Energy draw callback

talos.callbacks.PowerDraw records sampled GPU power draw in watts at epoch begin and end. Import it from talos.callbacks; talos.utils.power_draw_append adds the summary to a training History object.

Install the matching Keras or TensorFlow backend. Physical measurements additionally require supported NVIDIA hardware and nvidia-smi; an explicit provider makes callback behavior testable on a CPU. Endpoint averaging estimates watt-seconds rather than measuring energy continuously. The callback allows:

• record and analyze model energy consumption
• optimize towards energy efficient models

Example

Examples below use the held-out Iris setup in Scan → Minimal Example. Run that setup first; it defines scan_object, p, input_model, x, y, x_val, y_val, x_test, and y_test.

Before model.fit() in the input model:

from talos.callbacks import PowerDraw
# A provider fixture makes this CPU example runnable; these are declared watts.
power_draw = PowerDraw(provider=lambda: 10.0)

Then use power_draw as you would callbacks in general:

params = {key: values[0] for key, values in p.items()}
_, model = input_model(x, y, x_val, y_val, params)
history = model.fit(x, y, validation_data=(x_val, y_val), epochs=2, verbose=0,
                    callbacks=[power_draw])

To get the energy draw data into the experiment log:

history = talos.utils.power_draw_append(history, power_draw)
assert history.history['watts_min'] == [10.0]
assert history.history['Ws'][0] >= 0

NOTE: this line has to be after model.fit().

For physical sampling omit provider and select the GPU with PowerDraw(device=0). The fixture above verifies callback/log behavior only; it is not a measured hardware result.

Interface and units

The callback signature is PowerDraw(device=0, backend='keras', provider=None). device sets the -i device selector for nvidia-smi; backend selects the callback base. A provider is a zero-argument callable returning a numeric power reading in watts. Without a provider, each endpoint launches an nvidia-smi subprocess for that device.

.log contains epoch_begin and epoch_end lists in watts and a seconds list of monotonic elapsed epoch durations. Calling power_draw_append(history, power_draw) modifies history.history in place and returns that same History object. It adds one-element lists for the entire fit:

Field	Unit and calculation
watts_min	Watts; minimum of all endpoint readings.
watts_max	Watts; maximum of all endpoint readings.
seconds	Seconds; sum of recorded epoch durations.
Ws	Watt-seconds, equivalent to joules; sum of each epoch's mean endpoint power times its duration, rounded to two decimal places.

This estimate covers the selected GPU's epoch intervals. It does not include total machine power, dataset preparation, idle intervals or continuous power integration. The example's fixed provider proves the expected fields and nonnegative duration; it is not evidence of hardware energy use.

Failure boundaries

Append only after at least one completed epoch, with paired begin/end readings; an empty log has no minimum or maximum. Missing nvidia-smi, unavailable devices, subprocess failures and nonnumeric provider output propagate as errors. Native Torch loops must invoke the callback hooks explicitly with backend='torch'; Talos does not attach Keras callbacks to a Torch training loop automatically.

Read next

See monitoring for epoch logs and plots, and Scan to include these history summaries in a parameter sweep.