Stages: Preprocess, Forward and Postprocess

Each Pipeline has a preprocess, forward and postprocess part. We call those parts stages.

As the names suggest, the different stages are responsible for processing data in the different parts of the deep learning workflow:

• preprocess stands for pre-processing - for example: loading, reshaping and augmenting data

• forward corresponds to the model’s “forward” part - what happens in a PyTorch module, usually on the gpu

• postprocess stands for post-processing - for example converting the output of a model to a readable format

To define stages, use the special Transforms padl.batch and padl.unbatch in a composed Pipeline:

from padl import transform, batch, unbatch
from torchvision import transforms, models

transforms = transform(transforms)
models = transform(models)

@transform
def load_image(path):
    return Image.load()

@transform
def classify(x):
    # [...] lookup the most likely class
    return class

my_classifier_transform = (
    load_image                 # preprocessing ...
    >> transforms.ToTensor()   # 
    >> batch                   # ... stage
    >> models.resnet18()       # forward
    >> unbatch                 # postprocessing ...
    >> classify                # ... stage
)

The different stages of a Pipeline can be accessed via .pd_preprocess, .pd_forward and .pd_postprocess:

>>> my_classifier.pd_preprocess
load_image >> transforms.ToTensor() >> batch
>>> my_classifier.pd_forward
models.resnet18()
>>> my_classifier.pd_postprocess
unbatch >> classify

The Transforms in the preprocess and postprocess stages process the elements one-by-one whereas the Transforms in the forward stage process batches.

Continue in the next section to learn how to apply transforms to data for inference, evaluation and training.