API documentation
GATorch principal class is GA.
GATorch.GA
- class GATorch.GA[source]
- attach_model(model, loss=None, named_layer=True, profiler='pyjoules', disable_measurements=[])[source]
Attach a model to the GA profiler. You can add also a loss function. The profiler will then track the energy consumption of the attached model.
- Parameters
model (torch.nn.module) – PyTorch model you want to track.
loss (torch.nn.module) – Loss function to be passed for tracking the consumption of the loss computation.
named_layers (boolean) – Set to False to not track individual named layers.
profiler (string) – Typed of profiler used to track the energy consumption. Currently only ‘pyjoules’ is implemented.
disable_measurements (list) – Set which hardware components you are not intrested in tracking between cpu, ram and gpu.
- get_full_measurements()[source]
Get the full measurements collected by the profiler.
- Returns
The energy consumptions for each pass and model component.
- Return type
dict
- get_mean_measurements()[source]
Get the mean of the measurements collected by the profiler.
- Returns
The mean energy consumption for each model component.
- Return type
dict
- get_sum_measurements()[source]
Get the sum of the measurements collected by the profiler.
- Returns
The full energy consumption for each model component.
- Return type
dict
- to_pandas()[source]
Convert the energy measurements into a pandas.DataFrame object.
- Returns
The dataframe of the energy consumption for each model component.
- Return type
pandas.DataFrame
- to_csv(filename)[source]
Save the energy measurements into a csv file.
- Parameters
filename (string) – Name of the csv file.
- visualize_data(layers='all', complete_model=True, loss=False, phase='total', kind='line', smoothing=0.3, figsize=None, filename=None)[source]
Generate a matplotlib plot for the energy measurements.
- Parameters
layers (list or string) – Pass in a list which named layers you want to display. Pass ‘all’ if you want to see all of them.
complete_model (boolean) – Set to False if you don’t want to see the data for the complete model.
loss (boolean) – Set to True to display also the loss function data.
phase (string) – Select which phase to display between ‘total’, ‘forward’ or ‘backward’.
kind (string) – Select which type of plot to generate between ‘line’, ‘violin’ or ‘box’.
smoothing (float) – Only used for lineplots. Value between 0 and 1 used to add smoothing to the displayed data.
figsize (Tuple) – Size of the figure generated.
filename (string) – Pass a filename if you want to save the image created.
- Returns
The matplotlib axes containing the plot.
- Return type
Axes
- start_tracker_emissions(save_to_file=False)[source]
Start a tracker for carbon emission. Implemented using codecarbon.
- Parameters
save_to_file (boolean) – Save the tracking results in a file called ‘emissions’.
- stop_tracker_emissions()[source]
Stop the tracker for carbon emission.
- Returns
The total emissions produced by the model since the tracker was started.
- Return type
float
- set_tensorboard_stats(writer=None, experiment=0, named_layer=True, sample_size=500)[source]
Sets the tensorboard data that can be viewed through the tensorboard dashboard.
- Parameters
writer (torch.utils.tensorboard.SummaryWriter) – Optionally pass an already existing tensorboard writer.
experiment (int) – Set an identifier for the displayed data.
named_layer (boolean) – Indicate if the named layers should also be displayed in tensorboard.
sample_size (int) – Size of the sample size used to define a loss step. Used to compute the grap for energy per loss step.