blob42
/
langchain
forked from Archives/langchain
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

YoutubeAudioLoader and updates to OpenAIWhisperParser (#5772)

Browse Source 
This introduces the `YoutubeAudioLoader`, which will load blobs from a
YouTube url and write them. Blobs are then parsed by
`OpenAIWhisperParser()`, as show in this
[PR](https://github.com/hwchase17/langchain/pull/5580), but we extend
the parser to split audio such that each chuck meets the 25MB OpenAI
size limit. As shown in the notebook, this enables a very simple UX:

```
# Transcribe the video to text
loader = GenericLoader(YoutubeAudioLoader([url],save_dir),OpenAIWhisperParser())
docs = loader.load()
``` 

Tested on full set of Karpathy lecture videos:

```
# Karpathy lecture videos
urls = ["https://youtu.be/VMj-3S1tku0"
        "https://youtu.be/PaCmpygFfXo",
        "https://youtu.be/TCH_1BHY58I",
        "https://youtu.be/P6sfmUTpUmc",
        "https://youtu.be/q8SA3rM6ckI",
        "https://youtu.be/t3YJ5hKiMQ0",
        "https://youtu.be/kCc8FmEb1nY"]

# Directory to save audio files 
save_dir = "~/Downloads/YouTube"
 
# Transcribe the videos to text
loader = GenericLoader(YoutubeAudioLoader(urls,save_dir),OpenAIWhisperParser())
docs = loader.load()
```
searx_updates
Lance Martin 11 months ago committed by GitHub
parent 2a4b32dee2
commit 4092fd21dc
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
5 changed files with 393 additions and 6 deletions
Show Stats Download Patch File Download Diff File

296
docs/modules/indexes/document_loaders/examples/youtube_audio.ipynb
Unescape Escape View File

@ -0,0 +1,296 @@
{
"cells": [
{
"cell_type": "markdown",
"id": "e48afb8d",
"metadata": {},
"source": [
"# Loading documents from a YouTube url\n",
"\n",
"Building chat or QA applications on YouTube videos is a topic of high interest.\n",
"\n",
"Below we show how to easily go from a YouTube url to text to chat!\n",
"\n",
"We wil use the `OpenAIWhisperParser`, which will use the OpenAI Whisper API to transcribe audio to text.\n",
"\n",
"Note: You will need to have an `OPENAI_API_KEY` supplied."
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "5f34e934",
"metadata": {},
"outputs": [],
"source": [
"from langchain.document_loaders.generic import GenericLoader\n",
"from langchain.document_loaders.parsers import OpenAIWhisperParser\n",
"from langchain.document_loaders.blob_loaders.youtube_audio import YoutubeAudioLoader"
]
},
{
"cell_type": "markdown",
"id": "85fc12bd",
"metadata": {},
"source": [
"We will use `yt_dlp` to download audio for YouTube urls.\n",
"\n",
"We will use `pydub` to split downloaded audio files (such that we adhere to Whisper API's 25MB file size limit)."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "fb5a6606",
"metadata": {},
"outputs": [],
"source": [
"! pip install yt_dlp\n",
"! pip install pydub"
]
},
{
"cell_type": "markdown",
"id": "b0e119f4",
"metadata": {},
"source": [
"### YouTube url to text\n",
"\n",
"Use `YoutubeAudioLoader` to fetch / download the audio files.\n",
"\n",
"Then, ues `OpenAIWhisperParser()` to transcribe them to text.\n",
"\n",
"Let's take the first lecture of Andrej Karpathy's YouTube course as an example! "
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "23e1e134",
"metadata": {
"scrolled": false
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[youtube] Extracting URL: https://youtu.be/kCc8FmEb1nY\n",
"[youtube] kCc8FmEb1nY: Downloading webpage\n",
"[youtube] kCc8FmEb1nY: Downloading android player API JSON\n",
"[info] kCc8FmEb1nY: Downloading 1 format(s): 140\n",
"[dashsegments] Total fragments: 11\n",
"[download] Destination: /Users/31treehaus/Desktop/AI/langchain-fork/docs/modules/indexes/document_loaders/examples/Let's build GPT： from scratch, in code, spelled out..m4a\n",
"[download] 100% of 107.73MiB in 00:00:18 at 5.92MiB/s \n",
"[FixupM4a] Correcting container of \"/Users/31treehaus/Desktop/AI/langchain-fork/docs/modules/indexes/document_loaders/examples/Let's build GPT： from scratch, in code, spelled out..m4a\"\n",
"[ExtractAudio] Not converting audio /Users/31treehaus/Desktop/AI/langchain-fork/docs/modules/indexes/document_loaders/examples/Let's build GPT： from scratch, in code, spelled out..m4a; file is already in target format m4a\n",
"[youtube] Extracting URL: https://youtu.be/VMj-3S1tku0\n",
"[youtube] VMj-3S1tku0: Downloading webpage\n",
"[youtube] VMj-3S1tku0: Downloading android player API JSON\n",
"[info] VMj-3S1tku0: Downloading 1 format(s): 140\n",
"[download] /Users/31treehaus/Desktop/AI/langchain-fork/docs/modules/indexes/document_loaders/examples/The spelled-out intro to neural networks and backpropagation： building micrograd.m4a has already been downloaded\n",
"[download] 100% of 134.98MiB\n",
"[ExtractAudio] Not converting audio /Users/31treehaus/Desktop/AI/langchain-fork/docs/modules/indexes/document_loaders/examples/The spelled-out intro to neural networks and backpropagation： building micrograd.m4a; file is already in target format m4a\n"
]
}
],
"source": [
"# Two Karpathy lecture videos\n",
"urls = [\"https://youtu.be/kCc8FmEb1nY\",\n",
" \"https://youtu.be/VMj-3S1tku0\"]\n",
"\n",
"# Directory to save audio files \n",
"save_dir = \"~/Downloads/YouTube\"\n",
"\n",
"# Transcribe the videos to text\n",
"loader = GenericLoader(YoutubeAudioLoader(urls,save_dir),OpenAIWhisperParser())\n",
"docs = loader.load()"
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "72a94fd8",
"metadata": {
"scrolled": false
},
"outputs": [
{
"data": {
"text/plain": [
"\"Hello, my name is Andrej and I've been training deep neural networks for a bit more than a decade. And in this lecture I'd like to show you what neural network training looks like under the hood. So in particular we are going to start with a blank Jupyter notebook and by the end of this lecture we will define and train a neural net and you'll get to see everything that goes on under the hood and exactly sort of how that works on an intuitive level. Now specifically what I would like to do is I w\""
]
},
"execution_count": 3,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Returns a list of Documents, which can be easily viewed or parsed\n",
"docs[0].page_content[0:500]"
]
},
{
"cell_type": "markdown",
"id": "93be6b49",
"metadata": {},
"source": [
"### Building a chat app from YouTube video\n",
"\n",
"Given `Documents`, we can easily enable chat / question+answering."
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "1823f042",
"metadata": {},
"outputs": [],
"source": [
"from langchain.chains import RetrievalQA\n",
"from langchain.vectorstores import FAISS\n",
"from langchain.chat_models import ChatOpenAI\n",
"from langchain.embeddings import OpenAIEmbeddings\n",
"from langchain.text_splitter import RecursiveCharacterTextSplitter"
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "7257cda1",
"metadata": {},
"outputs": [],
"source": [
"# Combine doc\n",
"combined_docs = [doc.page_content for doc in docs]\n",
"text = \" \".join(combined_docs)"
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "147c0c55",
"metadata": {},
"outputs": [],
"source": [
"# Split them\n",
"text_splitter = RecursiveCharacterTextSplitter(chunk_size = 1500, chunk_overlap = 150)\n",
"splits = text_splitter.split_text(text)"
]
},
{
"cell_type": "code",
"execution_count": 7,
"id": "f3556703",
"metadata": {},
"outputs": [],
"source": [
"# Build an index\n",
"embeddings = OpenAIEmbeddings()\n",
"vectordb = FAISS.from_texts(splits,embeddings)"
]
},
{
"cell_type": "code",
"execution_count": 8,
"id": "beaa99db",
"metadata": {},
"outputs": [],
"source": [
"# Build a QA chain\n",
"qa_chain = RetrievalQA.from_chain_type(llm = ChatOpenAI(model_name=\"gpt-3.5-turbo\", temperature=0),\n",
" chain_type=\"stuff\",\n",
" retriever=vectordb.as_retriever())"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "f2239a62",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"\"We need to zero out the gradient before backprop at each step because the backward pass accumulates gradients in the grad attribute of each parameter. If we don't reset the grad to zero before each backward pass, the gradients will accumulate and add up, leading to incorrect updates and slower convergence. By resetting the grad to zero before each backward pass, we ensure that the gradients are calculated correctly and that the optimization process works as intended.\""
]
},
"execution_count": 9,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Ask a question!\n",
"query = \"Why do we need to zero out the gradient before backprop at each step?\"\n",
"qa_chain.run(query)"
]
},
{
"cell_type": "code",
"execution_count": 10,
"id": "a8d01098",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"'In the context of transformers, an encoder is a component that reads in a sequence of input tokens and generates a sequence of hidden representations. On the other hand, a decoder is a component that takes in a sequence of hidden representations and generates a sequence of output tokens. The main difference between the two is that the encoder is used to encode the input sequence into a fixed-length representation, while the decoder is used to decode the fixed-length representation into an output sequence. In machine translation, for example, the encoder reads in the source language sentence and generates a fixed-length representation, which is then used by the decoder to generate the target language sentence.'"
]
},
"execution_count": 10,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"query = \"What is the difference between an encoder and decoder?\"\n",
"qa_chain.run(query)"
]
},
{
"cell_type": "code",
"execution_count": 11,
"id": "fe1e77dd",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"'For any token, x is the input vector that contains the private information of that token, k and q are the key and query vectors respectively, which are produced by forwarding linear modules on x, and v is the vector that is calculated by propagating the same linear module on x again. The key vector represents what the token contains, and the query vector represents what the token is looking for. The vector v is the information that the token will communicate to other tokens if it finds them interesting, and it gets aggregated for the purposes of the self-attention mechanism.'"
]
},
"execution_count": 11,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"query = \"For any token, what are x, k, v, and q?\"\n",
"qa_chain.run(query)"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.9.16"
}
},
"nbformat": 4,
"nbformat_minor": 5
}

3
langchain/document_loaders/blob_loaders/__init__.py
Unescape Escape View File

@ -1,4 +1,5 @@
from langchain.document_loaders.blob_loaders.file_system import FileSystemBlobLoader
from langchain.document_loaders.blob_loaders.schema import Blob, BlobLoader
from langchain.document_loaders.blob_loaders.youtube_audio import YoutubeAudioLoader
__all__ = ["BlobLoader", "Blob", "FileSystemBlobLoader"]
__all__ = ["BlobLoader", "Blob", "FileSystemBlobLoader", "YoutubeAudioLoader"]

50
langchain/document_loaders/blob_loaders/youtube_audio.py
Unescape Escape View File

@ -0,0 +1,50 @@
from typing import Iterable, List
from langchain.document_loaders.blob_loaders import FileSystemBlobLoader
from langchain.document_loaders.blob_loaders.schema import Blob, BlobLoader
class YoutubeAudioLoader(BlobLoader):
"""Load YouTube urls as audio file(s)."""
def __init__(self, urls: List[str], save_dir: str):
if not isinstance(urls, list):
raise TypeError("urls must be a list")
self.urls = urls
self.save_dir = save_dir
def yield_blobs(self) -> Iterable[Blob]:
"""Yield audio blobs for each url."""
try:
import yt_dlp
except ImportError:
raise ValueError(
"yt_dlp package not found, please install it with "
"`pip install yt_dlp`"
)
# Use yt_dlp to download audio given a YouTube url
ydl_opts = {
"format": "m4a/bestaudio/best",
"noplaylist": True,
"outtmpl": self.save_dir + "/%(title)s.%(ext)s",
"postprocessors": [
{
"key": "FFmpegExtractAudio",
"preferredcodec": "m4a",
}
],
}
for url in self.urls:
# Download file
with yt_dlp.YoutubeDL(ydl_opts) as ydl:
ydl.download(url)
# Yield the written blobs
loader = FileSystemBlobLoader(self.save_dir, glob="*.m4a")
for blob in loader.yield_blobs():
yield blob

43
langchain/document_loaders/parsers/audio.py
Unescape Escape View File

@ -12,10 +12,45 @@ class OpenAIWhisperParser(BaseBlobParser):
def lazy_parse(self, blob: Blob) -> Iterator[Document]:
"""Lazily parse the blob."""
import openai
import io
try:
import openai
except ImportError:
raise ValueError(
"openai package not found, please install it with "
"`pip install openai`"
)
try:
from pydub import AudioSegment
except ImportError:
raise ValueError(
"pydub package not found, please install it with " "`pip install pydub`"
)
# Audio file from disk
audio = AudioSegment.from_file(blob.path)
# Define the duration of each chunk in minutes
# Need to meet 25MB size limit for Whisper API
chunk_duration = 20
chunk_duration_ms = chunk_duration * 60 * 1000
# Split the audio into chunk_duration_ms chunks
for split_number, i in enumerate(range(0, len(audio), chunk_duration_ms)):
# Audio chunk
chunk = audio[i : i + chunk_duration_ms]
file_obj = io.BytesIO(chunk.export(format="mp3").read())
if blob.source is not None:
file_obj.name = blob.source + f"_part_{split_number}.mp3"
else:
file_obj.name = f"part_{split_number}.mp3"
# Transcribe
print(f"Transcribing part {split_number+1}!")
transcript = openai.Audio.transcribe("whisper-1", file_obj)
with blob.as_bytes_io() as f:
transcript = openai.Audio.transcribe("whisper-1", f)
yield Document(
page_content=transcript.text, metadata={"source": blob.source}
page_content=transcript.text,
metadata={"source": blob.source, "chunk": split_number},
)

7
tests/unit_tests/document_loaders/blob_loaders/test_public_api.py
Unescape Escape View File

@ -3,4 +3,9 @@ from langchain.document_loaders.blob_loaders import __all__
def test_public_api() -> None:
"""Hard-code public API to help determine if we have broken it."""
assert sorted(__all__) == ["Blob", "BlobLoader", "FileSystemBlobLoader"]
assert sorted(__all__) == [
"Blob",
"BlobLoader",
"FileSystemBlobLoader",
"YoutubeAudioLoader",
]

Write Preview
Loading…
Cancel
Save