13. How to use Veles. MNIST example. Simple (user entry/medium level)

13.1. Preprocessing

For preprocessing of the data you should use Loaders (descendants of veles.loader.base.Loader).

Available Loaders can be found in Units. Parameters of the units. Some of Loaders could be used right away.

Each Loader, which can be used without deriving from it, must have MAPPING - unique loader’s name.

There are few types of Loaders. If Loader was derived from veles.loader.fullbatch.FullBatchLoader, data will be stored entire in memory. If Loader was derived from veles.loader.base.LoaderMSE, Loader has not only labels (string, int or double values), but also targets (matrixes, vectors). If Loader was derived from veles.loader.image.ImageLoader, functionality for preprocessing images could be used (scale, crop, add sobel, rotate, change color space, mirror, etc.)

Learn more: How to use Veles. MNIST example. Advanced (user medium/high level)

13.2. Training

There are 4 ways to create train Workflow:

1. Use existing Snapshot of Workflow and continue training.
2. Use existing Workflow and existing Configuration file.
3. Use existing Workflow and change Configuration file.
4. Create custom Workflow and Configuration file. (Advanced)

13.2.1. Use existing Snapshot

Use your Snapshot to continue training if training was interrupted.

Or download existing Snapshot from Amazon (paths to snapshots of the Model is in “snapshots” field in manifest.json)

Here is manifest.json of MNIST:

{
...
"snapshots":
["https://s3-eu-west-1.amazonaws.com/veles.forge/MNIST/mnist_validation_1.92_train_0.04.4.pickle.gz",
"https://s3-eu-west-1.amazonaws.com/veles.forge/MNIST/mnist_caffe_validation_0.86_train_0.23.4.pickle",
"https://s3-eu-west-1.amazonaws.com/veles.forge/MNIST/mnist_conv_validation_0.73_train_0.11.4.pickle"]
}

Or use path to the amazon snapshot as command line argument. For fully-connected MNISTWorkflow:

python3 -m veles -s -d 0 -w=https://s3-eu-west-1.amazonaws.com/veles.forge/MNIST/mnist_validation_1.92_train_0.04.4.pickle.gz veles/znicz/samples/MNIST/mnist.py -

For convolutional MNISTWorkflow:

python3 -m veles -s -d 0 -w=https://s3-eu-west-1.amazonaws.com/veles.forge/MNIST/mnist_conv_validation_0.73_train_0.11.4.pickle veles/znicz/samples/MNIST/mnist.py veles/znicz/samples/MNIST/mnist_conv_config.py

13.2.2. Use existing Workflow and existing Configuration file

Use existing Workflows from samples or download them from VelesForge.

To run veles from command line you need to set path to workflow and path to configuration files as arguments:

python3 -m veles path_to_workflow path_to_config

See Command line examples for command line examples. To run the MNIST workflow from command line for fully-connected Workflow:

python3 -m veles -s -d 0 veles/znicz/samples/MNIST/mnist.py -

For convolutional Workflow:

python3 -m veles -s -d 0 veles/znicz/samples/MNIST/mnist.py veles/znicz/samples/MNIST/mnist_conv_config.py

For convolutional Workflow with Caffe configuration:

python3 -m veles -s -d 0 veles/znicz/samples/MNIST/mnist.py veles/znicz/samples/MNIST/mnist_caffe_config.py

If Veles was installed for 1-2 users levels, set PYTHONPATH=”/usr/lib/python3/dist-packages” or use absolute paths to the Workflow and Configuration files:

python3 -m veles -s -d 0 /usr/lib/python3/dist-packages/veles/znicz/samples/MNIST/mnist.py -

Or copy samples from /usr/lib/python3/dist-packages/veles/znicz/samples to your local directory.

13.2.3. Use existing Workflow and change Configuration file

About configuration: Configuration

First copy samples to a local directory:

cp -r /usr/lib/python3/dist-packages/veles/znicz/samples /home/user/

or download Model from VelesForge.

There 2 ways to change configuration parameters:

1. In configuration file
2. From command line

To use the first way, open file /home/user/samples/MNIST/mnist_config.py and change it. To use the second way change necessary parameters right from the command line:

python3 -m veles -s -d 0 /usr/lib/python3/dist-packages/veles/znicz/samples/MNIST/mnist.py - root.mnistr.loader.minibatch_size=10 root.mnistr.loader.data_path=\"/path/to/new/dataset\"

MNIST workflow (veles.znicz.samples.mnist.MnistWorkflow) was derived from StandardWorkflow (veles.znicz.standard_workflow.StandardWorkflow).

See parameters of StandardWorkflow and how to work with it here: Parameters of the StandardWorkflow

Learn more: How to use Veles. MNIST example. Advanced (user medium/high level)

13.3. Testing

Snapshot of trained Workflow is required to run Workflow in testing mode. There are 5 ways to create test Workflow:

1. Use existing test Workflow or script
2. Use –test and –result-file arguments in command line
3. Use testing mode and write_results function (Advanced)
4. Create Workflow with extract_forward function (Advanced)
5. Create custom test Workflow (Advanced)

13.3.1. Use existing test Workflow or script

If test Workflow is exists it is located in the directory with train Workflow. See samples or download Model from VelesForge.

13.3.2. Use –test and –result-file arguments in command line

If Loader of trained Workflow has filled test set, run Workflow in testing mode with –test command line argument. Use –result-file argument to save the results of testing:

python3 -m veles -s -d 0 --test --result-file="/home/user/mnist_result.txt" /home/user/samples/MNIST/mnist.py -

File with results will be constructed from results of get_metric_values and get_metric_names functions of Units (IResultProvider must be implemented). Example:

@implementer(IResultProvider, ...)
class EvaluatorBase(...):
    ...
    def get_metric_names(self):
        ...

    def get_metric_values(self):
        ...

Learn more: How to use Veles. MNIST example. Advanced (user medium/high level)

13.4. How to run

13.4.1. Run with ipython notebook

Veles is usable from IPython or IPython Notebook. Open ipython notebook, import veles and run it:

import veles
launcher = veles(
    "veles/znicz/samples/MnistSimple/mnist.py", stealth=True,
    matplotlib_backend="WebAgg")

Arguments are the same as for the command line, but “-” symbol changes to “_” symbol and using of long form options is required.

To pause the process of execution:

launcher.pause()

To resume the process of execution:

launcher.resume()

To stop the process of execution:

launcher.stop()

To initialize the Workflow:

launcher.initialize()

To run the Workflow:

launcher.run()

To initialize and run the Workflow:

launcher.boot()

To get the Workflow:

launcher.workflow

To get Units:

launcher.workflow.units

To get specific Unit:

1. Get by name:

loader = launcher.workflow["MnistLoader"]

2. Get by the instance of Unit:

loader = launcher.workflow.loader

3. Get from Units list

launcher.workflow.units

[veles.plumbing.StartPoint “Start of MnistWorkflow”, veles.plumbing.EndPoint “End of MnistWorkflow”, <veles.plumbing.Repeater object at 0x7f8fc4f1def0>, <MnistSimple.loader_mnist.MnistLoader object at 0x7f8ff17c20f0>, ...]

loader = launcher.workflow.units[3]

13.4.2. Run from command line

See Command line examples.

13.4.3. Frontend option

Use frontend option for the interactive display of Veles options and the command line. Run in the terminal:

python3 -m veles --frontend

Compose the command line and click run button.

13.4.4. Manhole option

Use manhole option to run interactive mode at any time. Run:

python3 -m veles --manhole /home/user/samples/MNIST/mnist.py -

You will see something like this:

MANHOLE:Manhole UDS path: nc -U /tmp/manhole-7355
MANHOLE:Waiting for a new connection (in pid 7355) ...

To switch to the interactive console open new terminal and run the command with a Manhole UDS path:

nc -U /tmp/manhole-7355

You will see:

VELES interactive console
Type in 'workflow' or 'units' to start
veles [1]>

Change some attributes. For example, decrease learning rate in backward propagation units (gds) in 10 times:

veles [1]> for gd in workflow.gds:
      ...:     gd.learning_rate/=10
      ...:

veles [2]>

To stop interactive mode and continue execution type “exit()”:

veles [2]> exit()

13.4.5. Distributed training

See Distributed Training: Master/slave.

13.4.6. Training ensembles

See Training ensembles.

13.5. Optimization parameters

from veles.config import root
from veles.genetics import Range

root.mnistr.update({
    ...
    "loader": {"minibatch_size": Range(20, 1, 1000),
               "normalization_type": "linear",
               "data_path": "/path/to/dataset"},
    ...})

To optimize parameters of Workflow by Genetic Algorithm use Range (veles.genetics.config.Range) for every parameter, which you want to optimize. When optimization is off, the first parameter will be used by default. In MNIST example minibatch size will be equal 20. If optimization is on, the second and the third parameter will be used as range to optimize. In MNIST example minibatch size will be selected from 1 to 1000 by Genetic Algorithm.

See Automatic optimization of workflow parameters.

13.6. Export of the Model

To export Model as package use veles.workflow.package_export(). Set path to the exported package by package_name argument. precision is an optional parameter.

...
class MnistWorkflow(StandardWorkflow):
    def __init__(self, workflow, **kwargs):
        super(MnistWorkflow, self).__init__(workflow, **kwargs)
        self.export_wf = kwargs.get("export_wf", False)
        self.package_name = kwargs.get(
            "package_name", os.path.join(root.common.dirs.user, "mnist.zip"))

    ...

    def on_workflow_finished(self):
        super(MnistWorkflow, self).on_workflow_finished()
        if self.export_wf:
            self.package_export(self.package_name, precision=16)

def run(load, main):
    load(MnistWorkflow,
         ...)
    main()

13.7. Using plotters

To disable plotters during Workflow run:

python3 -m veles -p '' /home/user/samples/MNIST/mnist.py -

To choose WebAgg backend:

python3 -m veles -p 'WebAgg' /home/user/samples/MNIST/mnist.py -

To choose Qt4Agg backend:

python3 -m veles -p 'Qt4Agg' /home/user/samples/MNIST/mnist.py -

To disable plotting service:

python3 -m veles /home/user/samples/MNIST/mnist.py - root.common.disable.plotting=True

13.8. Using publisher

See Publishing reports.