CDECatapult / mlpredict

Python package to predict deep learning execution time

mlpredict

A python package to predict the execution time for one forward and backward pass a deep learning model.

To install mlpredict run

pip install -r requirements.txt

and

python setup.py install

from the root directory.

The mlpredict API can be used to create representations of deep neural networks and predict their execution time on different hardware.

Create a model representations

To build the representation of a deep neural network from scratch create an instance of the dnn class

dnn_repr = mlpredict.api.new_dnn(input_dimension,input_size)

and add layers

dnn_repr.add_layer(layer_type, layer_name, **kwargs)

The last layer can be removed

dnn_repr.remove_last_layer()

and the current network architecture can be displayed

dnn_repr.describe()

A completed model can be saved to a .json file using

dnn_repr.save(filename)

For a full working example see the jupyter notebook https://github.com/CDECatapult/mlpredict/blob/master/notebooks/Create_new_dnn.ipynb.

Import existing model representations

Exiting model representations can be imported using

dnn_repr = mlpredict.api.import_default(filename)

An imported representation can be modified and saved as described above in Create a model representations.

Predict execution time using mlpredict

A model representation dnn_repr created or imported through the mlpredict API can be used to predict the execution time of the corresponding tensorflow model on arbitrary GPUs.

time_total, layer, time_layer = dnn_repr.predict(gpu,
                                                 optimizer,
                                                 batchsize)

returns the total execution time, the layers and the time per layer. For a complete working example see https://github.com/CDECatapult/mlpredict/blob/master/notebooks/Full_model_prediction.ipynb