Skip to content

New samplers benchmark #248

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Oct 8, 2024
Merged

New samplers benchmark #248

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
267 changes: 46 additions & 221 deletions benchmarks/samplers/benchmark_samplers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,227 +1,52 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
from pathlib import Path
from time import perf_counter_ns

import abc
import argparse
import importlib
import os

import decord
import numpy as np
import torch

import torch.utils.benchmark as benchmark
from torchcodec.samplers import (
IndexBasedSamplerArgs,
TimeBasedSamplerArgs,
VideoArgs,
VideoClipSampler,
)
from torchmultimodal.fb.utils.video_utils import (
ClipSamplerType,
VideoClipSampler as tmm_vcs,
)
from torchvision.datasets.video_clip_sampler import ( # @manual=//pytorch/vision:internal_datasets
TVVideoClipDecoder,
UniformClipSamplingStrategy,
VideoClipSampler as ta_vcs,
)


class AbstractSampler:
def __init__(self):
pass

@abc.abstractmethod
def sample_frames_uniformly(self, video_file, clips_per_video):
pass


class TorchCodecTimeBasedSampler(AbstractSampler):
def __init__(self):
pass

def sample_frames_uniformly(self, video_file, clips_per_video):
arr = np.fromfile(video_file, dtype=np.uint8)
video_tensor = torch.from_numpy(arr)
video_input = VideoArgs()
sampler_input = TimeBasedSamplerArgs(
sampler_type="uniform", clips_per_video=clips_per_video, frames_per_clip=1
)
sampler = VideoClipSampler(video_input, sampler_input)
return sampler(video_tensor)


class TorchCodecIndexBasedSampler(AbstractSampler):
def __init__(self):
pass

def sample_frames_uniformly(self, video_file, clips_per_video):
arr = np.fromfile(video_file, dtype=np.uint8)
video_tensor = torch.from_numpy(arr)
video_input = VideoArgs()
sampler_input = IndexBasedSamplerArgs(
sampler_type="uniform", clips_per_video=clips_per_video, frames_per_clip=1
)
sampler = VideoClipSampler(video_input, sampler_input)
return sampler(video_tensor)


class TorchCodecIndexBasedSamplerWithStackedOutput(AbstractSampler):
"""
On large batch, torch stack has impact on performance, but it's not obvious locally.
"""

def __init__(self):
pass

def sample_frames_uniformly(self, video_file, clips_per_video):
arr = np.fromfile(video_file, dtype=np.uint8)
video_tensor = torch.from_numpy(arr)
video_input = VideoArgs()
sampler_input = IndexBasedSamplerArgs(
sampler_type="uniform", clips_per_video=clips_per_video, frames_per_clip=1
)
sampler = VideoClipSampler(video_input, sampler_input)
clips = sampler(video_tensor)
return torch.stack([clip[0] for clip in clips])


class DecordSampler(AbstractSampler):
def __init__(self):
pass

def sample_frames_uniformly(self, video_file, clips_per_video):
decord.bridge.set_bridge("torch")
av_reader = decord.VideoReader(video_file)
num_frames = len(av_reader)
frame_indices = np.linspace(0, num_frames - 1, clips_per_video, dtype=int)
frames = av_reader.get_batch(frame_indices)
return frames


class TorchMMSamplerWithTorchVisionBackend(AbstractSampler):
"""
Here we use TorchMultimodal sampler as it's updated version on top of torchvision decoder.
"""

def __init__(self):
pass

def sample_frames_uniformly(self, video_file, clips_per_video):
arr = np.fromfile(video_file, dtype=np.uint8)
video_tensor = torch.from_numpy(arr)
sampler = tmm_vcs(
clip_sampler_type=ClipSamplerType("UNIFORM"),
clips_per_video=clips_per_video,
frames_per_clip=1,
frame_dilation=1,
)
return sampler(video_tensor)


class TorchVisionNewSamplerWithTorchVisionBackend(AbstractSampler):
def __init__(self):
pass

def sample_frames_uniformly(self, video_file, clips_per_video):
clip_sampling_strategy = UniformClipSamplingStrategy(
clips_per_video=clips_per_video
)
decoder = TVVideoClipDecoder(clip_length_in_frames=1, read_audio_stream=False)
sampler = ta_vcs(clip_sampling_strategy, decoder)
return sampler(str(video_file))


def main():
"""Benchmarks the performance of different samplers"""

parser = argparse.ArgumentParser()
parser.add_argument(
"--bm_small_video_speed",
help="Benchmark small video decoding speed",
default=True,
action=argparse.BooleanOptionalAction,
)
parser.add_argument(
"--bm_large_video_speed",
help="Benchmark large video decoding speed",
default=True,
action=argparse.BooleanOptionalAction,
from torchcodec.decoders import VideoDecoder
from torchcodec.samplers import clips_at_random_indices


def bench(f, *args, num_exp=100, warmup=0, **kwargs):

for _ in range(warmup):
f(*args, **kwargs)

times = []
for _ in range(num_exp):
start = perf_counter_ns()
f(*args, **kwargs)
end = perf_counter_ns()
times.append(end - start)
return torch.tensor(times).float()


def report_stats(times, unit="ms"):
mul = {
"ns": 1,
"µs": 1e-3,
"ms": 1e-6,
"s": 1e-9,
}[unit]
times = times * mul
std = times.std().item()
med = times.median().item()
print(f"{med = :.2f}{unit} +- {std:.2f}")
return med


def sample(num_clips):
decoder = VideoDecoder(VIDEO_PATH)
clips_at_random_indices(
decoder,
num_clips=num_clips,
num_frames_per_clip=10,
num_indices_between_frames=2,
)
parser.add_argument(
"--bm_video_speed_min_run_seconds",
help="Benchmark minimum run time, in seconds, to wait per datapoint",
type=float,
default=5.0,
)
args = parser.parse_args()

small_video_path = importlib.resources.path(__package__, "nasa_13013.mp4")
small_video_path = os.fspath(str(small_video_path))

large_video_path = importlib.resources.path(__package__, "853.mp4")
large_video_path = os.fspath(str(large_video_path))

clips_per_video = 8

sampler_dict = {}
sampler_dict["TorchCodecTimeBasedSampler"] = TorchCodecTimeBasedSampler()
sampler_dict["TorchCodecIndexBasedSampler"] = TorchCodecIndexBasedSampler()
sampler_dict["TorchCodecIndexBasedSamplerWithStackedOutput"] = (
TorchCodecIndexBasedSamplerWithStackedOutput()
)
sampler_dict["DecordSampler"] = DecordSampler()
sampler_dict["TorchMMSamplerWithTorchVisionBackend"] = (
TorchMMSamplerWithTorchVisionBackend()
)
sampler_dict["TorchVisionNewSamplerWithTorchVisionBackend"] = (
TorchVisionNewSamplerWithTorchVisionBackend()
)

results = []

for sampler_name, sampler in sampler_dict.items():
if args.bm_small_video_speed:
sampler_result = benchmark.Timer(
stmt="sampler.sample_frames_uniformly(video_file, clips_per_video)",
globals={
"video_file": small_video_path,
"clips_per_video": clips_per_video,
"sampler": sampler,
},
label="uniform sampling latency for 700KB video",
sub_label=sampler_name,
description=f"uniform sampling {clips_per_video} frames",
)
results.append(
sampler_result.blocked_autorange(
min_run_time=args.bm_video_speed_min_run_seconds
)
)

if args.bm_large_video_speed:
if sampler_name == "TorchMMSamplerWithTorchVisionBackend":
continue
sampler_result = benchmark.Timer(
stmt="sampler.sample_frames_uniformly(video_file, clips_per_video)",
globals={
"video_file": large_video_path,
"clips_per_video": clips_per_video,
"sampler": sampler,
},
label="uniform sampling latency for 50MB video",
sub_label=sampler_name,
description=f"uniform sampling {clips_per_video} frames",
)
results.append(
sampler_result.blocked_autorange(
min_run_time=args.bm_video_speed_min_run_seconds
)
)
VIDEO_PATH = Path(__file__).parent / "../../test/resources/nasa_13013.mp4"

compare = benchmark.Compare(results)
compare.print()
times = bench(sample, num_clips=1, num_exp=30, warmup=2)
report_stats(times, unit="ms")
times = bench(sample, num_clips=50, num_exp=30, warmup=2)
report_stats(times, unit="ms")
Loading