This package contains small utility code to enable easy logging to TensorBoardX for experiments managed via seml.

pip install git+https://github.com/n-gao/seml_logger.git

from sacred import Experiment
from seml_logger import automain, Logger

ex = Experiment()

...

def naming_fn(dataset, **_):
    # This function shall return a name for the tensorboard run
    # you may also simply use a timestamp instead but you can also
    # grab any of the main parameters to construct a fitting name.
    return dataset

@automain(ex, naming_fn)
def main(..., dataset='MNIST', logger: Logger):
    ...

In addition to the manual logs, seml_logger also dumps all parameters in a file called config.json within the Tensorboard directory. The config file is also logged in Tensorboard directly via the Text functionality.

If the experiment crashes, i.e., the code returns with an exception or an error code != 0, the stacktrace is also logged via Tensorboard for easy access.

Finally, the results of the experiments are stored via pickle in results.pickle within the logging directory.

The Logger class automatically redirects attributes to tensorboardX.SummaryWriter, so you can directly use the logger as

for i in range(10):
    logger.add_scalar(i, global_step=i)

Additionally, several utility functions are implemented to log dictionaries of parameters directly.

To store parameters or intermediate variables we use numpy.save and numpy.savez. You may store arrays in the tensorboard directory via logger.store_array or logger.store_dict.

To log lots of numerical data, HDF5 presents a nicely accessible way of storing such data. To create an HDF5 dataset simply call

logger.create_dataset('name', shape=(...), ...)

The interface is identical to h5py.File.create_dataset. Further, after creating datasets, you may access them directly from the logger via indexing

# works after creating the `name` dataset.
logger['name']