Sidekick (Beta)

This code is the sidekick to the superhero that is the Peltarion Platform. Sidekick handles the mundane tasks like bundling up data into the Platform's preferred format or sending data examples to the deployment endpoints to get predictions.

Sidekick's aim is to make it easier to

• get data in
• get predictions out

We hope that sidekick will help more people experience the end-to-end flow of a deep learning project and appreciate the value that the Platform provides.

Installation of sidekick

Requirements Sidekick requires python 3.5+.

When installing sidekick we recommend using a separate virtual environment, see e.g. the tutorial Python virtual environments a primer.

Install package and dependencies with pip directly from GitHub:

pip install git+https://github.com/Peltarion/sidekick#egg=sidekick

Get data in - Create a Platform compatible dataset

When creating a dataset zip you can load the data in two separate ways. Both require loading the data in a Pandas DataFrame and assume all columns only contain one type of data with the same shape.

1) Load data in memory objects

Store objects directly in the Series (columns of your DataFrame). This works for all scalars (floats, integers and strings of one dimension) as well as Pillow images and numpy arrays.

Example

This is such an example with a progressbar enabled:

df.head()

float_column          image_column                numpy_column
    0.248851  <PIL.Image.Image ...  [0.18680, 0.61951, 0.83...
    0.523621  <PIL.Image.Image ...  [0.75213, 0.44948, 0.82...
    0.647844  <PIL.Image.Image ...  [0.41525, 0.63858, 0.34...
    0.447717  <PIL.Image.Image ...  [0.79373, 0.24514, 0.94...
    0.194222  <PIL.Image.Image ...  [0.12636, 0.40554, 0.66...

import sidekick

# Create dataset
sidekick.create_dataset(
    'path/to/dataset.zip',
    df,
    progress=True
)

2) Load data in paths to objects

Columns may also point to paths of object. Which columns are paths should be indicated in the path_columns. Like the in-memory version these may also be preprocessed.

Example

This is an example where all images are loaded from a path, preprocessed to have the same shape and type and then placed in the dataset.

df.head()

float_column string_column                                  image_file_column
    0.248851           foo  /var/folders/7t/80jfy0rd3l7f31xdd3rw0_jw0000gn...
    0.523621           foo  /var/folders/7t/80jfy0rd3l7f31xdd3rw0_jw0000gn...
    0.647844           foo  /var/folders/7t/80jfy0rd3l7f31xdd3rw0_jw0000gn...
    0.447717           foo  /var/folders/7t/80jfy0rd3l7f31xdd3rw0_jw0000gn...
    0.194222           foo  /var/folders/7t/80jfy0rd3l7f31xdd3rw0_jw0000gn...

import functools
import sidekick

# Create preprocessor for images, cropping to 32x32 and formatting as png
image_processor = functools.partial(
    sidekick.process_image, mode='center_crop_or_pad', size=(32, 32), format='png')

# Create dataset
sidekick.create_dataset(
    'path/to/dataset.zip',
    df,
    path_columns=['image_file_column'],
    preprocess={
        'image_file_column': image_processor
    }
)

Get predictions out - Use a deployed experiment

To connect to an enabled deployment use the sidekick.Deployment class. This class takes the information you find on the deployment page of an experiment.

Example

This example shows how to query an enabled deployment for image classification.

Use the url and token displayed in the dark box. Then, create a dictionary of the Feature Name and Type (here image and Image (32x32x3)) fields from the table of input and output parameters to specify dtypes_in and dtypes_out.

import sidekick

client = sidekick.Deployment(
    url='<url>',
    token='<token>',
    dtypes_in={'image': 'Image (32x32x3)'},
    dtypes_out={'image': 'Image (32x32x3)'}
)

This deployment client may now be used to get predictions for images

Test deployment with one sample - predict

To predict result of one image (here test.png) use predict.

Example

from PIL import Image

# Load image
image = Image.open('test.png')

# Get predictions from model
client.predict(image=image)

Note: If the feature name is not a valid python variable, e.g., Image.Input, use predict_many instead of predict.

Test deployment with many samples - predict_many

To efficiently predict the results of multiple input samples (here, test1.png, test2.png) use predict_many.

Example

client.predict_many([
    {'image': Image.open('test1.png')},
    {'image': Image.open('test2.png')}
])

Interactive exploration of data - predict_lazy

For interactive exploration of data it is useful to use the predict_lazy method, which returns a generator that lazily polls the deployment when needed. This allows you to immediatly start exploring the results instead of waiting for all predictions to finnish.

Example

client.predict_lazy([
    {'image': Image.open('test1.png')},
    {'image': Image.open('test2.png')}
])

Object type and field type compatibility

Which object types are compatible with each field type may be shown by printing the sidekick.encode.DTYPE_COMPATIBLITY dictionary.

print(sidekick.encode.DTYPE_COMPATIBILITY)

{'Float': {float},
 'Int': {int},
 'Numpy': {numpy.ndarray},
 'Image': {PIL.Image.Image}}

Examples

Examples of how to use sidekick are available at: examples/