Feature request: download all features but only load only part of DGS Corpus at a time?

[cleong110]

trafficstars

When attempting to load the DGS Corpus's default configuration on either my own workstation or in Colab, I run out of memory and crash.

Here are some screenshots

https://colab.research.google.com/drive/1_vWFvWo0ZMg5_6AFU6Ln2LPHwm9TW_Rz?usp=sharing for example, will crash given time. Is there a method such that all the features, video and pose and gloss, can all be downloaded, yet only some portion loaded into memory at a time?

[cleong110]

Possibly something like

dgs_corpus = tfds.load('dgs_corpus', split=["train:2%"])

would work

[cleong110]

If that works, I wonder if it would be good to: a. warn the user of projected memory usage somehow when they run the .load command? b. change the default load to not load the entire dataset into memory?

[cleong110]

you can do it, DGS Corpus! I believe in you!

[cleong110]

[cleong110]

[cleong110]

sigh

[cleong110]

Giving it a try on my personal workstation

[cleong110]

...nope, still "Killed"

[cleong110]

All of these crash on my workstation, using up all 33 GB:

# dgs_corpus = tfds.load('dgs_corpus') # Killed
# dgs_corpus = tfds.load('dgs_corpus', split=["train:2%"]) # Killed
# dgs_corpus = tfds.load('dgs_corpus', split=["train:100"]) # Killed
# dgs_corpus = tfds.load('dgs_corpus', split=["train:10"]) # still Killed

[cleong110]

Maybe one of these tricks can work?

https://www.tensorflow.org/guide/data_performance#reducing_memory_footprint

[cleong110]

Maybe a custom data generator? https://medium.com/analytics-vidhya/write-your-own-custom-data-generator-for-tensorflow-keras-1252b64e41c3

[cleong110]

https://www.tensorflow.org/datasets/performances#large_datasets

[cleong110]

Maybe something from here? https://github.com/tensorflow/tfjs/issues/7801

[cleong110]

Oh hey, this looks relevant, and I see a familiar name: https://github.com/huggingface/datasets/issues/741. It's transformers library though.

[cleong110]

Reading https://www.tensorflow.org/datasets/api_docs/python/tfds/load, maybe we can do as_dataset separately?

[cleong110]

https://www.reddit.com/r/deeplearning/comments/z8otan/if_the_dataset_is_too_big_to_fit_into_your_ram/ some more possibilities

[cleong110]

OK, so tf.Dataset does support streaming... https://stackoverflow.com/questions/63140320/how-to-use-sequence-generator-on-tf-data-dataset-object-to-fit-partial-data-into, so is it the split generation where the issue comes?

[cleong110]

Using the "manually add print statements to the site-packages in my conda env" method I kept following it all the way down and it gets killed in here:

https://github.com/sign-language-processing/datasets/blob/master/sign_language_datasets/datasets/dgs_corpus/dgs_corpus.py#L330,

and makes it to here, https://github.com/tensorflow/datasets/blob/v4.9.3/tensorflow_datasets/core/dataset_builder.py#L1584

and makes it to here

https://github.com/tensorflow/datasets/blob/v4.9.3/tensorflow_datasets/core/split_builder.py#L415

and gets killed around there somewhere

[cleong110]

https://github.com/gruns/icecream might be helpful, note to self

[cleong110]

Or, you know, I could look at one of these: https://stackify.com/top-5-python-memory-profilers/

[cleong110]

Or https://www.tensorflow.org/guide/profiler#memory_profile_tool

[cleong110]

https://colab.research.google.com/github/jakevdp/PythonDataScienceHandbook/blob/master/notebooks/01.07-Timing-and-Profiling.ipynb could also be of use

[cleong110]

I've done a lot of searching but tfds just doesn't seem to have a way to stream part of a large dataset that I can find.

[cleong110]

So I still can't figure out how to (1) only download a portion, (2) assuming it's all successfully downloaded, load only a portion into memory without the split generation using all available memory.

[AmitMY]

lots of comments... in the future it would be helpful if you keep editing the same comment or handful.

When you:

tfds.load('dgs_corpus', split=["train:2%"])

What happens is that first the entire dataset is being prepared, then only 2% of it is loaded. So you will need the exact same disk space.

Now since there are two processes here:

1. preparing the entire dataset
2. loading a part of the dataset

can you tell where the memory consumption is too high? my suspicion is number 1, but i don't know.

[cleong110]

Right, sorry, I forget that I'm not the only one getting spammed by all these, apologies.

[cleong110]

I'm also suspecting 1, based on the fact that I can sprinkle print statements all the way until https://github.com/tensorflow/datasets/blob/v4.9.3/tensorflow_datasets/core/split_builder.py#L415.

Edit: my big issue is that testing it currently requires me to run it until it crashes. Which, if I'm doing it on Google Colab, means that any modifications I've made to the code are then gone. I've got a workstation I can test locally on but don't have access as conveniently.

Edit again: Certainly the download_and_prepare is using a lot of RAM, though this particular instance of Colab has not yet crashed:

Edit 3: ... and it crashed after using all available RAM. so that step does seem to use high memory... but the file system was not actually deleted. OK I can work with that, perhaps

Edit 4: OK, trying it in a "high-memory" notebook in Colab Pro, I get this:

Edit 5: full stacktrace on the high-memory notebook:

ValueError                                Traceback (most recent call last)

[<ipython-input-5-6bad8ee20d7b>](https://localhost:8080/#) in <cell line: 1>()
----> 1 dgs_corpus = tfds.load('dgs_corpus', )

10 frames

[/usr/local/lib/python3.10/dist-packages/tensorflow_datasets/core/logging/__init__.py](https://localhost:8080/#) in __call__(self, function, instance, args, kwargs)
    166     metadata = self._start_call()
    167     try:
--> 168       return function(*args, **kwargs)
    169     except Exception:
    170       metadata.mark_error()

[/usr/local/lib/python3.10/dist-packages/tensorflow_datasets/core/load.py](https://localhost:8080/#) in load(name, split, data_dir, batch_size, shuffle_files, download, as_supervised, decoders, read_config, with_info, builder_kwargs, download_and_prepare_kwargs, as_dataset_kwargs, try_gcs)
    647       try_gcs,
    648   )
--> 649   _download_and_prepare_builder(dbuilder, download, download_and_prepare_kwargs)
    650 
    651   if as_dataset_kwargs is None:

[/usr/local/lib/python3.10/dist-packages/tensorflow_datasets/core/load.py](https://localhost:8080/#) in _download_and_prepare_builder(dbuilder, download, download_and_prepare_kwargs)
    506   if download:
    507     download_and_prepare_kwargs = download_and_prepare_kwargs or {}
--> 508     dbuilder.download_and_prepare(**download_and_prepare_kwargs)
    509 
    510 

[/usr/local/lib/python3.10/dist-packages/tensorflow_datasets/core/logging/__init__.py](https://localhost:8080/#) in __call__(self, function, instance, args, kwargs)
    166     metadata = self._start_call()
    167     try:
--> 168       return function(*args, **kwargs)
    169     except Exception:
    170       metadata.mark_error()

[/usr/local/lib/python3.10/dist-packages/tensorflow_datasets/core/dataset_builder.py](https://localhost:8080/#) in download_and_prepare(self, download_dir, download_config, file_format)
    697           self.info.read_from_directory(self.data_dir)
    698         else:
--> 699           self._download_and_prepare(
    700               dl_manager=dl_manager,
    701               download_config=download_config,

[/usr/local/lib/python3.10/dist-packages/tensorflow_datasets/core/dataset_builder.py](https://localhost:8080/#) in _download_and_prepare(self, dl_manager, download_config)
   1666       return
   1667 
-> 1668     split_infos = self._generate_splits(dl_manager, download_config)
   1669 
   1670     # Update the info object with the splits.

[/usr/local/lib/python3.10/dist-packages/tensorflow_datasets/core/dataset_builder.py](https://localhost:8080/#) in _generate_splits(self, dl_manager, download_config)
   1641       ):
   1642         filename_template = self._get_filename_template(split_name=split_name)
-> 1643         future = split_builder.submit_split_generation(
   1644             split_name=split_name,
   1645             generator=generator,

[/usr/local/lib/python3.10/dist-packages/tensorflow_datasets/core/split_builder.py](https://localhost:8080/#) in submit_split_generation(self, split_name, generator, filename_template, disable_shuffling)
    329     # `_build_from_xyz` method.
    330     if isinstance(generator, collections.abc.Iterable):
--> 331       return self._build_from_generator(**build_kwargs)
    332     else:  # Otherwise, beam required
    333       unknown_generator_type = TypeError(

[/usr/local/lib/python3.10/dist-packages/tensorflow_datasets/core/split_builder.py](https://localhost:8080/#) in _build_from_generator(self, split_name, generator, filename_template, disable_shuffling)
    400       except Exception as e:  # pylint: disable=broad-except
    401         utils.reraise(e, prefix=f'Failed to encode example:\n{example}\n')
--> 402       writer.write(key, example)
    403     shard_lengths, total_size = writer.finalize()
    404 

[/usr/local/lib/python3.10/dist-packages/tensorflow_datasets/core/writer.py](https://localhost:8080/#) in write(self, key, example)
    225       example: the Example to write to the shard.
    226     """
--> 227     serialized_example = self._serializer.serialize_example(example=example)
    228     self._shuffler.add(key, serialized_example)
    229     self._num_examples += 1

[/usr/local/lib/python3.10/dist-packages/tensorflow_datasets/core/example_serializer.py](https://localhost:8080/#) in serialize_example(self, example)
     96       serialize_proto: `str`, the serialized `tf.train.Example` proto
     97     """
---> 98     return self.get_tf_example(example).SerializeToString()
     99 
    100

Edit: and here's how many resources were used:

Update: OK, it seems that just these three files being encoded is enough to use many gigabytes.

[cleong110]

Well I am thoroughly stumped. I've narrowed it down to where in tfds the massive memory allocations are happening, but I still don't know why.

I just don't understand why it needs nearly 30 GiB to "encode" and then "serialize" the videos

Here's the memray report. memray_output_file.tar.gz

For some reason reading in the frames results in over 50k allocations? many GiB worth? This line right here, officer: https://github.com/google/etils/blob/main/etils/epath/abstract_path.py#L149 Called by this one https://github.com/tensorflow/datasets/blob/master/tensorflow_datasets/core/features/video_feature.py#L152

Also the serialize_example calls use huge memory as well

I don't know what to do or how to fix it. I know the UFC101 dataset doesn't have this issue. If anyone has thoughts let me know

This is all before we even get to the protobuf max size error

Update: well, I inspected the tmp folder that gets created, and there are indeed nearly 13k frames extracted from just one of the videos:

which ends up being nearly 5GB legitimately:

Note: perhaps something in here may be relevant?

• https://github.com/sign-language-processing/datasets/blob/master/sign_language_datasets/datasets/config.py#L61
• and https://www.tensorflow.org/datasets/api_docs/python/tfds/features/Video#example

Like maybe there's a setting in there to not load every file but just a list of paths?

[AmitMY]

Ok, so to me it seems like you are not using appropriate config

Like maybe there's a setting in there to not load every file but just a list of paths?

The "correct" config here would be:

config = DgsCorpusConfig(name="only-annotations", version="1.0.0", include_video=False, include_pose=None)
dgs_corpus = tfds.load('dgs_corpus', builder_kwargs=dict(config=config))

Which loads only the annotations.

You want to download the video but load them as paths? include_video=True, process_video=False

You want to load poses? include_pose="holistic" or include_pose="openpose"