[mxnet 2.0][item 4.8][RFC] Gluon Data API Extension and Fixes(Part 1)

[zhreshold]

Description

This is the part 1 of Gluon Data API extension and fixes, which mainly focus on cleaning up diverging usage of mxnet module/gluon.
Through long time evolution, there's currently two streams of data loading conventions implemented in mxnet

• Iterator: mxnet.io.DataIter(https://github.com/apache/incubator-mxnet/blob/ac88f1e87aa7da2c33f0cb89524d444ddc3a2ae8/python/mxnet/io/io.py#L180)
• Dataset + DataLoader: gluon.data.Dataset + gluon.data.DataLoader

In order to eliminate the confusion here and to reduce the maintenance efforts, the plan is to drop all old iterators and provide similar Dataset + Dataloader experience in gluon data API.

Things to be removed

iterators

• Base mxnet.io.DataIter(https://github.com/apache/incubator-mxnet/blob/ac88f1e87aa7da2c33f0cb89524d444ddc3a2ae8/python/mxnet/io/io.py#L180)
• mxnet.io.ResizeIter(https://github.com/apache/incubator-mxnet/blob/ac88f1e87aa7da2c33f0cb89524d444ddc3a2ae8/python/mxnet/io/io.py#L282)
• mxnet.io.PrefetchIter(https://github.com/apache/incubator-mxnet/blob/ac88f1e87aa7da2c33f0cb89524d444ddc3a2ae8/python/mxnet/io/io.py#L347)
• mxnet.io.NDArrayIter(https://github.com/apache/incubator-mxnet/blob/ac88f1e87aa7da2c33f0cb89524d444ddc3a2ae8/python/mxnet/io/io.py#L491)
• mxnet.io.MXDataIter(https://github.com/apache/incubator-mxnet/blob/ac88f1e87aa7da2c33f0cb89524d444ddc3a2ae8/python/mxnet/io/io.py#L800)
• mxnet.image.ImageIter(https://github.com/apache/incubator-mxnet/blob/ac88f1e87aa7da2c33f0cb89524d444ddc3a2ae8/python/mxnet/image/detection.py#L626)
• mxnet.image.ImageDetIter(https://github.com/apache/incubator-mxnet/blob/ac88f1e87aa7da2c33f0cb89524d444ddc3a2ae8/python/mxnet/image/detection.py#L626)

Augmenters from mxnet.image and mxnet.image.detection module

Random augmenters, e.g. (https://github.com/apache/incubator-mxnet/blob/ac88f1e87aa7da2c33f0cb89524d444ddc3a2ae8/python/mxnet/image/image.py#L615) will be removed.

Transform = args in gluon.data.Datasets

transform = is no longer supported, and can be replaced with dataset.transform or dataset.transform_first

Things to be added

Gluon Data Datasets

Dataset + Transfrom combo that simulate the removed Iterators

For example, NDArrayIter can be reimplemented as NDArrayDataset + empty transform function.

Gluon Data Augmentaters/Transforms

Data augmenters as mxnet.gluon.Block

Candidates TBD, useful candidates from GluonCV(https://github.com/dmlc/gluon-cv/tree/master/gluoncv/data/transforms) and GluonNLP(https://github.com/dmlc/gluon-nlp/blob/v0.8.x/src/gluonnlp/data/transforms.py)

mxnet.image

image processing functions will be absorbed from GluonCV(https://github.com/dmlc/gluon-cv/blob/master/gluoncv/data/transforms/image.py)

[ptrendx]

What about mx.io.ImageRecordIter? Also, what about the return type of those iterator - mx.io iterators return mx.io.DataBatch, will that be changed too?

@JanuszL FYI since DALI MXNet plugin produces mx.io.DataBatch and may be affected.

[zhreshold]

The old iterators will get a special gluon dataset wrapper which has no length and forbids random accessing or sampling from dataloader, they keep their original arguments during iteration