Logging development guidelines

Enabling logging

To enable logging support in your PennyLane work-flow simply run the following lines:

import pennylane as qml
qml.logging.enable_logging()

This will ensure all levels of the execution pipeline log function entries and outputs. These are logged to handlers pointing to the standard output. Note, since Python’s logging framework configuration functions overwrite previous logger configurations, it is expected that enable_logging() is called only when not using external logging configurations.

If you are defining custom logging configurations, you can extend the logging configuration options as defined in the section Customizing the logging configuration, or avoid calling enable_logging() in favour of tour custom configuration options.

Adding logging supports during development

To add logging support to components of PennyLane, we must define a module logger and ensure all scoped log statements use that logger. A logger can be imported and defined with:

import logging
logger = logging.getLogger(__name__)

which will be used within the given module, and track directories, filenames and function names, as we have defined the appropriate types within the formatter configuration (see pennylane.logging.DefaultFormatter). With the logger defined, we can selectively add to the logger by if-else statements, which compare the given module’s log-level to any log record message it receives. This step is not necessary, as the message will only output if the level is enabled, though if an expensive function call is required to build the string for the log-message, it can be faster to perform this check:

if logger.isEnabledFor(logging.DEBUG):
    logger.debug(
        """Entry with args=(arg_name_1=%s, arg_name_2=%s, ..., arg_name_n=%s)""",
        arg_name_1, arg_name_2, ..., arg_name_n,
    )

The above line can be added below the function/method entry point, and the provided arguments can be used to populate the log message. This allows us a way to track the inputs and calls through the stack in the order they are executed, as is the basis for following a trail of execution as needed.

All debug log-statements currently added to the PennyLane execution pipeline output log level messages with the originating function making that call, as this can allow us to track hard-to-find bug origins in the execution pipeline.

Customizing the logging configuration

As with any package that targets many domains, Python’s logging is as extensible and flexible as it is a challenge to configure for all user needs – ideally we define some good defaults that meet our development goals, and only deviate from them if required. For further details and customization options, consider reading the Python logging documentation how-to and “advanced” tutorial level.

PennyLane logging defaults are contained in a configuration file in the logging module titled log_config.toml, which is imported when logging is enabled. The file path can be accessed with qml.logging.config_path(). To change log-levels that are reporting on a package or module-wide basis, it is possible to do so by modifying the entries in the log_config.toml file, under the [loggers] section. In addition, if we want to send the logs elsewhere, we can adjust the [handlers] section, which controls what happens to each message. If we do not like the output format of the messages, we can adjust these through the [formatters] section. If we want to filter messages based on some criteria, we can add these to the respective handlers. As an example, we can go through the configuration file and explore the options.

Modifying the configuration options

To allow for good expressivity when requiring logging, we must often adjust several parts of the ecosystem to ensure messages are formatted a certain way, we control logging internally to PennyLane different to external packages, messages are sent to somewhere we can make them actionable, and we can remove messages that are not needed based on some criteria of interest. Let’s break the log_config.toml file into sections to discuss how these can be adjusted to suit needs:

###############################################################################
# Avoid interfering with existing loggers in dependent libraries
###############################################################################

disable_existing_loggers = false
version = 1

###############################################################################
# Bind formatters defined locally and those defined in pennylane.logging
###############################################################################

[formatters]
[formatters.qml_default_formatter]
"()" = "pennylane.logging.formatters.formatter.DefaultFormatter"

[formatters.qml_alt_formatter]
"()" = "pennylane.logging.formatters.formatter.AnotherLogFormatter"

[formatters.local_detailed]
format = "\u001B[38;2;45;145;210m[%(asctime)s][%(levelname)s][<PID %(process)d:%(processName)s>] - %(name)s.%(funcName)s()::\"%(message)s\"\u001B[0m"

[formatters.local_standard]
format = "[%(asctime)s] - %(name)s - %(levelname)s - %(message)s"

The first sections of the configuration file tell the logging infrastructure to avoid modification to existing log settings — this is set to true by default for backwards compatibility, though can be problematic if using external packages. It is recommended to keep this as false unless required otherwise.

###############################################################################
# Bind LogRecord filters defined in pennylane.logging module
###############################################################################

[filters]
# Filter to show messages from the same local process as the Python script
[filters.qml_LocalProcessFilter]
"()" = "pennylane.logging.filter.LocalProcessFilter"

# Filter to show debug level messages only
[filters.qml_DebugOnlyFilter]
"()" = "pennylane.logging.filter.DebugOnlyFilter"

The above section defines how to filter log messages (known as LogRecords), given some predicate. In this case, we have defined some classes, LocalProcessFilter and DebugOnlyFilter to filter based on process ID and on the severity of the incoming message. These can be used in the next section.

###############################################################################
# Bind handlers defined in the logging and in pennylane.logging modules
###############################################################################

[handlers]
[handlers.qml_debug_stream]
class = "logging.StreamHandler"
formatter = "qml_default_formatter"
level = "DEBUG"
stream = "ext://sys.stdout"

[handlers.qml_debug_stream_alt]
class = "logging.StreamHandler"
formatter = "qml_alt_formatter"
level = "DEBUG"
stream = "ext://sys.stdout"

[handlers.qml_debug_file]
class = "logging.handlers.RotatingFileHandler"
formatter = "local_standard"
level = "DEBUG"
filename ='qml_debug.log' # use `/tmp/filename.log` on Linux machines to avoid long-term persistence
maxBytes = 16777216 # 16MB per file before splitting
backupCount = 10 # Create 'qml_debug.log.1', ... 'qml_debug.log.backupCount' files and rollover when maxBytes is reached

[handlers.local_filtered_detailed_stdout]
class = "logging.StreamHandler"
formatter = "local_standard"
level = "DEBUG"
stream = "ext://sys.stdout"
filters = ["qml_LocalProcessFilter", "qml_DebugOnlyFilter"]

The above defines how LogRecord messages are handled, and directs them to the appropriate sink. The logging framework supports many such directions (see here for more info), but for this example we have defined stream handlers (printing to the screen via the standard output), and a file handler with a size cap at 16MB. Each handler can be customized by filters and formatters so that the consumed message fits the needs of the user.

###############################################################################
# Define logger controls for internal and external packages
###############################################################################

[loggers]

# Control JAX logging
[loggers.jax]
handlers = ["qml_debug_stream",]
level = "WARN"
propagate = false

# Control logging across pennylane
[loggers.pennylane]
handlers = ["qml_debug_stream",]
level = "DEBUG" # Set to TRACE for highest verbosity
propagate = false

# Control logging specifically in the pennylane.qnode module
# Note the required quotes to overcome TOML nesting issues
[loggers."pennylane.qnode"]
handlers = ["qml_debug_stream_alt",]
level = "DEBUG" # Set to TRACE for highest verbosity
propagate = false

###############################################################################

Finally, the loggers section which controls the individual loggers across the packages we are using. Python’s logging framework follows a parent-child hierarchy, where a logging configuration set at a parent level will set all child levels with the same features. In this instance, we have configured JAX, PennyLane and our script to all log into the qml_debug_stream handler we defined earlier, and modified the child logger "pennylane.qnode" (quotes needed due to TOML parsing limitations) to use a different logger, in this case qml_debug_stream_alt. We are free to define the module/package log-level here (we opt for DEBUG for all), and to also use multiple handlers per logger (such as for logging to the standard output and files through qml_debug_stream and qml_debug_file simultaneously). Given the complexity explosion with configuring these options, the default features in log_config.toml all use the same log-level, and handler, which can be adjusted based on developer needs.

Logging example with PennyLane and JAX’s JIT support

As mentioned above, we have added execution function entry logging supports, including some supports for each target interface. We can examine this support for both internal and external packages, where we enable logs for JAX, which has support for Python-native log messages. To enable logging specifically for JAX, we can modify the level parameter for the [loggers.jax] entry in the log_config.toml file as:

[loggers.jax]
handlers = ["qml_debug_stream"]
level = "DEBUG"
propagate = false

where handlers represents some arbitrary custom class we define to deal with the message, level the associated level we want that package to log at, and propagate tells the logger to keep the message at the given handler level, or also to throw it up to the parent logger interface — all of these are adhering to the logging API. We convert the highest supported log level from warning (less verbose) to debug (more verbose). We can at the same time change the PennyLane logging level to warnings and more severe, by making the following change:

[loggers.pennylane]
handlers = ["qml_debug_stream"]
level = "WARN"
propagate = false

Running the following example will produce lots of output about the JIT process, and surrounding operations:

import pennylane as qml
import jax, jax.numpy as jnp
from jax import jacfwd, jacrev
import logging

# Enable logging
qml.logging.enable_logging()

# Get logger for use by this script only.
logger = logging.getLogger(__name__)
dev_name = "default.qubit.jax"
num_wires = 2
num_shots = None

@jax.jit
def circuit(key, param):
    logger.info(f"Creating {dev_name} device with {num_wires} wires and {num_shots} shots with {key} PNRG")
    dev = qml.device(dev_name, wires=num_wires, shots=num_shots, prng_key=key)

    @qml.qnode(dev, interface="jax", diff_method="backprop")
    def my_circuit():
        qml.RX(param, wires=0)
        qml.CNOT(wires=[0, 1])
        return qml.expval(qml.PauliZ(0))

    logger.info(f"Created QNODE={my_circuit}")
    res =  my_circuit()
    logger.info(f"Created QNODE evaluation={res}")

    return res

key1 = jax.random.PRNGKey(0)
key2 = jax.random.PRNGKey(1)

logger.info(f"Running circuit with key={key1}")
circuit(key1, jnp.pi/2)
logger.info(f"Running circuit with key={key2}")
circuit(key2, jnp.pi/2)
logger.info(f"Calculating jacobian circuit with key={key1}")
logger.info(f"Jacobian={jacfwd(lambda x: circuit(key1, x))(jnp.pi/3)}")

We can examine the output of the log-statements, which shows debug level messages from JAX, and info-level messages for the given script. To modify the logger defined in the Python script, a new section can be added as:

# Control logging in the executing Python script
[loggers.__main__]
handlers = ["qml_debug_stream",]
level = "INFO"
propagate = false

To see PennyLane-wide debug messages, we can revert the PennyLane log level to debug, and rerun the script. There should be more output than previously observed.

Adding log-statements to the interface execution pipelines

Similarly, for autograd (TF and Torch also), we can run examples that tie-into the execution pipeline for devices without backprop supports:

import pennylane as qml
import logging

qml.logging.enable_logging()

logger = logging.getLogger(__name__)
dev_name = "lightning.qubit"
num_wires = 2
num_shots = None

def circuit(param):
    logger.info(f"Creating {dev_name} device with {num_wires} wires and {num_shots} shots")
    dev = qml.device(dev_name, wires=num_wires, shots=num_shots)

    @qml.qnode(dev, diff_method="adjoint")
    def my_circuit(param):
        qml.RX(param, wires=0)
        qml.CNOT(wires=[0, 1])
        return qml.expval(qml.PauliZ(0))

    logger.info(f"Created QNODE={my_circuit}")
    res =  my_circuit(param)
    logger.info(f"Created QNODE evaluation={res}")

    return res

par = qml.numpy.array([0.1,0.2])

logger.info(f"Running circuit with par={par[0]}")
circuit(par[0])
logger.info(f"Running circuit with par={par[1]}")
circuit(par[1])
logger.info(f"Calculating jacobian circuit with par={par}")
logger.info(f"Jacobian={qml.jacobian(circuit)(par[0])}")

By using lightning.qubit we can now treat the execution environment as a black-box, and see the log-level messages as they hit the custom functions as part of the execution pipeline.

The above features have been added for Torch, Tensorflow, JAX and autograd, and should produce a sufficient level of detail in the execution messages.

Log formatting

The logging-formatter ties-into the ANSI color-code system to improve visibility of standard output and error logging during execution. The ANSI codes accept RGB-coded code to change the text and background colors, allowing messages to be color coded for ease of readability. For example, to generate all such sequences in steps of 5 across each 8-bit range per color, we can use the following bash command:

for r in `seq 0 5 255`; do
    for g in `seq 0 5 255`; do
        for b in `seq 0 5 255`; do
            echo -e "\e[38;2;${r};${g};${b}m"'\\e[38;2;'"${r};${g};${b}"m" FOREGROUND\e[0m"
            echo -e "\e[48;2;${r};${g};${b}m"'\\e[48;2;'"${r};${g};${b}"m" BACKGROUND\e[0m"
        done
    done
done

The strings in the log messages are prepended with the appropriate ANSI codes to ensure different log-levels are highlighted in different ways when outputing to the standard output stream (stdout/stderr). These are defined in the pennylane.logging.formatter module, and can be customized to suit any colors, or messaging structure.

development/guide/logging

 

Download Python script

 

Download Notebook

 

View on GitHub