Skip to content
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.

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

Add adapter for HiSanta data #47

Draft
wants to merge 1 commit into
base: main
Choose a base branch
from
Draft
Changes from all commits
Commits
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
14 changes: 2 additions & 12 deletions poetry.lock

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

1 change: 1 addition & 0 deletions pyproject.toml
Original file line number Diff line number Diff line change
@@ -27,6 +27,7 @@ openai = "~1.33.0"
jiwer = "~3.0.4"
tensorboardx = "~2.6.2.2"
wandb = "~0.17.1"
google-cloud-storage = "~2.10"

[tool.poetry.group.dev.dependencies]
black = "~24.4.2"
84 changes: 81 additions & 3 deletions ultravox/data/datasets.py
Original file line number Diff line number Diff line change
@@ -1,14 +1,17 @@
import abc
import base64
import dataclasses
import datetime
import enum
import io
import json
import logging
import os
import tempfile
from typing import Any, Callable, Dict, List, Optional, Sequence
from typing import Any, Callable, Dict, Generator, List, Optional, Sequence

import datasets
import google.cloud.storage as gcs
import librosa
import numpy as np
import requests
@@ -60,7 +63,7 @@

# TODO(juberti): set these in the environment so they don't need to be hard-coded here.
os.environ["GOOGLE_APPLICATION_CREDENTIALS"] = "service_account.json"
os.environ["GOOGLE_CLOUD_PROJECT"] = "fixie-training"
os.environ["GOOGLE_CLOUD_PROJECT"] = "fixie-frame"


# Silence the spurious warnings coming from the MosaicML streaming library.
@@ -244,7 +247,7 @@ def _load_audio_dataset(
gcs_path += f"/{name}"
if split:
gcs_path += f"/{split}"
url = f"gs://fixie-datasets/mds/{gcs_path}"
url = f"gs://fixie-training-datasets/mds/{gcs_path}"
temp_dir = os.path.join(
tempfile.gettempdir(), f"mds_{gcs_path.replace('/', '_')}"
)
@@ -681,6 +684,80 @@ def _get_sample(self, idx, row) -> VoiceSample:
return self._get_transcribe_sample(idx, row, tcol="text")


class HiSantaDataset(data.IterableDataset):
"""
A proprietary dataset from post-processed conversations with Santa and
friends between 12/18 and 12/31 2023.
"""

class Subset(str, enum.Enum):
ALL = "all" # All recoverable data
BEST = "best" # Only the samples expected to have the best audio alignment

def __init__(self, args: VoiceDatasetArgs, subset: Subset = Subset.BEST) -> None:
super().__init__()
self._args = args # TODO(mdepinet): Respect whatever args we need to.
self._bucket = gcs.Client().get_bucket("hisanta-dataset")

self._conversations: list[str] = json.loads(
self._bucket.get_blob(f"{subset}.json").download_as_bytes()
)["conversations"]
"""List of references to conversation metadata JSON files in the bucket.
These all look like {conversation_id}/metadata.json."""
Comment on lines +705 to +706
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nit: Why use strings for comments?


self._system_prompts: list[dict] = json.loads(
self._bucket.get_blob("hisanta_prompts.json").download_as_bytes()
)
"""List of system prompts for each agent, with the time at which each became effective.
These all look like {"agent_id": AGENT_ID, "prompts": [{"start": ISO_TIMESTAMP, "prompt": PROMPT}]}
where the prompts are in chronological order."""

def __iter__(self) -> Generator[VoiceSample, Any, None]:
worker_info = data.get_worker_info()
start = worker_info.id if worker_info else 0
increment = worker_info.num_workers if worker_info else 1
for i in range(start, len(self._conversations), increment):
yield from self._from_conversation(i)
Comment on lines +719 to +720
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We'll probably have to experiment with how to form our samples here.
There are multiple issues to consider:

  1. How to do shuffle
  2. The length of each sample should be regulated: max_audio_duration_secs was an attempt at this, but generally the bottleneck is GPU memory


def _from_conversation(self, idx: int) -> Generator[VoiceSample, Any, None]:
conversation_id = self._conversations[idx].split("/")[0]
conversation = json.loads(
self._bucket.get_blob(self._conversations[idx]).download_as_bytes()
)
system_prompt = self._get_system_prompt(
conversation["agent_id"],
datetime.datetime.fromisoformat(conversation["call_time"]),
)
if not system_prompt:
return
history = [{"role": "system", "content": system_prompt}]
for message in conversation["messages"]:
if message["role"] == "assistant":
history.append({"role": "assistant", "content": message["message"]})
else:
audio = self._bucket.get_blob(
f"{conversation_id}/{message['speech']}"
).download_as_bytes()
yield VoiceSample(
messages=[*history, {"role": "user", "content": "<|audio|>"}],
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Current assumption is that the last message should be the assistant message.

audio=audio_from_buf(audio),
sample_rate=SAMPLE_RATE,
audio_transcript=message["message"],
)
history.append({"role": "user", "content": message["message"]})

def _get_system_prompt(
self, agent_id: str, call_time: datetime.datetime
) -> str | None:
for agent_prompts in self._system_prompts:
if agent_prompts["agent_id"] != agent_id:
continue
for prompt in reversed(agent_prompts["prompts"]):
if datetime.datetime.fromisoformat(prompt["start"]) <= call_time:
return prompt["prompt"]
return None # UGC agent


def create_dataset(name: str, args: VoiceDatasetArgs) -> data.IterableDataset:
DATASET_MAP: Dict[str, Any] = {
"anyinstruct": AnyInstructAnswerDataset,
@@ -690,6 +767,7 @@ def create_dataset(name: str, args: VoiceDatasetArgs) -> data.IterableDataset:
"boolq_in": BoolQInputDataset,
"boolq_extended": BoolQWithExtendedAnswerDataset,
"gigaspeech": GigaSpeechDataset,
"hisanta": HiSantaDataset,
"librispeech": LibriSpeechDataset,
"voxpopuli": VoxPopuliDataset,
"commonvoice": CommonVoiceDataset,
Loading