Streaming refacto

[rom1504]

Refactor img2dataset in 3 parts:

• core: provides the core functionalities as functions with no user interface
• batch: looks like the current CLI, can download N millions samples, by splitting in shards and running distributed if needed.
• service: a fastapi API, that can download a shard. It can also queue. It also contains a load balancer to be able to use N workers

The main ideas here are:

• make the core functionality be a simple function/process that takes one shard and process it
• bundle that as batch as usual
• now provides also a service which can provide similar but more flexible functionalities: streaming download, potentially more flexible distribution, ...

The implementation is mostly working but requires refinement before considering merging.

Was implemented some 8 months ago, but I'm thinking to finish it now

[rom1504]

#339 related issue

[rom1504]

one alternative I'm considering (but maybe that can simply live next to batch and service instead of replacing them): implement things as tasks and then use a distributed queue like celery to run things. I don't like making installing a service like reddis part of user requirements though

[rom1504]

maybe ray https://stackoverflow.com/a/54738967 is an alternative to celery

[rom1504]

https://docs.ray.io/en/latest/ray-core/walkthrough.html

[rom1504]

https://github.com/veonua/laion-yt-dlp example for celery

[rom1504]

I think this branch is an interesting POC but it's too many changes

I think a PR that would only extract the core as a standalone function / process would be a good start. It may need dropping the keys based on shard first

[rom1504]

https://github.com/rom1504/img2dataset/blob/streaming_refacto/img2dataset/core/downloader.py#L29-L51 I'm not sure I like the pydantic part

[rom1504]

maybe https://abseil.io/docs/python/quickstart is a good alternative

[rom1504]

https://docs.celeryq.dev/en/stable/userguide/tasks.html celery take on how to configure tasks

[rom1504]

ray's take https://docs.ray.io/en/latest/ray-core/tasks.html

[rom1504]

https://docs.celeryq.dev/en/stable/userguide/calling.html#calling-serializers

[rom1504]

the flags should be much more separated process by process

[rom1504]

https://github.com/rom1504/img2dataset/blob/streaming_refacto/img2dataset/core/resizer.py#L98 gets duplicated compared to https://github.com/rom1504/img2dataset/blob/streaming_refacto/img2dataset/core/resizer.py#L13C7-L13C22 ; how can we improve it

[rom1504]

https://github.com/rom1504/img2dataset/blob/streaming_refacto/img2dataset/core/downloader.py#L133 fiddle/omegaconf might automate this line

[rom1504]

next step here: try to POC a few different ideas of parameter definition instead of pydantic
feels like something simpler is possible

[rom1504]

anything can be generated into a fastapi model at the end if necessary https://github.com/rom1504/img2dataset/blob/streaming_refacto/img2dataset/service/service.py#L37

[rom1504]

maybe check a few projects using ray or celery and see if something helps

[rom1504]

configuration for processes

[rom1504]

https://docs.ray.io/en/latest/ray-core/patterns/pipelining.html

[rom1504]

one idea:

• config for the process (init), include feature names
• input features (image)
• output features (image)

input and output calls are generated based on feature names

so have a generic config for processes

then apply that here but also any2dataset/video2dataset

[rom1504]

try to write that out as json maybe

so all actual config here would be one json per process

we would also have support top level for the flat args at top level for legacy reasons

[rom1504]

write it out and see how that plays out for all the projects

[rom1504]

being able to configure these processes properly here should make all the wrapper around it (service, batch, distributors) come along naturally

[rom1504]

https://developers.google.com/mediapipe/api/solutions/python/mp/calculators/core/flow_limiter_calculator_pb2/FlowLimiterCalculatorOptions

[rom1504]

https://github.com/google/mediapipe/blob/master/mediapipe/calculators/tensorflow/matrix_to_tensor_calculator_options.proto

[rom1504]

https://github.com/google/mediapipe/blob/93290388178395730c9bed0be0042836cf710465/mediapipe/framework/calculator_context.h#L37-L39

[rom1504]

https://developers.google.com/mediapipe/framework/framework_concepts/calculators.md

[rom1504]

https://beam.apache.org/documentation/programming-guide/#pardo

[rom1504]

https://spark.apache.org/docs/3.1.3/api/python/reference/api/pyspark.sql.functions.udf.html

[rom1504]

resizer options:

        image_size,
        resize_mode,
        resize_only_if_bigger,
        upscale_interpolation="lanczos",
        downscale_interpolation="area",
        encode_quality=95,
        encode_format="jpg",
        skip_reencode=False,
        disable_all_reencoding=False,
        min_image_size=0,
        max_image_area=float("inf"),
        max_aspect_ratio=float("inf"),
        blurrer=None, <- this should be replaced by blurrer config or moved out ?

img_stream, blurring_bbox_list=None <- input feature
img_str, width, height, original_width, original_height, None <- output feature

this is a process that applies to one image

json for this?

[rom1504]

maybe a typed thing is better here?

• proto
• pydantic
• what else ?

[rom1504]

https://github.com/rom1504/img2dataset/tree/main/img2dataset

[rom1504]

https://github.com/iejMac/video2dataset/tree/main/video2dataset/subsamplers

[rom1504]

write down some of these and some potential schema and find what sticks

[rom1504]

https://scanner-research.github.io/guide/graphs.html#graphs

[rom1504]

https://github.com/videoflow/videoflow/tree/master

[rom1504]

https://www.anyscale.com/blog/streaming-distributed-execution-across-cpus-and-gpus