[GSK-2310, GSK-2537] Create a dataset300wlp using tensorflow datasets

[Inokinoki]

Add an optional DataLoader300WLP dataloader

[linear]

GSK-2310 Create a Dataset300WLP using TensorFlow datasets

linked to: GSK-2251

https://www.tensorflow.org/datasets/catalog/the300w_lp

import tensorflow_datasets as tfds

splits = tfds.load("the300w_lp")
ds = splits["train"]
datarows = ds.take(1)
for r in datarows:
  print(r["landmarks_2d"])
  print(r["image"])

Image format:shape=(1, 450, 450, 3), dtype=uint8

Landmark format:shape=(1, 68, 2), dtype=float32

[rabah-khalek]

No need to put too much time on it but is there a way to pick an image from a specific dataset? The tensorflow page mention:

With 300W, 300W-LP adopt the proposed face profiling to generate 61,225 samples across large poses (1,786 from IBUG, 5,207 from AFW, 16,556 from LFPW and 37,676 from HELEN, XM2VTS is not used).

But probably the combine everything... it's not really a bottleneck, but would be nice.

[Inokinoki]

Loading is composed by the following steps:

builder = tfds.builder('the300w_lp')
# 1. Create the tfrecord files (no-op if already exists)
builder.download_and_prepare()
# 2. Load the `tf.data.Dataset`
ds = builder.as_dataset(split='train', shuffle_files=True)

The file names are already dropped at step 1. Instead, the JPEG image data are integrated directly into a group of tfrecords binary files.

I tried to get the file name from the builder:

>>> g = builder._generate_examples("/Users/inoki/Downloads/tensorflow_datasets/downloads/extracted/ZIP.ucexport_download_id_0B7OEHD3T4eCkVGs0TkhUWFN6JQw2bEF61I9yUitin_g9uysqV5RYA61KUwppa7axPuc/300W_LP/")
<generator object Builder._generate_examples at 0x2810035a0>
>>> next(g)
('IBUG_image_100_03_1.jpg', {'image': '/Users/inoki/Downloads/tensorflow_datasets/downloads/extracted/ZIP.ucexport_download_id_0B7OEHD3T4eCkVGs0TkhUWFN6JQw2bEF61I9yUitin_g9uysqV5RYA61KUwppa7axPuc/300W_LP/IBUG/IBUG_image_100_03_1.jpg', 'landmarks_origin': array(...

It works but there is not an easy way to integrate with tfdsoperations.

I think that we can write a customized loader, to directly build the dataset from their ZIP file, with dataset name and flip as metadata. That way, we can drop dependency requirements of tensorflow and tfds. WDYT? @rabah-khalek

Their zip file url is hardcoded in the source code:

_DATASET_URL = "https://drive.google.com/uc?export=download&id=0B7OEHD3T4eCkVGs0TkhUWFN6N1k"

[rabah-khalek]

We recently upgraded the property index_sampler to an abstract one. We need to know the len of the data, do you think it's possible to retrieve this information?

[Inokinoki]

It's working on M2 Mac!

[linear]

GSK-2537 Fixing tensorflow-io-gcs-filesystem compatibility

[rabah-khalek]

great work @Inokinoki !