atksh atksh

Example of using OpenAI functions in completions with Python decorators

This example illustrates a way to utilize a function dynamically while querying an OpenAI GPT model. It uses the newly released functions support in the completion endpoints OpenAI provides.

The general concept is based on using a decorator to extract information from a function so it can be presented to the language model for use, and then pass the result of that function back to the completion endpoint for language augmentation.

In general, a wide variety of functions can be swapped in for use by the model. By changing the get_top_stories function, plus the prompt in run_conversation, you should be able to get the model to run your function without changing any of the other code.

Configuration

To use this, create a config.py file and add a variable with your OpenAI token:

Encoding high-quality HEVC content with FFmpeg - based NVENC encoder on supported hardware:

tl;dr

https://ffmpeg.zeranoe.com/forum/viewtopic.php?t=4766

ffmpeg -hide_banner -i input -c:v hevc_nvenc -preset hq -rc constqp -global_quality 21 -c:a libfdk_aac -ar:a 48000 -channel_layout:a 5.1 -ab:a 640k output

ffmpeg -hide_banner -i input -c:v hevc_nvenc -preset hq -rc vbr_hq -b:v 30000k -qmin 17 -qmax 21 -c:a libfdk_aac -ar:a 48000 -channel_layout:a 5.1 -ab:a 640k output

概要

Optunaというハイパーパラメータ最適化ツールを使って、FFmpegでの動画エンコードパラメータの最適化を試してみた結果のメモ
具体的には、決められた制約(後述)下で、画質(SSIM)を最大化するようなパラメータ群を自動で見つけ出すのが目的
結果としては、
- 画質的には、FFmpegが提供しているプリセットの中で二番目に重いもの(slower)より若干良い程度のパラメータ群が見つかった
- また、Optunaが見つけたパラメータ群の方がslowerに比べて、CPU負荷が小さかった

	import dataclasses as dc
	from jax import tree_util as jt


	def register_jax_dataclass(cls):
	"""Registers a dataclass as a JAX pytree."""
	if not dc.is_dataclass(cls):
	raise TypeError('%s is not a dataclass.' % cls)

	keys = [field.name for field in dc.fields(cls)]

	import numpy as np
	from tensorflow.python.summary.event_accumulator import EventAccumulator

	import matplotlib as mpl
	import matplotlib.pyplot as plt

	def plot_tensorflow_log(path):

	# Loading too much data is slow...
	tf_size_guidance = {