
weights= proj.w.flatten()

add_image(tag, img, step=None, equalize=False)[source]

Logs an image.

Parameters: tag – name of the figure in tensorboard. img – array for the image. step – time index. equalize – rescales the pixels between 0 and 1 using the min and max values of the array.

The image must be a numpy array of size (height, width) for monochrome images or (height, width, 3) for colored images. The values should either be integers between 0 and 255 or floats between 0 and 1. The parameter equalize forces the values to be between 0 and 1 by equalizing using the min/max values.

Example:

with Logger() as logger:
for trial in range(100):
simulate(1000.0)
img = pop.r.reshape((10, 10))
logger.add_image("Population / Firing rate", img, trial, equalize=True)

add_images(tag, img, step=None, equalize=False, equalize_per_image=False)[source]

Logs a set of images (e.g. receptive fields).

Parameters: tag – name of the figure in tensorboard. img – array for the images. step – time index. equalize – rescales the pixels between 0 and 1 using the min and max values of the array. equalize_per_image – whether the rescaling should be using the global min/max values of the array, or per image. Has no effect if equalize of False.

The numpy array must be of size (number, height, width) for monochrome images or (number, height, width, 3) for colored images. The values should either be integers between 0 and 255 or floats between 0 and 1. The parameter equalize forces the values to be between 0 and 1 by equalizing using the min/max values.

Example:

with Logger() as logger:
for trial in range(100):
simulate(1000.0)
weights= proj.w.reshape(100, 10, 10) # 100 post neurons, 10*10 pre neurons

add_parameters(params, metrics)[source]

Logs parameters of a simulation.

Parameters: params – dictionary of parameters. metrics – dictionary of metrics.

This should be run only once per simulation, generally at the end. This allows to compare different runs of the same network using different parameter values and study how they influence the global output metrics, such as accuracy, error rate, reaction speed, etc.

Example:

with Logger() as logger:
# ...
logger.add_parameters({'learning_rate': lr, 'tau': tau}, {'accuracy': accuracy})

add_scalar(tag, value, step=None)[source]

Logs a single scalar value, e.g. a success rate at various stages of learning.

Parameters: tag – name of the figure in tensorboard. value – value. step – time index.

Example:

with Logger() as logger:
for trial in range(100):
simulate(1000.0)
accuracy = ...

add_scalars(tag, value, step=None)[source]

Logs multiple scalar values to be displayed in the same figure, e.g. several metrics or neural activities.

Parameters: tag – name of the figure in tensorboard. value – dictionary of values. step – time index.

Example:

with Logger() as logger:
for trial in range(100):
simulate(1000.0)
act1 = pop.r[0]
act2 = pop.r[1]

close()[source]
flush()[source]